Training in progress, step 100, checkpoint

.gitattributes

@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
 checkpoint-50/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+checkpoint-100/tokenizer.json filter=lfs diff=lfs merge=lfs -text

checkpoint-100/config.json

@@ -0,0 +1,29 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 8960,
+  "max_position_embeddings": 131072,
+  "max_window_layers": 21,
+  "model_type": "qwen2",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 10000,
+  "sliding_window": 4096,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.3",
+  "use_cache": false,
+  "use_mrope": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-100/generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151646,
+  "do_sample": true,
+  "eos_token_id": 151643,
+  "temperature": 0.6,
+  "top_p": 0.95,
+  "transformers_version": "4.51.3"
+}

checkpoint-100/global_step100/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2dd970f479c53af33bbe281636fb393b324a5b59446f9544efe3b488c7bb351
+size 5331274140

checkpoint-100/global_step100/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:664b2098192b09ed4a439e92d9c16e9cbb770dc63255a17a6a326b91b1229d96
+size 5331276572

checkpoint-100/global_step100/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f56a43d8ecad68271b64b290df4fea17c00bbfdd3d49b7b62510c5eb5b20665
+size 5331276892

checkpoint-100/global_step100/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd2bb1f7042d7df8ce7a2aae59e269dbb975a14eba30f04b21d20474390f76ca
+size 5331273884

checkpoint-100/global_step100/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a11aa81d048fbe7360e965fd5d68d6d44608ade5417a16903af2aa12ffe07cf
+size 3554267640

checkpoint-100/latest

@@ -0,0 +1 @@
+global_step100

checkpoint-100/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:169c2b3f46ff468195686e5cbfe35ded2e5ca03e24eb746876ff95a5fdc4b8c6
+size 3554214752

checkpoint-100/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0d1a91a61e9e1ed69964104ac8f3aa26499a6e8e9d2d722eb85fd9a8e3ffb73
+size 14960

checkpoint-100/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3ca66af15bb10392f9a774ffaa873492d302a2ddfb8b6eac8eee1a4de2cae7d
+size 14960

checkpoint-100/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e434082aea9c3bea58ff81d3f8a44b08d0622929a033989dd9f58d4f1868619
+size 14960

checkpoint-100/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8cb6657acd6447ce0c00b88d9826ad435e4f9a7f857570e059e5c9afd7c4c13c
+size 14960

checkpoint-100/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b0d6ed3e119807e165d53a18d2ec22befd359c1465f6aeaa69d1d7eb1452246
+size 1064

checkpoint-100/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<｜begin▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-100/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4256422650d141f228fe954acee98679da412984c29a569877eefd3af69315a
+size 11422959

checkpoint-100/tokenizer_config.json

@@ -0,0 +1,195 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<｜end▁of▁sentence｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<｜User｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151645": {
+      "content": "<｜Assistant｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151646": {
+      "content": "<｜begin▁of▁sentence｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|EOT|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151648": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151649": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "bos_token": "<｜begin▁of▁sentence｜>",
+  "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %}{%- for message in messages %}{%- if message['role'] == 'system' %}{% set ns.system_prompt = message['content'] %}{%- endif %}{%- endfor %}{{bos_token}}{{ns.system_prompt}}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is none %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls']%}{%- if not ns.is_first %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin��>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- set ns.is_first = true -%}{%- else %}{{'\\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- endfor %}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is not none %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'\\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜><think>\\n'}}{% endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<｜end▁of▁sentence｜>",
+  "extra_special_tokens": {},
+  "legacy": true,
+  "model_max_length": 16384,
+  "pad_token": "<｜end▁of▁sentence｜>",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizerFast",
+  "unk_token": null,
+  "use_default_system_prompt": false
+}

checkpoint-100/trainer_state.json

@@ -0,0 +1,2534 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.11428571428571428,
+  "eval_steps": 500,
+  "global_step": 100,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1734.0,
+      "completions/mean_length": 1702.03125,
+      "completions/mean_terminated_length": 993.6190795898438,
+      "completions/min_length": 483.0,
+      "completions/min_terminated_length": 483.0,
+      "epoch": 0.001142857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2837146520614624,
+      "learning_rate": 0.0,
+      "loss": -0.0,
+      "num_tokens": 118418.0,
+      "reward": -0.09800112247467041,
+      "reward_std": 0.3028089702129364,
+      "rewards/cosine_scaled_reward/mean": -0.09800112992525101,
+      "rewards/cosine_scaled_reward/std": 0.37953105568885803,
+      "step": 1
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1894.0,
+      "completions/mean_length": 1738.90625,
+      "completions/mean_terminated_length": 949.0,
+      "completions/min_length": 435.0,
+      "completions/min_terminated_length": 435.0,
+      "epoch": 0.002285714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24220912158489227,
+      "learning_rate": 2e-08,
+      "loss": -0.0,
+      "num_tokens": 239748.0,
+      "reward": 0.020556632429361343,
+      "reward_std": 0.3545936942100525,
+      "rewards/cosine_scaled_reward/mean": 0.020556632429361343,
+      "rewards/cosine_scaled_reward/std": 0.4492928683757782,
+      "step": 2
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.921875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 869.0,
+      "completions/mean_length": 1946.515625,
+      "completions/mean_terminated_length": 749.0,
+      "completions/min_length": 609.0,
+      "completions/min_terminated_length": 609.0,
+      "epoch": 0.0034285714285714284,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24765528738498688,
+      "learning_rate": 4e-08,
+      "loss": -0.0,
+      "num_tokens": 374797.0,
+      "reward": -0.20057085156440735,
+      "reward_std": 0.13691216707229614,
+      "rewards/cosine_scaled_reward/mean": -0.20057085156440735,
+      "rewards/cosine_scaled_reward/std": 0.16282624006271362,
+      "step": 3
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.578125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1983.0,
+      "completions/mean_length": 1592.0,
+      "completions/mean_terminated_length": 967.1111450195312,
+      "completions/min_length": 516.0,
+      "completions/min_terminated_length": 516.0,
+      "epoch": 0.004571428571428572,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.28862521052360535,
+      "learning_rate": 6e-08,
+      "loss": 0.0,
+      "num_tokens": 486493.0,
+      "reward": -0.19111667573451996,
+      "reward_std": 0.19739457964897156,
+      "rewards/cosine_scaled_reward/mean": -0.19111669063568115,
+      "rewards/cosine_scaled_reward/std": 0.22545036673545837,
+      "step": 4
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.890625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1939.0,
+      "completions/mean_length": 1976.578125,
+      "completions/mean_terminated_length": 1395.0001220703125,
+      "completions/min_length": 610.0,
+      "completions/min_terminated_length": 610.0,
+      "epoch": 0.005714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23521216213703156,
+      "learning_rate": 8e-08,
+      "loss": 0.0,
+      "num_tokens": 623810.0,
+      "reward": -0.2342512309551239,
+      "reward_std": 0.16005605459213257,
+      "rewards/cosine_scaled_reward/mean": -0.2342512309551239,
+      "rewards/cosine_scaled_reward/std": 0.20709452033042908,
+      "step": 5
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.8125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1987.0,
+      "completions/mean_length": 1840.125,
+      "completions/mean_terminated_length": 939.3333740234375,
+      "completions/min_length": 552.0,
+      "completions/min_terminated_length": 552.0,
+      "epoch": 0.006857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2831529676914215,
+      "learning_rate": 1e-07,
+      "loss": 0.0,
+      "num_tokens": 753226.0,
+      "reward": -0.1443408578634262,
+      "reward_std": 0.25838011503219604,
+      "rewards/cosine_scaled_reward/mean": -0.1443408727645874,
+      "rewards/cosine_scaled_reward/std": 0.3164331316947937,
+      "step": 6
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2002.0,
+      "completions/mean_length": 1974.265625,
+      "completions/mean_terminated_length": 1458.125,
+      "completions/min_length": 1153.0,
+      "completions/min_terminated_length": 1153.0,
+      "epoch": 0.008,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22311581671237946,
+      "learning_rate": 1.2e-07,
+      "loss": 0.0,
+      "num_tokens": 889987.0,
+      "reward": -0.15585696697235107,
+      "reward_std": 0.21075330674648285,
+      "rewards/cosine_scaled_reward/mean": -0.15585698187351227,
+      "rewards/cosine_scaled_reward/std": 0.3327982723712921,
+      "step": 7
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1411.0,
+      "completions/mean_length": 1701.46875,
+      "completions/mean_terminated_length": 815.888916015625,
+      "completions/min_length": 346.0,
+      "completions/min_terminated_length": 346.0,
+      "epoch": 0.009142857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23218390345573425,
+      "learning_rate": 1.4e-07,
+      "loss": -0.0,
+      "num_tokens": 1009297.0,
+      "reward": -0.019736051559448242,
+      "reward_std": 0.22464922070503235,
+      "rewards/cosine_scaled_reward/mean": -0.01973605342209339,
+      "rewards/cosine_scaled_reward/std": 0.46309077739715576,
+      "step": 8
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.859375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1836.0,
+      "completions/mean_length": 1936.96875,
+      "completions/mean_terminated_length": 1258.4444580078125,
+      "completions/min_length": 839.0,
+      "completions/min_terminated_length": 839.0,
+      "epoch": 0.010285714285714285,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2455250322818756,
+      "learning_rate": 1.6e-07,
+      "loss": -0.0,
+      "num_tokens": 1144719.0,
+      "reward": -0.22108668088912964,
+      "reward_std": 0.20550987124443054,
+      "rewards/cosine_scaled_reward/mean": -0.22108666598796844,
+      "rewards/cosine_scaled_reward/std": 0.27375248074531555,
+      "step": 9
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1579.0,
+      "completions/mean_length": 1662.0625,
+      "completions/mean_terminated_length": 813.0,
+      "completions/min_length": 389.0,
+      "completions/min_terminated_length": 389.0,
+      "epoch": 0.011428571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.26574036478996277,
+      "learning_rate": 1.8e-07,
+      "loss": -0.0,
+      "num_tokens": 1261923.0,
+      "reward": -0.140568345785141,
+      "reward_std": 0.2796468734741211,
+      "rewards/cosine_scaled_reward/mean": -0.140568345785141,
+      "rewards/cosine_scaled_reward/std": 0.35179150104522705,
+      "step": 10
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.921875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1335.0,
+      "completions/mean_length": 1970.859375,
+      "completions/mean_terminated_length": 1060.5999755859375,
+      "completions/min_length": 906.0,
+      "completions/min_terminated_length": 906.0,
+      "epoch": 0.012571428571428572,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24890610575675964,
+      "learning_rate": 2e-07,
+      "loss": -0.0,
+      "num_tokens": 1399730.0,
+      "reward": -0.2551690638065338,
+      "reward_std": 0.16209062933921814,
+      "rewards/cosine_scaled_reward/mean": -0.2551690638065338,
+      "rewards/cosine_scaled_reward/std": 0.2319207787513733,
+      "step": 11
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.65625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2031.0,
+      "completions/mean_length": 1798.71875,
+      "completions/mean_terminated_length": 1322.8182373046875,
+      "completions/min_length": 724.0,
+      "completions/min_terminated_length": 724.0,
+      "epoch": 0.013714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2804766595363617,
+      "learning_rate": 2.1999999999999998e-07,
+      "loss": 0.0,
+      "num_tokens": 1525792.0,
+      "reward": -0.19796784222126007,
+      "reward_std": 0.30078738927841187,
+      "rewards/cosine_scaled_reward/mean": -0.19796785712242126,
+      "rewards/cosine_scaled_reward/std": 0.3346545696258545,
+      "step": 12
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1800.0,
+      "completions/mean_length": 1816.890625,
+      "completions/mean_terminated_length": 1123.5625,
+      "completions/min_length": 583.0,
+      "completions/min_terminated_length": 583.0,
+      "epoch": 0.014857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2471778392791748,
+      "learning_rate": 2.4e-07,
+      "loss": -0.0,
+      "num_tokens": 1653113.0,
+      "reward": -0.17365078628063202,
+      "reward_std": 0.23729698359966278,
+      "rewards/cosine_scaled_reward/mean": -0.17365078628063202,
+      "rewards/cosine_scaled_reward/std": 0.2726025879383087,
+      "step": 13
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1529.0,
+      "completions/mean_length": 1815.046875,
+      "completions/mean_terminated_length": 1171.0,
+      "completions/min_length": 639.0,
+      "completions/min_terminated_length": 639.0,
+      "epoch": 0.016,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22734108567237854,
+      "learning_rate": 2.6e-07,
+      "loss": 0.0,
+      "num_tokens": 1779884.0,
+      "reward": -0.086978480219841,
+      "reward_std": 0.2551291584968567,
+      "rewards/cosine_scaled_reward/mean": -0.0869784876704216,
+      "rewards/cosine_scaled_reward/std": 0.4508184790611267,
+      "step": 14
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1354.0,
+      "completions/mean_length": 1705.421875,
+      "completions/mean_terminated_length": 758.2941284179688,
+      "completions/min_length": 429.0,
+      "completions/min_terminated_length": 429.0,
+      "epoch": 0.017142857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25105422735214233,
+      "learning_rate": 2.8e-07,
+      "loss": -0.0,
+      "num_tokens": 1899951.0,
+      "reward": 0.025415867567062378,
+      "reward_std": 0.13560885190963745,
+      "rewards/cosine_scaled_reward/mean": 0.025415875017642975,
+      "rewards/cosine_scaled_reward/std": 0.4663754105567932,
+      "step": 15
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 1.0,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 0.0,
+      "completions/mean_length": 2048.0,
+      "completions/mean_terminated_length": 0.0,
+      "completions/min_length": 2048.0,
+      "completions/min_terminated_length": 0.0,
+      "epoch": 0.018285714285714287,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23334357142448425,
+      "learning_rate": 3e-07,
+      "loss": -0.0,
+      "num_tokens": 2041463.0,
+      "reward": -0.2220873385667801,
+      "reward_std": 0.17581966519355774,
+      "rewards/cosine_scaled_reward/mean": -0.2220873236656189,
+      "rewards/cosine_scaled_reward/std": 0.1694367378950119,
+      "step": 16
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.546875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1898.0,
+      "completions/mean_length": 1524.9375,
+      "completions/mean_terminated_length": 893.6551513671875,
+      "completions/min_length": 343.0,
+      "completions/min_terminated_length": 343.0,
+      "epoch": 0.019428571428571427,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.33780622482299805,
+      "learning_rate": 3.2e-07,
+      "loss": -0.0,
+      "num_tokens": 2149579.0,
+      "reward": -0.026115939021110535,
+      "reward_std": 0.3175298571586609,
+      "rewards/cosine_scaled_reward/mean": -0.026115931570529938,
+      "rewards/cosine_scaled_reward/std": 0.4766712486743927,
+      "step": 17
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1635.0,
+      "completions/mean_length": 1771.34375,
+      "completions/mean_terminated_length": 1116.105224609375,
+      "completions/min_length": 538.0,
+      "completions/min_terminated_length": 538.0,
+      "epoch": 0.02057142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23123449087142944,
+      "learning_rate": 3.4000000000000003e-07,
+      "loss": -0.0,
+      "num_tokens": 2273321.0,
+      "reward": -0.15853706002235413,
+      "reward_std": 0.27896177768707275,
+      "rewards/cosine_scaled_reward/mean": -0.15853706002235413,
+      "rewards/cosine_scaled_reward/std": 0.3426607847213745,
+      "step": 18
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2013.0,
+      "completions/mean_length": 1811.953125,
+      "completions/mean_terminated_length": 1159.3529052734375,
+      "completions/min_length": 484.0,
+      "completions/min_terminated_length": 484.0,
+      "epoch": 0.021714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25707289576530457,
+      "learning_rate": 3.6e-07,
+      "loss": -0.0,
+      "num_tokens": 2400542.0,
+      "reward": -0.052606794983148575,
+      "reward_std": 0.31571486592292786,
+      "rewards/cosine_scaled_reward/mean": -0.052606794983148575,
+      "rewards/cosine_scaled_reward/std": 0.44901713728904724,
+      "step": 19
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.65625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1807.0,
+      "completions/mean_length": 1632.953125,
+      "completions/mean_terminated_length": 840.5909423828125,
+      "completions/min_length": 379.0,
+      "completions/min_terminated_length": 379.0,
+      "epoch": 0.022857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25764355063438416,
+      "learning_rate": 3.7999999999999996e-07,
+      "loss": 0.0,
+      "num_tokens": 2516403.0,
+      "reward": -0.07391424477100372,
+      "reward_std": 0.2678168714046478,
+      "rewards/cosine_scaled_reward/mean": -0.07391423732042313,
+      "rewards/cosine_scaled_reward/std": 0.3888758718967438,
+      "step": 20
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1854.0,
+      "completions/mean_length": 1820.125,
+      "completions/mean_terminated_length": 1136.5,
+      "completions/min_length": 344.0,
+      "completions/min_terminated_length": 344.0,
+      "epoch": 0.024,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.27439141273498535,
+      "learning_rate": 4e-07,
+      "loss": 0.0,
+      "num_tokens": 2643699.0,
+      "reward": -0.16270118951797485,
+      "reward_std": 0.22588439285755157,
+      "rewards/cosine_scaled_reward/mean": -0.16270118951797485,
+      "rewards/cosine_scaled_reward/std": 0.39143073558807373,
+      "step": 21
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.40625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1741.0,
+      "completions/mean_length": 1271.359375,
+      "completions/mean_terminated_length": 739.9736938476562,
+      "completions/min_length": 282.0,
+      "completions/min_terminated_length": 282.0,
+      "epoch": 0.025142857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.37971845269203186,
+      "learning_rate": 4.1999999999999995e-07,
+      "loss": -0.0,
+      "num_tokens": 2734082.0,
+      "reward": -0.00552794337272644,
+      "reward_std": 0.23386958241462708,
+      "rewards/cosine_scaled_reward/mean": -0.005527939647436142,
+      "rewards/cosine_scaled_reward/std": 0.4625597596168518,
+      "step": 22
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.609375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1935.0,
+      "completions/mean_length": 1670.296875,
+      "completions/mean_terminated_length": 1081.0799560546875,
+      "completions/min_length": 472.0,
+      "completions/min_terminated_length": 472.0,
+      "epoch": 0.026285714285714287,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.28573453426361084,
+      "learning_rate": 4.3999999999999997e-07,
+      "loss": 0.0,
+      "num_tokens": 2851773.0,
+      "reward": -0.18269123136997223,
+      "reward_std": 0.2168647199869156,
+      "rewards/cosine_scaled_reward/mean": -0.18269124627113342,
+      "rewards/cosine_scaled_reward/std": 0.2703794836997986,
+      "step": 23
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1802.0,
+      "completions/mean_length": 1757.296875,
+      "completions/mean_terminated_length": 1068.7894287109375,
+      "completions/min_length": 327.0,
+      "completions/min_terminated_length": 327.0,
+      "epoch": 0.027428571428571427,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2553797662258148,
+      "learning_rate": 4.6e-07,
+      "loss": 0.0,
+      "num_tokens": 2975168.0,
+      "reward": -0.23130035400390625,
+      "reward_std": 0.35076260566711426,
+      "rewards/cosine_scaled_reward/mean": -0.23130035400390625,
+      "rewards/cosine_scaled_reward/std": 0.3866168260574341,
+      "step": 24
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1584.0,
+      "completions/mean_length": 1744.28125,
+      "completions/mean_terminated_length": 833.125,
+      "completions/min_length": 504.0,
+      "completions/min_terminated_length": 504.0,
+      "epoch": 0.02857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2636294960975647,
+      "learning_rate": 4.8e-07,
+      "loss": -0.0,
+      "num_tokens": 3097098.0,
+      "reward": -0.19239474833011627,
+      "reward_std": 0.2867633104324341,
+      "rewards/cosine_scaled_reward/mean": -0.19239474833011627,
+      "rewards/cosine_scaled_reward/std": 0.347222238779068,
+      "step": 25
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.796875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2033.0,
+      "completions/mean_length": 1932.09375,
+      "completions/mean_terminated_length": 1477.3846435546875,
+      "completions/min_length": 895.0,
+      "completions/min_terminated_length": 895.0,
+      "epoch": 0.029714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22351376712322235,
+      "learning_rate": 5e-07,
+      "loss": -0.0,
+      "num_tokens": 3231384.0,
+      "reward": -0.006307817995548248,
+      "reward_std": 0.2015555500984192,
+      "rewards/cosine_scaled_reward/mean": -0.006307825446128845,
+      "rewards/cosine_scaled_reward/std": 0.4079793393611908,
+      "step": 26
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.8125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1974.0,
+      "completions/mean_length": 1899.25,
+      "completions/mean_terminated_length": 1254.666748046875,
+      "completions/min_length": 545.0,
+      "completions/min_terminated_length": 545.0,
+      "epoch": 0.030857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2670150697231293,
+      "learning_rate": 5.2e-07,
+      "loss": -0.0,
+      "num_tokens": 3363224.0,
+      "reward": -0.22071197628974915,
+      "reward_std": 0.2118011713027954,
+      "rewards/cosine_scaled_reward/mean": -0.22071197628974915,
+      "rewards/cosine_scaled_reward/std": 0.2716290354728699,
+      "step": 27
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1947.0,
+      "completions/mean_length": 1767.609375,
+      "completions/mean_terminated_length": 926.4375,
+      "completions/min_length": 438.0,
+      "completions/min_terminated_length": 438.0,
+      "epoch": 0.032,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25918784737586975,
+      "learning_rate": 5.4e-07,
+      "loss": -0.0,
+      "num_tokens": 3486687.0,
+      "reward": -0.10919298231601715,
+      "reward_std": 0.2716072201728821,
+      "rewards/cosine_scaled_reward/mean": -0.10919298231601715,
+      "rewards/cosine_scaled_reward/std": 0.44544270634651184,
+      "step": 28
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.890625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1951.0,
+      "completions/mean_length": 1932.203125,
+      "completions/mean_terminated_length": 989.2857666015625,
+      "completions/min_length": 603.0,
+      "completions/min_terminated_length": 603.0,
+      "epoch": 0.03314285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24401192367076874,
+      "learning_rate": 5.6e-07,
+      "loss": 0.0,
+      "num_tokens": 3620820.0,
+      "reward": -0.19096782803535461,
+      "reward_std": 0.15806984901428223,
+      "rewards/cosine_scaled_reward/mean": -0.19096782803535461,
+      "rewards/cosine_scaled_reward/std": 0.181764155626297,
+      "step": 29
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1928.0,
+      "completions/mean_length": 1880.71875,
+      "completions/mean_terminated_length": 1334.2667236328125,
+      "completions/min_length": 604.0,
+      "completions/min_terminated_length": 604.0,
+      "epoch": 0.03428571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22094956040382385,
+      "learning_rate": 5.8e-07,
+      "loss": -0.0,
+      "num_tokens": 3751722.0,
+      "reward": -0.21267297863960266,
+      "reward_std": 0.24843861162662506,
+      "rewards/cosine_scaled_reward/mean": -0.21267297863960266,
+      "rewards/cosine_scaled_reward/std": 0.29802343249320984,
+      "step": 30
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.78125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1348.0,
+      "completions/mean_length": 1786.234375,
+      "completions/mean_terminated_length": 851.357177734375,
+      "completions/min_length": 355.0,
+      "completions/min_terminated_length": 355.0,
+      "epoch": 0.03542857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2912121117115021,
+      "learning_rate": 6e-07,
+      "loss": -0.0,
+      "num_tokens": 3876537.0,
+      "reward": -0.2621557414531708,
+      "reward_std": 0.18612943589687347,
+      "rewards/cosine_scaled_reward/mean": -0.2621557414531708,
+      "rewards/cosine_scaled_reward/std": 0.22891530394554138,
+      "step": 31
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.859375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1841.0,
+      "completions/mean_length": 1948.765625,
+      "completions/mean_terminated_length": 1342.3333740234375,
+      "completions/min_length": 536.0,
+      "completions/min_terminated_length": 536.0,
+      "epoch": 0.036571428571428574,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2303810715675354,
+      "learning_rate": 6.2e-07,
+      "loss": 0.0,
+      "num_tokens": 4011610.0,
+      "reward": -0.1655973494052887,
+      "reward_std": 0.2392224669456482,
+      "rewards/cosine_scaled_reward/mean": -0.1655973345041275,
+      "rewards/cosine_scaled_reward/std": 0.3260692358016968,
+      "step": 32
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.90625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1972.0,
+      "completions/mean_length": 1984.0,
+      "completions/mean_terminated_length": 1365.3333740234375,
+      "completions/min_length": 965.0,
+      "completions/min_terminated_length": 965.0,
+      "epoch": 0.037714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23169051110744476,
+      "learning_rate": 6.4e-07,
+      "loss": 0.0,
+      "num_tokens": 4149802.0,
+      "reward": -0.22799505293369293,
+      "reward_std": 0.24000275135040283,
+      "rewards/cosine_scaled_reward/mean": -0.22799506783485413,
+      "rewards/cosine_scaled_reward/std": 0.30748653411865234,
+      "step": 33
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.609375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1977.0,
+      "completions/mean_length": 1700.859375,
+      "completions/mean_terminated_length": 1159.3199462890625,
+      "completions/min_length": 433.0,
+      "completions/min_terminated_length": 433.0,
+      "epoch": 0.038857142857142854,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2647433578968048,
+      "learning_rate": 6.6e-07,
+      "loss": 0.0,
+      "num_tokens": 4268209.0,
+      "reward": -0.07232969254255295,
+      "reward_std": 0.3570185899734497,
+      "rewards/cosine_scaled_reward/mean": -0.07232969999313354,
+      "rewards/cosine_scaled_reward/std": 0.4520716369152069,
+      "step": 34
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1458.0,
+      "completions/mean_length": 1884.625,
+      "completions/mean_terminated_length": 741.0,
+      "completions/min_length": 358.0,
+      "completions/min_terminated_length": 358.0,
+      "epoch": 0.04,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2681647539138794,
+      "learning_rate": 6.800000000000001e-07,
+      "loss": 0.0,
+      "num_tokens": 4400321.0,
+      "reward": -0.21119418740272522,
+      "reward_std": 0.2156996876001358,
+      "rewards/cosine_scaled_reward/mean": -0.21119415760040283,
+      "rewards/cosine_scaled_reward/std": 0.304564893245697,
+      "step": 35
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.96875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2002.0,
+      "completions/mean_length": 2032.765625,
+      "completions/mean_terminated_length": 1560.5,
+      "completions/min_length": 1119.0,
+      "completions/min_terminated_length": 1119.0,
+      "epoch": 0.04114285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25201615691185,
+      "learning_rate": 7e-07,
+      "loss": -0.0,
+      "num_tokens": 4541530.0,
+      "reward": -0.2148258090019226,
+      "reward_std": 0.1970210075378418,
+      "rewards/cosine_scaled_reward/mean": -0.2148257941007614,
+      "rewards/cosine_scaled_reward/std": 0.21921320259571075,
+      "step": 36
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.859375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2005.0,
+      "completions/mean_length": 1954.5,
+      "completions/mean_terminated_length": 1383.111083984375,
+      "completions/min_length": 901.0,
+      "completions/min_terminated_length": 901.0,
+      "epoch": 0.04228571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.29214274883270264,
+      "learning_rate": 7.2e-07,
+      "loss": 0.0,
+      "num_tokens": 4677642.0,
+      "reward": -0.23519155383110046,
+      "reward_std": 0.14085054397583008,
+      "rewards/cosine_scaled_reward/mean": -0.23519155383110046,
+      "rewards/cosine_scaled_reward/std": 0.17065586149692535,
+      "step": 37
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2016.0,
+      "completions/mean_length": 1949.1875,
+      "completions/mean_terminated_length": 1257.5,
+      "completions/min_length": 1042.0,
+      "completions/min_terminated_length": 1042.0,
+      "epoch": 0.04342857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2337840050458908,
+      "learning_rate": 7.4e-07,
+      "loss": -0.0,
+      "num_tokens": 4814102.0,
+      "reward": -0.16185586154460907,
+      "reward_std": 0.19152981042861938,
+      "rewards/cosine_scaled_reward/mean": -0.16185584664344788,
+      "rewards/cosine_scaled_reward/std": 0.3005273640155792,
+      "step": 38
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1953.0,
+      "completions/mean_length": 1810.515625,
+      "completions/mean_terminated_length": 666.2727661132812,
+      "completions/min_length": 246.0,
+      "completions/min_terminated_length": 246.0,
+      "epoch": 0.044571428571428574,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.246645987033844,
+      "learning_rate": 7.599999999999999e-07,
+      "loss": 0.0,
+      "num_tokens": 4940759.0,
+      "reward": -0.10980962216854095,
+      "reward_std": 0.18094567954540253,
+      "rewards/cosine_scaled_reward/mean": -0.10980962216854095,
+      "rewards/cosine_scaled_reward/std": 0.3624936640262604,
+      "step": 39
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.65625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1754.0,
+      "completions/mean_length": 1700.796875,
+      "completions/mean_terminated_length": 1037.95458984375,
+      "completions/min_length": 524.0,
+      "completions/min_terminated_length": 524.0,
+      "epoch": 0.045714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.26321786642074585,
+      "learning_rate": 7.799999999999999e-07,
+      "loss": -0.0,
+      "num_tokens": 5059682.0,
+      "reward": -0.14547404646873474,
+      "reward_std": 0.22270715236663818,
+      "rewards/cosine_scaled_reward/mean": -0.14547404646873474,
+      "rewards/cosine_scaled_reward/std": 0.4000875651836395,
+      "step": 40
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1939.0,
+      "completions/mean_length": 1860.328125,
+      "completions/mean_terminated_length": 1415.8421630859375,
+      "completions/min_length": 982.0,
+      "completions/min_terminated_length": 982.0,
+      "epoch": 0.046857142857142854,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21273446083068848,
+      "learning_rate": 8e-07,
+      "loss": -0.0,
+      "num_tokens": 5189895.0,
+      "reward": -0.24220962822437286,
+      "reward_std": 0.27360057830810547,
+      "rewards/cosine_scaled_reward/mean": -0.24220961332321167,
+      "rewards/cosine_scaled_reward/std": 0.33429500460624695,
+      "step": 41
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1547.0,
+      "completions/mean_length": 1694.375,
+      "completions/mean_terminated_length": 539.2000122070312,
+      "completions/min_length": 131.0,
+      "completions/min_terminated_length": 131.0,
+      "epoch": 0.048,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3549652099609375,
+      "learning_rate": 8.199999999999999e-07,
+      "loss": -0.0,
+      "num_tokens": 5308695.0,
+      "reward": -0.22589105367660522,
+      "reward_std": 0.16009008884429932,
+      "rewards/cosine_scaled_reward/mean": -0.22589105367660522,
+      "rewards/cosine_scaled_reward/std": 0.17985297739505768,
+      "step": 42
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.796875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2006.0,
+      "completions/mean_length": 1824.75,
+      "completions/mean_terminated_length": 948.923095703125,
+      "completions/min_length": 473.0,
+      "completions/min_terminated_length": 473.0,
+      "epoch": 0.04914285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25625720620155334,
+      "learning_rate": 8.399999999999999e-07,
+      "loss": -0.0,
+      "num_tokens": 5437095.0,
+      "reward": -0.10874830186367035,
+      "reward_std": 0.2326180636882782,
+      "rewards/cosine_scaled_reward/mean": -0.10874830186367035,
+      "rewards/cosine_scaled_reward/std": 0.3275902569293976,
+      "step": 43
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1751.0,
+      "completions/mean_length": 1673.734375,
+      "completions/mean_terminated_length": 787.3157958984375,
+      "completions/min_length": 484.0,
+      "completions/min_terminated_length": 484.0,
+      "epoch": 0.05028571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3032245934009552,
+      "learning_rate": 8.599999999999999e-07,
+      "loss": 0.0,
+      "num_tokens": 5554910.0,
+      "reward": -0.1157154068350792,
+      "reward_std": 0.2323075234889984,
+      "rewards/cosine_scaled_reward/mean": -0.1157153993844986,
+      "rewards/cosine_scaled_reward/std": 0.4071435034275055,
+      "step": 44
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.9375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1931.0,
+      "completions/mean_length": 2031.03125,
+      "completions/mean_terminated_length": 1776.5,
+      "completions/min_length": 1421.0,
+      "completions/min_terminated_length": 1421.0,
+      "epoch": 0.05142857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2320922464132309,
+      "learning_rate": 8.799999999999999e-07,
+      "loss": 0.0,
+      "num_tokens": 5696552.0,
+      "reward": -0.22731460630893707,
+      "reward_std": 0.19835877418518066,
+      "rewards/cosine_scaled_reward/mean": -0.22731460630893707,
+      "rewards/cosine_scaled_reward/std": 0.28479474782943726,
+      "step": 45
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 981.0,
+      "completions/mean_length": 1890.3125,
+      "completions/mean_terminated_length": 786.5,
+      "completions/min_length": 490.0,
+      "completions/min_terminated_length": 490.0,
+      "epoch": 0.052571428571428575,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2494276612997055,
+      "learning_rate": 9e-07,
+      "loss": 0.0,
+      "num_tokens": 5828700.0,
+      "reward": -0.23243775963783264,
+      "reward_std": 0.18319474160671234,
+      "rewards/cosine_scaled_reward/mean": -0.23243777453899384,
+      "rewards/cosine_scaled_reward/std": 0.20973731577396393,
+      "step": 46
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1398.0,
+      "completions/mean_length": 1672.09375,
+      "completions/mean_terminated_length": 711.4444580078125,
+      "completions/min_length": 303.0,
+      "completions/min_terminated_length": 303.0,
+      "epoch": 0.053714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3419908881187439,
+      "learning_rate": 9.2e-07,
+      "loss": 0.0,
+      "num_tokens": 5946114.0,
+      "reward": -0.16157878935337067,
+      "reward_std": 0.24494563043117523,
+      "rewards/cosine_scaled_reward/mean": -0.16157880425453186,
+      "rewards/cosine_scaled_reward/std": 0.39992472529411316,
+      "step": 47
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1501.0,
+      "completions/mean_length": 1787.171875,
+      "completions/mean_terminated_length": 935.1333618164062,
+      "completions/min_length": 687.0,
+      "completions/min_terminated_length": 687.0,
+      "epoch": 0.054857142857142854,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25991642475128174,
+      "learning_rate": 9.399999999999999e-07,
+      "loss": 0.0,
+      "num_tokens": 6071037.0,
+      "reward": -0.1829870045185089,
+      "reward_std": 0.2542135417461395,
+      "rewards/cosine_scaled_reward/mean": -0.1829870045185089,
+      "rewards/cosine_scaled_reward/std": 0.30597779154777527,
+      "step": 48
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.5625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1981.0,
+      "completions/mean_length": 1565.34375,
+      "completions/mean_terminated_length": 944.7857666015625,
+      "completions/min_length": 322.0,
+      "completions/min_terminated_length": 322.0,
+      "epoch": 0.056,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.27452352643013,
+      "learning_rate": 9.6e-07,
+      "loss": 0.0,
+      "num_tokens": 6181283.0,
+      "reward": -0.22301900386810303,
+      "reward_std": 0.25131016969680786,
+      "rewards/cosine_scaled_reward/mean": -0.22301900386810303,
+      "rewards/cosine_scaled_reward/std": 0.2918049991130829,
+      "step": 49
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1947.0,
+      "completions/mean_length": 1775.515625,
+      "completions/mean_terminated_length": 885.4000244140625,
+      "completions/min_length": 280.0,
+      "completions/min_terminated_length": 280.0,
+      "epoch": 0.05714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22758428752422333,
+      "learning_rate": 9.8e-07,
+      "loss": 0.0,
+      "num_tokens": 6305732.0,
+      "reward": -0.10754476487636566,
+      "reward_std": 0.18711507320404053,
+      "rewards/cosine_scaled_reward/mean": -0.10754477977752686,
+      "rewards/cosine_scaled_reward/std": 0.39105597138404846,
+      "step": 50
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.578125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1939.0,
+      "completions/mean_length": 1507.5625,
+      "completions/mean_terminated_length": 766.9629516601562,
+      "completions/min_length": 440.0,
+      "completions/min_terminated_length": 440.0,
+      "epoch": 0.05828571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.29917222261428833,
+      "learning_rate": 1e-06,
+      "loss": 0.0,
+      "num_tokens": 6412424.0,
+      "reward": -0.08381433039903641,
+      "reward_std": 0.23327183723449707,
+      "rewards/cosine_scaled_reward/mean": -0.08381432294845581,
+      "rewards/cosine_scaled_reward/std": 0.40033307671546936,
+      "step": 51
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1933.0,
+      "completions/mean_length": 1746.015625,
+      "completions/mean_terminated_length": 840.0625,
+      "completions/min_length": 347.0,
+      "completions/min_terminated_length": 347.0,
+      "epoch": 0.05942857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2824826240539551,
+      "learning_rate": 9.999890338174275e-07,
+      "loss": -0.0,
+      "num_tokens": 6535521.0,
+      "reward": -0.2775638699531555,
+      "reward_std": 0.17903020977973938,
+      "rewards/cosine_scaled_reward/mean": -0.2775638699531555,
+      "rewards/cosine_scaled_reward/std": 0.38567760586738586,
+      "step": 52
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1862.0,
+      "completions/mean_length": 1765.71875,
+      "completions/mean_terminated_length": 918.875,
+      "completions/min_length": 460.0,
+      "completions/min_terminated_length": 460.0,
+      "epoch": 0.060571428571428575,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2795548737049103,
+      "learning_rate": 9.999561358041868e-07,
+      "loss": -0.0,
+      "num_tokens": 6659359.0,
+      "reward": -0.18778130412101746,
+      "reward_std": 0.24159184098243713,
+      "rewards/cosine_scaled_reward/mean": -0.18778130412101746,
+      "rewards/cosine_scaled_reward/std": 0.2979832589626312,
+      "step": 53
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2034.0,
+      "completions/mean_length": 1877.984375,
+      "completions/mean_terminated_length": 1367.9375,
+      "completions/min_length": 536.0,
+      "completions/min_terminated_length": 536.0,
+      "epoch": 0.061714285714285715,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.236692875623703,
+      "learning_rate": 9.999013075636804e-07,
+      "loss": 0.0,
+      "num_tokens": 6790694.0,
+      "reward": -0.09228484332561493,
+      "reward_std": 0.3374499976634979,
+      "rewards/cosine_scaled_reward/mean": -0.09228484332561493,
+      "rewards/cosine_scaled_reward/std": 0.4543565809726715,
+      "step": 54
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.859375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 720.0,
+      "completions/mean_length": 1855.140625,
+      "completions/mean_terminated_length": 676.5555419921875,
+      "completions/min_length": 597.0,
+      "completions/min_terminated_length": 597.0,
+      "epoch": 0.06285714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24421174824237823,
+      "learning_rate": 9.998245517681593e-07,
+      "loss": 0.0,
+      "num_tokens": 6919711.0,
+      "reward": -0.19803781807422638,
+      "reward_std": 0.1785231977701187,
+      "rewards/cosine_scaled_reward/mean": -0.19803781807422638,
+      "rewards/cosine_scaled_reward/std": 0.3721012771129608,
+      "step": 55
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2027.0,
+      "completions/mean_length": 1840.703125,
+      "completions/mean_terminated_length": 1163.533447265625,
+      "completions/min_length": 657.0,
+      "completions/min_terminated_length": 657.0,
+      "epoch": 0.064,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23600365221500397,
+      "learning_rate": 9.997258721585931e-07,
+      "loss": -0.0,
+      "num_tokens": 7048476.0,
+      "reward": -0.09674595296382904,
+      "reward_std": 0.21479913592338562,
+      "rewards/cosine_scaled_reward/mean": -0.09674594551324844,
+      "rewards/cosine_scaled_reward/std": 0.4473191201686859,
+      "step": 56
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.90625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1681.0,
+      "completions/mean_length": 1997.5625,
+      "completions/mean_terminated_length": 1510.0,
+      "completions/min_length": 1387.0,
+      "completions/min_terminated_length": 1387.0,
+      "epoch": 0.06514285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.19085553288459778,
+      "learning_rate": 9.996052735444862e-07,
+      "loss": 0.0,
+      "num_tokens": 7187888.0,
+      "reward": -0.2958947420120239,
+      "reward_std": 0.1703263819217682,
+      "rewards/cosine_scaled_reward/mean": -0.2958947420120239,
+      "rewards/cosine_scaled_reward/std": 0.18720079958438873,
+      "step": 57
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.578125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1996.0,
+      "completions/mean_length": 1591.40625,
+      "completions/mean_terminated_length": 965.7037353515625,
+      "completions/min_length": 415.0,
+      "completions/min_terminated_length": 415.0,
+      "epoch": 0.06628571428571428,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.26593509316444397,
+      "learning_rate": 9.994627618036452e-07,
+      "loss": -0.0,
+      "num_tokens": 7299834.0,
+      "reward": -0.0999627411365509,
+      "reward_std": 0.32584434747695923,
+      "rewards/cosine_scaled_reward/mean": -0.0999627485871315,
+      "rewards/cosine_scaled_reward/std": 0.4625846743583679,
+      "step": 58
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1490.0,
+      "completions/mean_length": 1744.640625,
+      "completions/mean_terminated_length": 905.941162109375,
+      "completions/min_length": 463.0,
+      "completions/min_terminated_length": 463.0,
+      "epoch": 0.06742857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.250278115272522,
+      "learning_rate": 9.992983438818915e-07,
+      "loss": -0.0,
+      "num_tokens": 7421955.0,
+      "reward": -0.16149799525737762,
+      "reward_std": 0.21139998733997345,
+      "rewards/cosine_scaled_reward/mean": -0.16149798035621643,
+      "rewards/cosine_scaled_reward/std": 0.3698217272758484,
+      "step": 59
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.796875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1265.0,
+      "completions/mean_length": 1805.390625,
+      "completions/mean_terminated_length": 853.6154174804688,
+      "completions/min_length": 520.0,
+      "completions/min_terminated_length": 520.0,
+      "epoch": 0.06857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23754432797431946,
+      "learning_rate": 9.991120277927223e-07,
+      "loss": 0.0,
+      "num_tokens": 7548388.0,
+      "reward": -0.2758587598800659,
+      "reward_std": 0.18496021628379822,
+      "rewards/cosine_scaled_reward/mean": -0.2758587598800659,
+      "rewards/cosine_scaled_reward/std": 0.22098895907402039,
+      "step": 60
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.84375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1578.0,
+      "completions/mean_length": 1863.265625,
+      "completions/mean_terminated_length": 865.7000122070312,
+      "completions/min_length": 528.0,
+      "completions/min_terminated_length": 528.0,
+      "epoch": 0.06971428571428571,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21628443896770477,
+      "learning_rate": 9.989038226169207e-07,
+      "loss": -0.0,
+      "num_tokens": 7679157.0,
+      "reward": -0.11532291769981384,
+      "reward_std": 0.24975456297397614,
+      "rewards/cosine_scaled_reward/mean": -0.11532291769981384,
+      "rewards/cosine_scaled_reward/std": 0.32742080092430115,
+      "step": 61
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.65625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1909.0,
+      "completions/mean_length": 1690.296875,
+      "completions/mean_terminated_length": 1007.4091186523438,
+      "completions/min_length": 543.0,
+      "completions/min_terminated_length": 543.0,
+      "epoch": 0.07085714285714285,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2538006901741028,
+      "learning_rate": 9.98673738502114e-07,
+      "loss": -0.0,
+      "num_tokens": 7797568.0,
+      "reward": -0.08548027276992798,
+      "reward_std": 0.1828608512878418,
+      "rewards/cosine_scaled_reward/mean": -0.08548027276992798,
+      "rewards/cosine_scaled_reward/std": 0.31418856978416443,
+      "step": 62
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.515625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2045.0,
+      "completions/mean_length": 1631.5,
+      "completions/mean_terminated_length": 1188.1290283203125,
+      "completions/min_length": 528.0,
+      "completions/min_terminated_length": 528.0,
+      "epoch": 0.072,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.27852457761764526,
+      "learning_rate": 9.98421786662277e-07,
+      "loss": 0.0,
+      "num_tokens": 7912240.0,
+      "reward": 0.03178010880947113,
+      "reward_std": 0.39872580766677856,
+      "rewards/cosine_scaled_reward/mean": 0.03178010135889053,
+      "rewards/cosine_scaled_reward/std": 0.4946252107620239,
+      "step": 63
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.65625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1871.0,
+      "completions/mean_length": 1675.578125,
+      "completions/mean_terminated_length": 964.5909423828125,
+      "completions/min_length": 472.0,
+      "completions/min_terminated_length": 472.0,
+      "epoch": 0.07314285714285715,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.267963171005249,
+      "learning_rate": 9.981479793771866e-07,
+      "loss": 0.0,
+      "num_tokens": 8030429.0,
+      "reward": -0.19527338445186615,
+      "reward_std": 0.2819081246852875,
+      "rewards/cosine_scaled_reward/mean": -0.19527339935302734,
+      "rewards/cosine_scaled_reward/std": 0.3602358102798462,
+      "step": 64
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1593.0,
+      "completions/mean_length": 1671.53125,
+      "completions/mean_terminated_length": 709.4444580078125,
+      "completions/min_length": 295.0,
+      "completions/min_terminated_length": 295.0,
+      "epoch": 0.07428571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25890231132507324,
+      "learning_rate": 9.97852329991824e-07,
+      "loss": -0.0,
+      "num_tokens": 8148111.0,
+      "reward": -0.17763729393482208,
+      "reward_std": 0.1911587119102478,
+      "rewards/cosine_scaled_reward/mean": -0.17763729393482208,
+      "rewards/cosine_scaled_reward/std": 0.4043731391429901,
+      "step": 65
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.5,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1958.0,
+      "completions/mean_length": 1330.71875,
+      "completions/mean_terminated_length": 613.4375,
+      "completions/min_length": 195.0,
+      "completions/min_terminated_length": 195.0,
+      "epoch": 0.07542857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.4998987913131714,
+      "learning_rate": 9.975348529157229e-07,
+      "loss": -0.0,
+      "num_tokens": 8243509.0,
+      "reward": -0.012211084365844727,
+      "reward_std": 0.25645655393600464,
+      "rewards/cosine_scaled_reward/mean": -0.012211091816425323,
+      "rewards/cosine_scaled_reward/std": 0.4760035276412964,
+      "step": 66
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.921875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1362.0,
+      "completions/mean_length": 1970.15625,
+      "completions/mean_terminated_length": 1051.5999755859375,
+      "completions/min_length": 715.0,
+      "completions/min_terminated_length": 715.0,
+      "epoch": 0.07657142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22345179319381714,
+      "learning_rate": 9.971955636222684e-07,
+      "loss": -0.0,
+      "num_tokens": 8380335.0,
+      "reward": -0.27880045771598816,
+      "reward_std": 0.169667050242424,
+      "rewards/cosine_scaled_reward/mean": -0.27880045771598816,
+      "rewards/cosine_scaled_reward/std": 0.18985651433467865,
+      "step": 67
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1628.0,
+      "completions/mean_length": 1223.421875,
+      "completions/mean_terminated_length": 728.6749877929688,
+      "completions/min_length": 338.0,
+      "completions/min_terminated_length": 338.0,
+      "epoch": 0.07771428571428571,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3179270625114441,
+      "learning_rate": 9.968344786479415e-07,
+      "loss": -0.0,
+      "num_tokens": 8467890.0,
+      "reward": -0.09876523166894913,
+      "reward_std": 0.25151342153549194,
+      "rewards/cosine_scaled_reward/mean": -0.09876523166894913,
+      "rewards/cosine_scaled_reward/std": 0.4221951961517334,
+      "step": 68
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.5625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1582.0,
+      "completions/mean_length": 1574.234375,
+      "completions/mean_terminated_length": 965.107177734375,
+      "completions/min_length": 518.0,
+      "completions/min_terminated_length": 518.0,
+      "epoch": 0.07885714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.292468398809433,
+      "learning_rate": 9.964516155915151e-07,
+      "loss": -0.0,
+      "num_tokens": 8578985.0,
+      "reward": -0.20737677812576294,
+      "reward_std": 0.23497402667999268,
+      "rewards/cosine_scaled_reward/mean": -0.20737677812576294,
+      "rewards/cosine_scaled_reward/std": 0.3156755864620209,
+      "step": 69
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.890625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 866.0,
+      "completions/mean_length": 1893.546875,
+      "completions/mean_terminated_length": 635.857177734375,
+      "completions/min_length": 518.0,
+      "completions/min_terminated_length": 518.0,
+      "epoch": 0.08,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2380189746618271,
+      "learning_rate": 9.960469931131936e-07,
+      "loss": 0.0,
+      "num_tokens": 8711628.0,
+      "reward": -0.25674766302108765,
+      "reward_std": 0.1897822916507721,
+      "rewards/cosine_scaled_reward/mean": -0.25674766302108765,
+      "rewards/cosine_scaled_reward/std": 0.2669999301433563,
+      "step": 70
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1986.0,
+      "completions/mean_length": 1716.5625,
+      "completions/mean_terminated_length": 1037.90478515625,
+      "completions/min_length": 367.0,
+      "completions/min_terminated_length": 367.0,
+      "epoch": 0.08114285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2737840414047241,
+      "learning_rate": 9.956206309337066e-07,
+      "loss": 0.0,
+      "num_tokens": 8832208.0,
+      "reward": -0.11062799394130707,
+      "reward_std": 0.27241969108581543,
+      "rewards/cosine_scaled_reward/mean": -0.11062799394130707,
+      "rewards/cosine_scaled_reward/std": 0.43007227778434753,
+      "step": 71
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.59375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2043.0,
+      "completions/mean_length": 1669.015625,
+      "completions/mean_terminated_length": 1115.115478515625,
+      "completions/min_length": 369.0,
+      "completions/min_terminated_length": 369.0,
+      "epoch": 0.08228571428571428,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.27252519130706787,
+      "learning_rate": 9.951725498333448e-07,
+      "loss": 0.0,
+      "num_tokens": 8949377.0,
+      "reward": -0.2082766592502594,
+      "reward_std": 0.1827523410320282,
+      "rewards/cosine_scaled_reward/mean": -0.2082766592502594,
+      "rewards/cosine_scaled_reward/std": 0.18022844195365906,
+      "step": 72
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 1.0,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 0.0,
+      "completions/mean_length": 2048.0,
+      "completions/mean_terminated_length": 0.0,
+      "completions/min_length": 2048.0,
+      "completions/min_terminated_length": 0.0,
+      "epoch": 0.08342857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20380868017673492,
+      "learning_rate": 9.947027716509488e-07,
+      "loss": 0.0,
+      "num_tokens": 9091177.0,
+      "reward": -0.29910945892333984,
+      "reward_std": 0.12098947167396545,
+      "rewards/cosine_scaled_reward/mean": -0.29910945892333984,
+      "rewards/cosine_scaled_reward/std": 0.1714438796043396,
+      "step": 73
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.796875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1975.0,
+      "completions/mean_length": 1885.90625,
+      "completions/mean_terminated_length": 1250.0,
+      "completions/min_length": 725.0,
+      "completions/min_terminated_length": 725.0,
+      "epoch": 0.08457142857142858,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.262696772813797,
+      "learning_rate": 9.942113192828444e-07,
+      "loss": -0.0,
+      "num_tokens": 9221803.0,
+      "reward": -0.15267591178417206,
+      "reward_std": 0.23455429077148438,
+      "rewards/cosine_scaled_reward/mean": -0.15267591178417206,
+      "rewards/cosine_scaled_reward/std": 0.41386422514915466,
+      "step": 74
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1987.0,
+      "completions/mean_length": 1865.65625,
+      "completions/mean_terminated_length": 1318.625,
+      "completions/min_length": 966.0,
+      "completions/min_terminated_length": 966.0,
+      "epoch": 0.08571428571428572,
+      "frac_reward_zero_std": 0.125,
+      "grad_norm": 0.19511669874191284,
+      "learning_rate": 9.93698216681727e-07,
+      "loss": 0.0,
+      "num_tokens": 9352165.0,
+      "reward": -0.09251219034194946,
+      "reward_std": 0.182725191116333,
+      "rewards/cosine_scaled_reward/mean": -0.09251218289136887,
+      "rewards/cosine_scaled_reward/std": 0.47868576645851135,
+      "step": 75
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1665.0,
+      "completions/mean_length": 1767.421875,
+      "completions/mean_terminated_length": 991.7058715820312,
+      "completions/min_length": 436.0,
+      "completions/min_terminated_length": 436.0,
+      "epoch": 0.08685714285714285,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2677210569381714,
+      "learning_rate": 9.931634888554935e-07,
+      "loss": 0.0,
+      "num_tokens": 9475680.0,
+      "reward": -0.2391628623008728,
+      "reward_std": 0.16363291442394257,
+      "rewards/cosine_scaled_reward/mean": -0.2391628623008728,
+      "rewards/cosine_scaled_reward/std": 0.18309317529201508,
+      "step": 76
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.78125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2002.0,
+      "completions/mean_length": 1851.609375,
+      "completions/mean_terminated_length": 1150.21435546875,
+      "completions/min_length": 411.0,
+      "completions/min_terminated_length": 411.0,
+      "epoch": 0.088,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21603599190711975,
+      "learning_rate": 9.926071618660237e-07,
+      "loss": 0.0,
+      "num_tokens": 9605759.0,
+      "reward": -0.15259909629821777,
+      "reward_std": 0.212618887424469,
+      "rewards/cosine_scaled_reward/mean": -0.15259911119937897,
+      "rewards/cosine_scaled_reward/std": 0.2940331995487213,
+      "step": 77
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.796875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1857.0,
+      "completions/mean_length": 1851.640625,
+      "completions/mean_terminated_length": 1081.3077392578125,
+      "completions/min_length": 670.0,
+      "completions/min_terminated_length": 670.0,
+      "epoch": 0.08914285714285715,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20200739800930023,
+      "learning_rate": 9.9202926282791e-07,
+      "loss": 0.0,
+      "num_tokens": 9734984.0,
+      "reward": -0.18924658000469208,
+      "reward_std": 0.24043609201908112,
+      "rewards/cosine_scaled_reward/mean": -0.18924658000469208,
+      "rewards/cosine_scaled_reward/std": 0.38954904675483704,
+      "step": 78
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1807.0,
+      "completions/mean_length": 1608.40625,
+      "completions/mean_terminated_length": 875.75,
+      "completions/min_length": 249.0,
+      "completions/min_terminated_length": 249.0,
+      "epoch": 0.09028571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.31782233715057373,
+      "learning_rate": 9.91429819907136e-07,
+      "loss": -0.0,
+      "num_tokens": 9848018.0,
+      "reward": -0.1820095181465149,
+      "reward_std": 0.25530290603637695,
+      "rewards/cosine_scaled_reward/mean": -0.1820095181465149,
+      "rewards/cosine_scaled_reward/std": 0.31191888451576233,
+      "step": 79
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.890625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2047.0,
+      "completions/mean_length": 1976.84375,
+      "completions/mean_terminated_length": 1397.4285888671875,
+      "completions/min_length": 501.0,
+      "completions/min_terminated_length": 501.0,
+      "epoch": 0.09142857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24291428923606873,
+      "learning_rate": 9.908088623197048e-07,
+      "loss": 0.0,
+      "num_tokens": 9984928.0,
+      "reward": -0.253532737493515,
+      "reward_std": 0.19657698273658752,
+      "rewards/cosine_scaled_reward/mean": -0.2535327672958374,
+      "rewards/cosine_scaled_reward/std": 0.2723200023174286,
+      "step": 80
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1586.0,
+      "completions/mean_length": 1705.015625,
+      "completions/mean_terminated_length": 828.5,
+      "completions/min_length": 314.0,
+      "completions/min_terminated_length": 314.0,
+      "epoch": 0.09257142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3213472068309784,
+      "learning_rate": 9.901664203302124e-07,
+      "loss": -0.0,
+      "num_tokens": 10105321.0,
+      "reward": -0.1452670842409134,
+      "reward_std": 0.16492897272109985,
+      "rewards/cosine_scaled_reward/mean": -0.1452670842409134,
+      "rewards/cosine_scaled_reward/std": 0.20188800990581512,
+      "step": 81
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1601.0,
+      "completions/mean_length": 1707.203125,
+      "completions/mean_terminated_length": 1009.3809814453125,
+      "completions/min_length": 406.0,
+      "completions/min_terminated_length": 406.0,
+      "epoch": 0.09371428571428571,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25050660967826843,
+      "learning_rate": 9.895025252503755e-07,
+      "loss": -0.0,
+      "num_tokens": 10224910.0,
+      "reward": -0.07721791416406631,
+      "reward_std": 0.26486068964004517,
+      "rewards/cosine_scaled_reward/mean": -0.07721789926290512,
+      "rewards/cosine_scaled_reward/std": 0.4591779112815857,
+      "step": 82
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1707.0,
+      "completions/mean_length": 1616.609375,
+      "completions/mean_terminated_length": 897.625,
+      "completions/min_length": 441.0,
+      "completions/min_terminated_length": 441.0,
+      "epoch": 0.09485714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2958391010761261,
+      "learning_rate": 9.888172094375033e-07,
+      "loss": 0.0,
+      "num_tokens": 10339461.0,
+      "reward": -0.05388225242495537,
+      "reward_std": 0.23644787073135376,
+      "rewards/cosine_scaled_reward/mean": -0.053882256150245667,
+      "rewards/cosine_scaled_reward/std": 0.376263827085495,
+      "step": 83
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1666.0,
+      "completions/mean_length": 1838.828125,
+      "completions/mean_terminated_length": 831.0,
+      "completions/min_length": 499.0,
+      "completions/min_terminated_length": 499.0,
+      "epoch": 0.096,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23179632425308228,
+      "learning_rate": 9.881105062929221e-07,
+      "loss": -0.0,
+      "num_tokens": 10467842.0,
+      "reward": -0.15529119968414307,
+      "reward_std": 0.30153706669807434,
+      "rewards/cosine_scaled_reward/mean": -0.15529119968414307,
+      "rewards/cosine_scaled_reward/std": 0.4041438102722168,
+      "step": 84
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2021.0,
+      "completions/mean_length": 1840.71875,
+      "completions/mean_terminated_length": 1384.7000732421875,
+      "completions/min_length": 751.0,
+      "completions/min_terminated_length": 751.0,
+      "epoch": 0.09714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.18997986614704132,
+      "learning_rate": 9.873824502603459e-07,
+      "loss": -0.0,
+      "num_tokens": 10595968.0,
+      "reward": -0.09931906312704086,
+      "reward_std": 0.2868148386478424,
+      "rewards/cosine_scaled_reward/mean": -0.09931905567646027,
+      "rewards/cosine_scaled_reward/std": 0.32533466815948486,
+      "step": 85
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1373.0,
+      "completions/mean_length": 1669.25,
+      "completions/mean_terminated_length": 772.2105102539062,
+      "completions/min_length": 495.0,
+      "completions/min_terminated_length": 495.0,
+      "epoch": 0.09828571428571428,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2640744745731354,
+      "learning_rate": 9.866330768241983e-07,
+      "loss": 0.0,
+      "num_tokens": 10713656.0,
+      "reward": -0.09163744747638702,
+      "reward_std": 0.25668954849243164,
+      "rewards/cosine_scaled_reward/mean": -0.09163745492696762,
+      "rewards/cosine_scaled_reward/std": 0.34459924697875977,
+      "step": 86
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.609375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1999.0,
+      "completions/mean_length": 1626.609375,
+      "completions/mean_terminated_length": 969.239990234375,
+      "completions/min_length": 385.0,
+      "completions/min_terminated_length": 385.0,
+      "epoch": 0.09942857142857142,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2926872968673706,
+      "learning_rate": 9.85862422507884e-07,
+      "loss": -0.0,
+      "num_tokens": 10827879.0,
+      "reward": -0.20403151214122772,
+      "reward_std": 0.28549331426620483,
+      "rewards/cosine_scaled_reward/mean": -0.20403149724006653,
+      "rewards/cosine_scaled_reward/std": 0.32589223980903625,
+      "step": 87
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1864.0,
+      "completions/mean_length": 1735.953125,
+      "completions/mean_terminated_length": 1049.4500732421875,
+      "completions/min_length": 351.0,
+      "completions/min_terminated_length": 351.0,
+      "epoch": 0.10057142857142858,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.26702597737312317,
+      "learning_rate": 9.850705248720068e-07,
+      "loss": 0.0,
+      "num_tokens": 10949492.0,
+      "reward": 0.03890814632177353,
+      "reward_std": 0.3359295129776001,
+      "rewards/cosine_scaled_reward/mean": 0.03890814632177353,
+      "rewards/cosine_scaled_reward/std": 0.45631229877471924,
+      "step": 88
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.515625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1979.0,
+      "completions/mean_length": 1619.453125,
+      "completions/mean_terminated_length": 1163.258056640625,
+      "completions/min_length": 623.0,
+      "completions/min_terminated_length": 623.0,
+      "epoch": 0.10171428571428572,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21441271901130676,
+      "learning_rate": 9.8425742251254e-07,
+      "loss": -0.0,
+      "num_tokens": 11064137.0,
+      "reward": -0.0988616794347763,
+      "reward_std": 0.3224140405654907,
+      "rewards/cosine_scaled_reward/mean": -0.09886167198419571,
+      "rewards/cosine_scaled_reward/std": 0.41691890358924866,
+      "step": 89
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.53125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1753.0,
+      "completions/mean_length": 1592.1875,
+      "completions/mean_terminated_length": 1075.60009765625,
+      "completions/min_length": 459.0,
+      "completions/min_terminated_length": 459.0,
+      "epoch": 0.10285714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.30840882658958435,
+      "learning_rate": 9.83423155058946e-07,
+      "loss": -0.0,
+      "num_tokens": 11176037.0,
+      "reward": -0.22794684767723083,
+      "reward_std": 0.19634509086608887,
+      "rewards/cosine_scaled_reward/mean": -0.22794684767723083,
+      "rewards/cosine_scaled_reward/std": 0.2059042751789093,
+      "step": 90
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1933.0,
+      "completions/mean_length": 1857.125,
+      "completions/mean_terminated_length": 1369.3333740234375,
+      "completions/min_length": 876.0,
+      "completions/min_terminated_length": 876.0,
+      "epoch": 0.104,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24654391407966614,
+      "learning_rate": 9.825677631722435e-07,
+      "loss": 0.0,
+      "num_tokens": 11305461.0,
+      "reward": -0.06898833811283112,
+      "reward_std": 0.24478675425052643,
+      "rewards/cosine_scaled_reward/mean": -0.06898833811283112,
+      "rewards/cosine_scaled_reward/std": 0.4049251973628998,
+      "step": 91
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1291.0,
+      "completions/mean_length": 1659.859375,
+      "completions/mean_terminated_length": 865.0952758789062,
+      "completions/min_length": 459.0,
+      "completions/min_terminated_length": 459.0,
+      "epoch": 0.10514285714285715,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2405616194009781,
+      "learning_rate": 9.816912885430258e-07,
+      "loss": 0.0,
+      "num_tokens": 11421684.0,
+      "reward": -0.21882590651512146,
+      "reward_std": 0.186202734708786,
+      "rewards/cosine_scaled_reward/mean": -0.21882590651512146,
+      "rewards/cosine_scaled_reward/std": 0.20097385346889496,
+      "step": 92
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.984375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2046.0,
+      "completions/mean_length": 2047.96875,
+      "completions/mean_terminated_length": 2046.0,
+      "completions/min_length": 2046.0,
+      "completions/min_terminated_length": 2046.0,
+      "epoch": 0.10628571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24273920059204102,
+      "learning_rate": 9.807937738894303e-07,
+      "loss": 0.0,
+      "num_tokens": 11564098.0,
+      "reward": -0.25700533390045166,
+      "reward_std": 0.11929697543382645,
+      "rewards/cosine_scaled_reward/mean": -0.2570053040981293,
+      "rewards/cosine_scaled_reward/std": 0.1724296510219574,
+      "step": 93
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1979.0,
+      "completions/mean_length": 1864.96875,
+      "completions/mean_terminated_length": 983.0909423828125,
+      "completions/min_length": 619.0,
+      "completions/min_terminated_length": 619.0,
+      "epoch": 0.10742857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2541305422782898,
+      "learning_rate": 9.798752629550546e-07,
+      "loss": -0.0,
+      "num_tokens": 11693224.0,
+      "reward": -0.12399546802043915,
+      "reward_std": 0.15344232320785522,
+      "rewards/cosine_scaled_reward/mean": -0.12399546802043915,
+      "rewards/cosine_scaled_reward/std": 0.4378487467765808,
+      "step": 94
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.921875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1800.0,
+      "completions/mean_length": 1977.0625,
+      "completions/mean_terminated_length": 1140.0,
+      "completions/min_length": 755.0,
+      "completions/min_terminated_length": 755.0,
+      "epoch": 0.10857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21893343329429626,
+      "learning_rate": 9.78935800506826e-07,
+      "loss": 0.0,
+      "num_tokens": 11830284.0,
+      "reward": -0.2706957459449768,
+      "reward_std": 0.1604195535182953,
+      "rewards/cosine_scaled_reward/mean": -0.2706957459449768,
+      "rewards/cosine_scaled_reward/std": 0.17591074109077454,
+      "step": 95
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1688.0,
+      "completions/mean_length": 1656.359375,
+      "completions/mean_terminated_length": 1003.625,
+      "completions/min_length": 541.0,
+      "completions/min_terminated_length": 541.0,
+      "epoch": 0.10971428571428571,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24892951548099518,
+      "learning_rate": 9.779754323328192e-07,
+      "loss": 0.0,
+      "num_tokens": 11947427.0,
+      "reward": -0.05472355708479881,
+      "reward_std": 0.22797656059265137,
+      "rewards/cosine_scaled_reward/mean": -0.05472356453537941,
+      "rewards/cosine_scaled_reward/std": 0.4557226002216339,
+      "step": 96
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2021.0,
+      "completions/mean_length": 1808.265625,
+      "completions/mean_terminated_length": 1240.4736328125,
+      "completions/min_length": 601.0,
+      "completions/min_terminated_length": 601.0,
+      "epoch": 0.11085714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.26835066080093384,
+      "learning_rate": 9.769942052400235e-07,
+      "loss": 0.0,
+      "num_tokens": 12073420.0,
+      "reward": -0.10791188478469849,
+      "reward_std": 0.2891411781311035,
+      "rewards/cosine_scaled_reward/mean": -0.10791188478469849,
+      "rewards/cosine_scaled_reward/std": 0.3751998543739319,
+      "step": 97
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.5625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1930.0,
+      "completions/mean_length": 1552.390625,
+      "completions/mean_terminated_length": 915.1785888671875,
+      "completions/min_length": 402.0,
+      "completions/min_terminated_length": 402.0,
+      "epoch": 0.112,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.27451470494270325,
+      "learning_rate": 9.759921670520634e-07,
+      "loss": -0.0,
+      "num_tokens": 12183837.0,
+      "reward": -0.1808183193206787,
+      "reward_std": 0.24214914441108704,
+      "rewards/cosine_scaled_reward/mean": -0.1808183193206787,
+      "rewards/cosine_scaled_reward/std": 0.3102630078792572,
+      "step": 98
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1318.0,
+      "completions/mean_length": 1721.15625,
+      "completions/mean_terminated_length": 653.4666748046875,
+      "completions/min_length": 298.0,
+      "completions/min_terminated_length": 298.0,
+      "epoch": 0.11314285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2736988067626953,
+      "learning_rate": 9.749693666068663e-07,
+      "loss": 0.0,
+      "num_tokens": 12305159.0,
+      "reward": -0.10280460864305496,
+      "reward_std": 0.21398380398750305,
+      "rewards/cosine_scaled_reward/mean": -0.10280461609363556,
+      "rewards/cosine_scaled_reward/std": 0.4072605073451996,
+      "step": 99
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.546875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1780.0,
+      "completions/mean_length": 1564.6875,
+      "completions/mean_terminated_length": 981.3793334960938,
+      "completions/min_length": 498.0,
+      "completions/min_terminated_length": 498.0,
+      "epoch": 0.11428571428571428,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.26440170407295227,
+      "learning_rate": 9.739258537542835e-07,
+      "loss": 0.0,
+      "num_tokens": 12415011.0,
+      "reward": 0.07170121371746063,
+      "reward_std": 0.38168632984161377,
+      "rewards/cosine_scaled_reward/mean": 0.07170121371746063,
+      "rewards/cosine_scaled_reward/std": 0.519091784954071,
+      "step": 100
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 500,
+  "num_input_tokens_seen": 12415011,
+  "num_train_epochs": 1,
+  "save_steps": 50,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 0.0,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-100/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b847efd20062b0cb1aebb8c261e138e047198d5f1277135c040d30f434a732d
+size 8824

checkpoint-100/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)