brthor commited on
Commit
11fff9b
·
verified ·
1 Parent(s): 105b56f

Upload full trainer checkpoint step=4052 (profile=llama3_2_1b_base_grpo, run_type=grpo)

Browse files
.gitattributes CHANGED
@@ -130,3 +130,4 @@ meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-14196/tokenizer.json filter
130
  meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-1013/tokenizer.json filter=lfs diff=lfs merge=lfs -text
131
  meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-2026/tokenizer.json filter=lfs diff=lfs merge=lfs -text
132
  meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-3039/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 
 
130
  meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-1013/tokenizer.json filter=lfs diff=lfs merge=lfs -text
131
  meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-2026/tokenizer.json filter=lfs diff=lfs merge=lfs -text
132
  meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-3039/tokenizer.json filter=lfs diff=lfs merge=lfs -text
133
+ meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/tokenizer.json filter=lfs diff=lfs merge=lfs -text
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/README.md ADDED
@@ -0,0 +1,210 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: unsloth/Llama-3.2-1B
3
+ library_name: peft
4
+ pipeline_tag: text-generation
5
+ tags:
6
+ - base_model:adapter:unsloth/Llama-3.2-1B
7
+ - grpo
8
+ - lora
9
+ - transformers
10
+ - trl
11
+ - unsloth
12
+ ---
13
+
14
+ # Model Card for Model ID
15
+
16
+ <!-- Provide a quick summary of what the model is/does. -->
17
+
18
+
19
+
20
+ ## Model Details
21
+
22
+ ### Model Description
23
+
24
+ <!-- Provide a longer summary of what this model is. -->
25
+
26
+
27
+
28
+ - **Developed by:** [More Information Needed]
29
+ - **Funded by [optional]:** [More Information Needed]
30
+ - **Shared by [optional]:** [More Information Needed]
31
+ - **Model type:** [More Information Needed]
32
+ - **Language(s) (NLP):** [More Information Needed]
33
+ - **License:** [More Information Needed]
34
+ - **Finetuned from model [optional]:** [More Information Needed]
35
+
36
+ ### Model Sources [optional]
37
+
38
+ <!-- Provide the basic links for the model. -->
39
+
40
+ - **Repository:** [More Information Needed]
41
+ - **Paper [optional]:** [More Information Needed]
42
+ - **Demo [optional]:** [More Information Needed]
43
+
44
+ ## Uses
45
+
46
+ <!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
47
+
48
+ ### Direct Use
49
+
50
+ <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
51
+
52
+ [More Information Needed]
53
+
54
+ ### Downstream Use [optional]
55
+
56
+ <!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
57
+
58
+ [More Information Needed]
59
+
60
+ ### Out-of-Scope Use
61
+
62
+ <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
63
+
64
+ [More Information Needed]
65
+
66
+ ## Bias, Risks, and Limitations
67
+
68
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
69
+
70
+ [More Information Needed]
71
+
72
+ ### Recommendations
73
+
74
+ <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
75
+
76
+ Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
77
+
78
+ ## How to Get Started with the Model
79
+
80
+ Use the code below to get started with the model.
81
+
82
+ [More Information Needed]
83
+
84
+ ## Training Details
85
+
86
+ ### Training Data
87
+
88
+ <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
89
+
90
+ [More Information Needed]
91
+
92
+ ### Training Procedure
93
+
94
+ <!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
95
+
96
+ #### Preprocessing [optional]
97
+
98
+ [More Information Needed]
99
+
100
+
101
+ #### Training Hyperparameters
102
+
103
+ - **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
104
+
105
+ #### Speeds, Sizes, Times [optional]
106
+
107
+ <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
108
+
109
+ [More Information Needed]
110
+
111
+ ## Evaluation
112
+
113
+ <!-- This section describes the evaluation protocols and provides the results. -->
114
+
115
+ ### Testing Data, Factors & Metrics
116
+
117
+ #### Testing Data
118
+
119
+ <!-- This should link to a Dataset Card if possible. -->
120
+
121
+ [More Information Needed]
122
+
123
+ #### Factors
124
+
125
+ <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
126
+
127
+ [More Information Needed]
128
+
129
+ #### Metrics
130
+
131
+ <!-- These are the evaluation metrics being used, ideally with a description of why. -->
132
+
133
+ [More Information Needed]
134
+
135
+ ### Results
136
+
137
+ [More Information Needed]
138
+
139
+ #### Summary
140
+
141
+
142
+
143
+ ## Model Examination [optional]
144
+
145
+ <!-- Relevant interpretability work for the model goes here -->
146
+
147
+ [More Information Needed]
148
+
149
+ ## Environmental Impact
150
+
151
+ <!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
152
+
153
+ Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
154
+
155
+ - **Hardware Type:** [More Information Needed]
156
+ - **Hours used:** [More Information Needed]
157
+ - **Cloud Provider:** [More Information Needed]
158
+ - **Compute Region:** [More Information Needed]
159
+ - **Carbon Emitted:** [More Information Needed]
160
+
161
+ ## Technical Specifications [optional]
162
+
163
+ ### Model Architecture and Objective
164
+
165
+ [More Information Needed]
166
+
167
+ ### Compute Infrastructure
168
+
169
+ [More Information Needed]
170
+
171
+ #### Hardware
172
+
173
+ [More Information Needed]
174
+
175
+ #### Software
176
+
177
+ [More Information Needed]
178
+
179
+ ## Citation [optional]
180
+
181
+ <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
182
+
183
+ **BibTeX:**
184
+
185
+ [More Information Needed]
186
+
187
+ **APA:**
188
+
189
+ [More Information Needed]
190
+
191
+ ## Glossary [optional]
192
+
193
+ <!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
194
+
195
+ [More Information Needed]
196
+
197
+ ## More Information [optional]
198
+
199
+ [More Information Needed]
200
+
201
+ ## Model Card Authors [optional]
202
+
203
+ [More Information Needed]
204
+
205
+ ## Model Card Contact
206
+
207
+ [More Information Needed]
208
+ ### Framework versions
209
+
210
+ - PEFT 0.18.1
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/adapter_config.json ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": {
6
+ "base_model_class": "LlamaForCausalLM",
7
+ "parent_library": "transformers.models.llama.modeling_llama",
8
+ "unsloth_fixed": true
9
+ },
10
+ "base_model_name_or_path": "unsloth/Llama-3.2-1B",
11
+ "bias": "none",
12
+ "corda_config": null,
13
+ "ensure_weight_tying": false,
14
+ "eva_config": null,
15
+ "exclude_modules": null,
16
+ "fan_in_fan_out": false,
17
+ "inference_mode": true,
18
+ "init_lora_weights": true,
19
+ "layer_replication": null,
20
+ "layers_pattern": null,
21
+ "layers_to_transform": null,
22
+ "loftq_config": {},
23
+ "lora_alpha": 16,
24
+ "lora_bias": false,
25
+ "lora_dropout": 0.0,
26
+ "megatron_config": null,
27
+ "megatron_core": "megatron.core",
28
+ "modules_to_save": null,
29
+ "peft_type": "LORA",
30
+ "peft_version": "0.18.1",
31
+ "qalora_group_size": 16,
32
+ "r": 16,
33
+ "rank_pattern": {},
34
+ "revision": null,
35
+ "target_modules": [
36
+ "down_proj",
37
+ "q_proj",
38
+ "up_proj",
39
+ "gate_proj",
40
+ "o_proj",
41
+ "k_proj",
42
+ "v_proj"
43
+ ],
44
+ "target_parameters": null,
45
+ "task_type": "CAUSAL_LM",
46
+ "trainable_token_indices": null,
47
+ "use_dora": false,
48
+ "use_qalora": false,
49
+ "use_rslora": false
50
+ }
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a2849dea97bed69e994584ea2751c409c0921717c747841c950fc494b53b711f
3
+ size 45118424
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/chat_template.jinja ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ messages[0]['content'] + '
2
+
3
+ ' }}{% set loop_messages = messages[1:] %}{% else %}{{ '' + '
4
+
5
+ ' }}{% set loop_messages = messages %}{% endif %}{% for message in loop_messages %}{% if message['role'] == 'user' %}{{ '### Instruction:
6
+ ' + message['content'] + '
7
+
8
+ ' }}{% elif message['role'] == 'assistant' %}{{ '### Response:
9
+ ' + message['content'] + eos_token + '
10
+
11
+ ' }}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '### Response:
12
+ ' }}{% endif %}
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3ea7a6d4b32ab68f0bc370504eb4487dcf2e67f2bb7accfb042077f8c3f36bb4
3
+ size 23162187
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/rng_state.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:92f155cfa537a8b11137016754f83954b318aaddc149deb01d8f510cee0e19f3
3
+ size 14645
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3215b21f192c8251ba270bdce153a16e749942f9794c9bcb315608fb08702a7a
3
+ size 1465
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a65c6c5f9764771aa485e6a1f5e63d7d9af8477fe0777148c17476ecb2e09a05
3
+ size 17210099
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/tokenizer_config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "bos_token": "<|begin_of_text|>",
4
+ "clean_up_tokenization_spaces": true,
5
+ "eos_token": "<|end_of_text|>",
6
+ "from_slow": true,
7
+ "is_local": false,
8
+ "legacy": false,
9
+ "model_input_names": [
10
+ "input_ids",
11
+ "attention_mask"
12
+ ],
13
+ "model_max_length": 131072,
14
+ "pad_token": "<|finetune_right_pad_id|>",
15
+ "padding_side": "right",
16
+ "tokenizer_class": "TokenizersBackend",
17
+ "unk_token": null
18
+ }
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/trainer_state.json ADDED
@@ -0,0 +1,2257 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": null,
3
+ "best_metric": null,
4
+ "best_model_checkpoint": null,
5
+ "epoch": 0.007197030248130584,
6
+ "eval_steps": 500,
7
+ "global_step": 4052,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "clip_ratio/high_max": 0.0,
14
+ "clip_ratio/high_mean": 0.0,
15
+ "clip_ratio/low_mean": 0.0,
16
+ "clip_ratio/low_min": 0.0,
17
+ "clip_ratio/region_mean": 0.0,
18
+ "completion_length": 65.68,
19
+ "completions/clipped_ratio": 0.0,
20
+ "completions/max_length": 65.88,
21
+ "completions/max_terminated_length": 65.88,
22
+ "completions/mean_length": 65.68,
23
+ "completions/mean_terminated_length": 65.68,
24
+ "completions/min_length": 65.48,
25
+ "completions/min_terminated_length": 65.48,
26
+ "epoch": 8.88083693007229e-05,
27
+ "frac_reward_zero_std": 0.56,
28
+ "grad_norm": 0.5864518284797668,
29
+ "kl": 0.2340605702996254,
30
+ "learning_rate": 5.439488466064254e-09,
31
+ "loss": 0.0004230526089668274,
32
+ "num_tokens": 46552.0,
33
+ "reward": 3.2782554626464844e-09,
34
+ "reward_std": 0.47037978172302247,
35
+ "rewards/TranscriptCorrectionGrpoReward/mean": 3.2782554626464844e-09,
36
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4703797769546509,
37
+ "step": 50
38
+ },
39
+ {
40
+ "clip_ratio/high_max": 0.0,
41
+ "clip_ratio/high_mean": 0.0,
42
+ "clip_ratio/low_mean": 0.0,
43
+ "clip_ratio/low_min": 0.0,
44
+ "clip_ratio/region_mean": 0.0,
45
+ "completion_length": 53.46,
46
+ "completions/clipped_ratio": 0.0,
47
+ "completions/max_length": 53.6,
48
+ "completions/max_terminated_length": 53.6,
49
+ "completions/mean_length": 53.46,
50
+ "completions/mean_terminated_length": 53.46,
51
+ "completions/min_length": 53.24,
52
+ "completions/min_terminated_length": 53.24,
53
+ "epoch": 0.0001776167386014458,
54
+ "frac_reward_zero_std": 0.68,
55
+ "grad_norm": 7.498858030885458e-05,
56
+ "kl": 0.293030250146985,
57
+ "learning_rate": 1.0989986900823695e-08,
58
+ "loss": 0.0005571873486042023,
59
+ "num_tokens": 85556.0,
60
+ "reward": 4.76837158203125e-09,
61
+ "reward_std": 0.3420943832397461,
62
+ "rewards/TranscriptCorrectionGrpoReward/mean": 4.76837158203125e-09,
63
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
64
+ "step": 100
65
+ },
66
+ {
67
+ "clip_ratio/high_max": 0.0,
68
+ "clip_ratio/high_mean": 0.0,
69
+ "clip_ratio/low_mean": 0.0002914614137262106,
70
+ "clip_ratio/low_min": 0.0002914614137262106,
71
+ "clip_ratio/region_mean": 0.0002914614137262106,
72
+ "completion_length": 62.48,
73
+ "completions/clipped_ratio": 0.0,
74
+ "completions/max_length": 62.6,
75
+ "completions/max_terminated_length": 62.6,
76
+ "completions/mean_length": 62.48,
77
+ "completions/mean_terminated_length": 62.48,
78
+ "completions/min_length": 62.24,
79
+ "completions/min_terminated_length": 62.24,
80
+ "epoch": 0.0002664251079021687,
81
+ "frac_reward_zero_std": 0.56,
82
+ "grad_norm": 0.2601912319660187,
83
+ "kl": 0.2793836747482419,
84
+ "learning_rate": 1.6540485335583138e-08,
85
+ "loss": 0.0005555346608161927,
86
+ "num_tokens": 130380.0,
87
+ "reward": 1.7881393432617187e-09,
88
+ "reward_std": 0.4703797769546509,
89
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
90
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.47037978172302247,
91
+ "step": 150
92
+ },
93
+ {
94
+ "clip_ratio/high_max": 0.0,
95
+ "clip_ratio/high_mean": 0.0,
96
+ "clip_ratio/low_mean": 0.0,
97
+ "clip_ratio/low_min": 0.0,
98
+ "clip_ratio/region_mean": 0.0,
99
+ "completion_length": 62.165,
100
+ "completions/clipped_ratio": 0.0,
101
+ "completions/max_length": 62.4,
102
+ "completions/max_terminated_length": 62.4,
103
+ "completions/mean_length": 62.165,
104
+ "completions/mean_terminated_length": 62.165,
105
+ "completions/min_length": 61.84,
106
+ "completions/min_terminated_length": 61.84,
107
+ "epoch": 0.0003552334772028916,
108
+ "frac_reward_zero_std": 0.48,
109
+ "grad_norm": 0.39951568841934204,
110
+ "kl": 0.2763713626563549,
111
+ "learning_rate": 2.2090983770342578e-08,
112
+ "loss": 0.0009129984676837921,
113
+ "num_tokens": 174957.0,
114
+ "reward": 8.940696716308593e-10,
115
+ "reward_std": 0.5559033823013305,
116
+ "rewards/TranscriptCorrectionGrpoReward/mean": 8.940696716308593e-10,
117
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5559033775329589,
118
+ "step": 200
119
+ },
120
+ {
121
+ "clip_ratio/high_max": 0.0,
122
+ "clip_ratio/high_mean": 0.0,
123
+ "clip_ratio/low_mean": 0.0,
124
+ "clip_ratio/low_min": 0.0,
125
+ "clip_ratio/region_mean": 0.0,
126
+ "completion_length": 65.825,
127
+ "completions/clipped_ratio": 0.0,
128
+ "completions/max_length": 66.04,
129
+ "completions/max_terminated_length": 66.04,
130
+ "completions/mean_length": 65.825,
131
+ "completions/mean_terminated_length": 65.825,
132
+ "completions/min_length": 65.56,
133
+ "completions/min_terminated_length": 65.56,
134
+ "epoch": 0.0004440418465036145,
135
+ "frac_reward_zero_std": 0.52,
136
+ "grad_norm": 5.004892227589153e-05,
137
+ "kl": 0.2560868472605944,
138
+ "learning_rate": 2.764148220510202e-08,
139
+ "loss": 0.001215285211801529,
140
+ "num_tokens": 222266.0,
141
+ "reward": 4.470348358154297e-09,
142
+ "reward_std": 0.5131415796279907,
143
+ "rewards/TranscriptCorrectionGrpoReward/mean": 4.470348358154297e-09,
144
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5131415748596191,
145
+ "step": 250
146
+ },
147
+ {
148
+ "clip_ratio/high_max": 0.0,
149
+ "clip_ratio/high_mean": 0.0,
150
+ "clip_ratio/low_mean": 0.0,
151
+ "clip_ratio/low_min": 0.0,
152
+ "clip_ratio/region_mean": 0.0,
153
+ "completion_length": 68.75,
154
+ "completions/clipped_ratio": 0.0,
155
+ "completions/max_length": 69.04,
156
+ "completions/max_terminated_length": 69.04,
157
+ "completions/mean_length": 68.75,
158
+ "completions/mean_terminated_length": 68.75,
159
+ "completions/min_length": 68.56,
160
+ "completions/min_terminated_length": 68.56,
161
+ "epoch": 0.0005328502158043374,
162
+ "frac_reward_zero_std": 0.6,
163
+ "grad_norm": 1.926846834976459e-06,
164
+ "kl": 0.239115035161376,
165
+ "learning_rate": 3.3191980639861464e-08,
166
+ "loss": 0.001264408677816391,
167
+ "num_tokens": 271736.0,
168
+ "reward": 6.258487701416015e-09,
169
+ "reward_std": 0.4276179838180542,
170
+ "rewards/TranscriptCorrectionGrpoReward/mean": 6.258487701416015e-09,
171
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
172
+ "step": 300
173
+ },
174
+ {
175
+ "clip_ratio/high_max": 4.8543689772486686e-05,
176
+ "clip_ratio/high_mean": 4.8543689772486686e-05,
177
+ "clip_ratio/low_mean": 0.0,
178
+ "clip_ratio/low_min": 0.0,
179
+ "clip_ratio/region_mean": 4.8543689772486686e-05,
180
+ "completion_length": 61.07,
181
+ "completions/clipped_ratio": 0.0,
182
+ "completions/max_length": 61.2,
183
+ "completions/max_terminated_length": 61.2,
184
+ "completions/mean_length": 61.07,
185
+ "completions/mean_terminated_length": 61.07,
186
+ "completions/min_length": 60.96,
187
+ "completions/min_terminated_length": 60.96,
188
+ "epoch": 0.0006216585851050603,
189
+ "frac_reward_zero_std": 0.72,
190
+ "grad_norm": 2.6762343168229563e-06,
191
+ "kl": 0.2665533754974604,
192
+ "learning_rate": 3.8742479074620904e-08,
193
+ "loss": 0.0013578705489635467,
194
+ "num_tokens": 316350.0,
195
+ "reward": 2.384185791015625e-09,
196
+ "reward_std": 0.29933258533477786,
197
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.384185791015625e-09,
198
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.2993325901031494,
199
+ "step": 350
200
+ },
201
+ {
202
+ "clip_ratio/high_max": 0.0,
203
+ "clip_ratio/high_mean": 0.0,
204
+ "clip_ratio/low_mean": 0.0,
205
+ "clip_ratio/low_min": 0.0,
206
+ "clip_ratio/region_mean": 0.0,
207
+ "completion_length": 64.22,
208
+ "completions/clipped_ratio": 0.0,
209
+ "completions/max_length": 64.6,
210
+ "completions/max_terminated_length": 64.6,
211
+ "completions/mean_length": 64.22,
212
+ "completions/mean_terminated_length": 64.22,
213
+ "completions/min_length": 63.92,
214
+ "completions/min_terminated_length": 63.92,
215
+ "epoch": 0.0007104669544057832,
216
+ "frac_reward_zero_std": 0.6,
217
+ "grad_norm": 1.1867748526128707e-06,
218
+ "kl": 0.2582776914909482,
219
+ "learning_rate": 4.429297750938035e-08,
220
+ "loss": 0.000348069965839386,
221
+ "num_tokens": 362546.0,
222
+ "reward": 0.0,
223
+ "reward_std": 0.42761797904968263,
224
+ "rewards/TranscriptCorrectionGrpoReward/mean": 0.0,
225
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
226
+ "step": 400
227
+ },
228
+ {
229
+ "clip_ratio/high_max": 0.0,
230
+ "clip_ratio/high_mean": 0.0,
231
+ "clip_ratio/low_mean": 0.0,
232
+ "clip_ratio/low_min": 0.0,
233
+ "clip_ratio/region_mean": 0.0,
234
+ "completion_length": 65.25,
235
+ "completions/clipped_ratio": 0.0,
236
+ "completions/max_length": 65.36,
237
+ "completions/max_terminated_length": 65.36,
238
+ "completions/mean_length": 65.25,
239
+ "completions/mean_terminated_length": 65.25,
240
+ "completions/min_length": 65.04,
241
+ "completions/min_terminated_length": 65.04,
242
+ "epoch": 0.000799275323706506,
243
+ "frac_reward_zero_std": 0.6,
244
+ "grad_norm": 0.0002190974773839116,
245
+ "kl": 0.24411882523447276,
246
+ "learning_rate": 4.984347594413978e-08,
247
+ "loss": 0.0009650686383247376,
248
+ "num_tokens": 409412.0,
249
+ "reward": 1.7881393432617187e-09,
250
+ "reward_std": 0.4276179838180542,
251
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
252
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4276179838180542,
253
+ "step": 450
254
+ },
255
+ {
256
+ "clip_ratio/high_max": 0.0,
257
+ "clip_ratio/high_mean": 0.0,
258
+ "clip_ratio/low_mean": 0.0,
259
+ "clip_ratio/low_min": 0.0,
260
+ "clip_ratio/region_mean": 0.0,
261
+ "completion_length": 65.765,
262
+ "completions/clipped_ratio": 0.0,
263
+ "completions/max_length": 65.92,
264
+ "completions/max_terminated_length": 65.92,
265
+ "completions/mean_length": 65.765,
266
+ "completions/mean_terminated_length": 65.765,
267
+ "completions/min_length": 65.72,
268
+ "completions/min_terminated_length": 65.72,
269
+ "epoch": 0.000888083693007229,
270
+ "frac_reward_zero_std": 0.88,
271
+ "grad_norm": 0.03857799246907234,
272
+ "kl": 0.23137393046170474,
273
+ "learning_rate": 5.5393974378899236e-08,
274
+ "loss": 0.00039650365710258484,
275
+ "num_tokens": 456021.0,
276
+ "reward": -5.960464477539063e-10,
277
+ "reward_std": 0.1282853889465332,
278
+ "rewards/TranscriptCorrectionGrpoReward/mean": -5.960464477539063e-10,
279
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.12828539371490477,
280
+ "step": 500
281
+ },
282
+ {
283
+ "clip_ratio/high_max": 0.0,
284
+ "clip_ratio/high_mean": 0.0,
285
+ "clip_ratio/low_mean": 0.0,
286
+ "clip_ratio/low_min": 0.0,
287
+ "clip_ratio/region_mean": 0.0,
288
+ "completion_length": 66.25,
289
+ "completions/clipped_ratio": 0.0,
290
+ "completions/max_length": 66.64,
291
+ "completions/max_terminated_length": 66.64,
292
+ "completions/mean_length": 66.25,
293
+ "completions/mean_terminated_length": 66.25,
294
+ "completions/min_length": 66.0,
295
+ "completions/min_terminated_length": 66.0,
296
+ "epoch": 0.0009768920623079518,
297
+ "frac_reward_zero_std": 0.68,
298
+ "grad_norm": 1.3617865988635458e-05,
299
+ "kl": 0.2465892593190074,
300
+ "learning_rate": 6.094447281365867e-08,
301
+ "loss": 0.0013641029596328734,
302
+ "num_tokens": 503519.0,
303
+ "reward": 0.0,
304
+ "reward_std": 0.3420943832397461,
305
+ "rewards/TranscriptCorrectionGrpoReward/mean": 0.0,
306
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
307
+ "step": 550
308
+ },
309
+ {
310
+ "clip_ratio/high_max": 0.0,
311
+ "clip_ratio/high_mean": 0.0,
312
+ "clip_ratio/low_mean": 8.620689623057842e-05,
313
+ "clip_ratio/low_min": 8.620689623057842e-05,
314
+ "clip_ratio/region_mean": 8.620689623057842e-05,
315
+ "completion_length": 70.03,
316
+ "completions/clipped_ratio": 0.0,
317
+ "completions/max_length": 70.16,
318
+ "completions/max_terminated_length": 70.16,
319
+ "completions/mean_length": 70.03,
320
+ "completions/mean_terminated_length": 70.03,
321
+ "completions/min_length": 69.8,
322
+ "completions/min_terminated_length": 69.8,
323
+ "epoch": 0.0010657004316086747,
324
+ "frac_reward_zero_std": 0.52,
325
+ "grad_norm": 0.00015144512872211635,
326
+ "kl": 0.22738730356097223,
327
+ "learning_rate": 6.649497124841812e-08,
328
+ "loss": 0.0008844786882400513,
329
+ "num_tokens": 553309.0,
330
+ "reward": 3.2782554626464844e-09,
331
+ "reward_std": 0.5131415700912476,
332
+ "rewards/TranscriptCorrectionGrpoReward/mean": 3.2782554626464844e-09,
333
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5131415748596191,
334
+ "step": 600
335
+ },
336
+ {
337
+ "clip_ratio/high_max": 0.0,
338
+ "clip_ratio/high_mean": 0.0,
339
+ "clip_ratio/low_mean": 0.0,
340
+ "clip_ratio/low_min": 0.0,
341
+ "clip_ratio/region_mean": 0.0,
342
+ "completion_length": 71.28,
343
+ "completions/clipped_ratio": 0.0,
344
+ "completions/max_length": 71.76,
345
+ "completions/max_terminated_length": 71.76,
346
+ "completions/mean_length": 71.28,
347
+ "completions/mean_terminated_length": 71.28,
348
+ "completions/min_length": 70.52,
349
+ "completions/min_terminated_length": 70.52,
350
+ "epoch": 0.0011545088009093977,
351
+ "frac_reward_zero_std": 0.44,
352
+ "grad_norm": 8.714160685485695e-06,
353
+ "kl": 0.23357330340892077,
354
+ "learning_rate": 7.204546968317756e-08,
355
+ "loss": -0.00036261290311813356,
356
+ "num_tokens": 603181.0,
357
+ "reward": -8.940696716308593e-10,
358
+ "reward_std": 0.5986651754379273,
359
+ "rewards/TranscriptCorrectionGrpoReward/mean": -8.940696716308593e-10,
360
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5986651754379273,
361
+ "step": 650
362
+ },
363
+ {
364
+ "clip_ratio/high_max": 0.0,
365
+ "clip_ratio/high_mean": 0.0,
366
+ "clip_ratio/low_mean": 0.0,
367
+ "clip_ratio/low_min": 0.0,
368
+ "clip_ratio/region_mean": 0.0,
369
+ "completion_length": 71.675,
370
+ "completions/clipped_ratio": 0.0,
371
+ "completions/max_length": 71.88,
372
+ "completions/max_terminated_length": 71.88,
373
+ "completions/mean_length": 71.675,
374
+ "completions/mean_terminated_length": 71.675,
375
+ "completions/min_length": 71.52,
376
+ "completions/min_terminated_length": 71.52,
377
+ "epoch": 0.0012433171702101206,
378
+ "frac_reward_zero_std": 0.6,
379
+ "grad_norm": 0.1935173124074936,
380
+ "kl": 0.22929298244416713,
381
+ "learning_rate": 7.7595968117937e-08,
382
+ "loss": 0.0007024538516998291,
383
+ "num_tokens": 654188.0,
384
+ "reward": 2.9802322387695314e-09,
385
+ "reward_std": 0.4276179885864258,
386
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.9802322387695314e-09,
387
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
388
+ "step": 700
389
+ },
390
+ {
391
+ "clip_ratio/high_max": 0.0,
392
+ "clip_ratio/high_mean": 0.0,
393
+ "clip_ratio/low_mean": 0.0,
394
+ "clip_ratio/low_min": 0.0,
395
+ "clip_ratio/region_mean": 0.0,
396
+ "completion_length": 65.605,
397
+ "completions/clipped_ratio": 0.0,
398
+ "completions/max_length": 65.92,
399
+ "completions/max_terminated_length": 65.92,
400
+ "completions/mean_length": 65.605,
401
+ "completions/mean_terminated_length": 65.605,
402
+ "completions/min_length": 65.36,
403
+ "completions/min_terminated_length": 65.36,
404
+ "epoch": 0.0013321255395108435,
405
+ "frac_reward_zero_std": 0.64,
406
+ "grad_norm": 0.41991785168647766,
407
+ "kl": 0.23588136799633502,
408
+ "learning_rate": 8.314646655269643e-08,
409
+ "loss": 0.0017469561100006104,
410
+ "num_tokens": 700877.0,
411
+ "reward": 5.960464477539063e-10,
412
+ "reward_std": 0.3848561763763428,
413
+ "rewards/TranscriptCorrectionGrpoReward/mean": 5.960464477539063e-10,
414
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.38485618591308596,
415
+ "step": 750
416
+ },
417
+ {
418
+ "clip_ratio/high_max": 0.0,
419
+ "clip_ratio/high_mean": 0.0,
420
+ "clip_ratio/low_mean": 0.0,
421
+ "clip_ratio/low_min": 0.0,
422
+ "clip_ratio/region_mean": 0.0,
423
+ "completion_length": 67.435,
424
+ "completions/clipped_ratio": 0.0,
425
+ "completions/max_length": 67.6,
426
+ "completions/max_terminated_length": 67.6,
427
+ "completions/mean_length": 67.435,
428
+ "completions/mean_terminated_length": 67.435,
429
+ "completions/min_length": 67.28,
430
+ "completions/min_terminated_length": 67.28,
431
+ "epoch": 0.0014209339088115665,
432
+ "frac_reward_zero_std": 0.48,
433
+ "grad_norm": 0.31073635816574097,
434
+ "kl": 0.2509309166669846,
435
+ "learning_rate": 8.869696498745589e-08,
436
+ "loss": 0.0010700391232967377,
437
+ "num_tokens": 749028.0,
438
+ "reward": 6.556510925292969e-09,
439
+ "reward_std": 0.5559033823013305,
440
+ "rewards/TranscriptCorrectionGrpoReward/mean": 6.556510925292969e-09,
441
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5559033727645875,
442
+ "step": 800
443
+ },
444
+ {
445
+ "clip_ratio/high_max": 0.0,
446
+ "clip_ratio/high_mean": 0.0,
447
+ "clip_ratio/low_mean": 0.00014605543576180934,
448
+ "clip_ratio/low_min": 0.00014605543576180934,
449
+ "clip_ratio/region_mean": 0.00014605543576180934,
450
+ "completion_length": 75.25,
451
+ "completions/clipped_ratio": 0.0,
452
+ "completions/max_length": 75.36,
453
+ "completions/max_terminated_length": 75.36,
454
+ "completions/mean_length": 75.25,
455
+ "completions/mean_terminated_length": 75.25,
456
+ "completions/min_length": 75.08,
457
+ "completions/min_terminated_length": 75.08,
458
+ "epoch": 0.0015097422781122894,
459
+ "frac_reward_zero_std": 0.44,
460
+ "grad_norm": 0.04336247965693474,
461
+ "kl": 0.21717634629458188,
462
+ "learning_rate": 9.424746342221532e-08,
463
+ "loss": 0.0009317247569561004,
464
+ "num_tokens": 802606.0,
465
+ "reward": 4.172325134277344e-09,
466
+ "reward_std": 0.5986651706695557,
467
+ "rewards/TranscriptCorrectionGrpoReward/mean": 4.172325134277344e-09,
468
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5986651706695557,
469
+ "step": 850
470
+ },
471
+ {
472
+ "clip_ratio/high_max": 0.0,
473
+ "clip_ratio/high_mean": 0.0,
474
+ "clip_ratio/low_mean": 0.0,
475
+ "clip_ratio/low_min": 0.0,
476
+ "clip_ratio/region_mean": 0.0,
477
+ "completion_length": 74.23,
478
+ "completions/clipped_ratio": 0.0,
479
+ "completions/max_length": 74.84,
480
+ "completions/max_terminated_length": 74.84,
481
+ "completions/mean_length": 74.23,
482
+ "completions/mean_terminated_length": 74.23,
483
+ "completions/min_length": 74.0,
484
+ "completions/min_terminated_length": 74.0,
485
+ "epoch": 0.001598550647413012,
486
+ "frac_reward_zero_std": 0.52,
487
+ "grad_norm": 0.4461303651332855,
488
+ "kl": 0.23601274881511927,
489
+ "learning_rate": 9.979796185697477e-08,
490
+ "loss": 0.0018890520930290222,
491
+ "num_tokens": 855396.0,
492
+ "reward": 0.0,
493
+ "reward_std": 0.5131415748596191,
494
+ "rewards/TranscriptCorrectionGrpoReward/mean": 0.0,
495
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5131415796279907,
496
+ "step": 900
497
+ },
498
+ {
499
+ "clip_ratio/high_max": 0.0,
500
+ "clip_ratio/high_mean": 0.0,
501
+ "clip_ratio/low_mean": 0.0,
502
+ "clip_ratio/low_min": 0.0,
503
+ "clip_ratio/region_mean": 0.0,
504
+ "completion_length": 69.48,
505
+ "completions/clipped_ratio": 0.0,
506
+ "completions/max_length": 69.64,
507
+ "completions/max_terminated_length": 69.64,
508
+ "completions/mean_length": 69.48,
509
+ "completions/mean_terminated_length": 69.48,
510
+ "completions/min_length": 69.32,
511
+ "completions/min_terminated_length": 69.32,
512
+ "epoch": 0.001687359016713735,
513
+ "frac_reward_zero_std": 0.6,
514
+ "grad_norm": 0.3977927565574646,
515
+ "kl": 0.2380372828245163,
516
+ "learning_rate": 1.053484602917342e-07,
517
+ "loss": 0.0015320682525634766,
518
+ "num_tokens": 905028.0,
519
+ "reward": 8.940696716308593e-10,
520
+ "reward_std": 0.42761797904968263,
521
+ "rewards/TranscriptCorrectionGrpoReward/mean": 8.940696716308593e-10,
522
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
523
+ "step": 950
524
+ },
525
+ {
526
+ "clip_ratio/high_max": 0.0,
527
+ "clip_ratio/high_mean": 0.0,
528
+ "clip_ratio/low_mean": 0.0,
529
+ "clip_ratio/low_min": 0.0,
530
+ "clip_ratio/region_mean": 0.0,
531
+ "completion_length": 61.315,
532
+ "completions/clipped_ratio": 0.0,
533
+ "completions/max_length": 61.6,
534
+ "completions/max_terminated_length": 61.6,
535
+ "completions/mean_length": 61.315,
536
+ "completions/mean_terminated_length": 61.315,
537
+ "completions/min_length": 60.96,
538
+ "completions/min_terminated_length": 60.96,
539
+ "epoch": 0.001776167386014458,
540
+ "frac_reward_zero_std": 0.68,
541
+ "grad_norm": 0.0018230041023343801,
542
+ "kl": 0.27308079957962034,
543
+ "learning_rate": 1.1089895872649365e-07,
544
+ "loss": -0.00010064475238323211,
545
+ "num_tokens": 949379.0,
546
+ "reward": -8.940696716308593e-10,
547
+ "reward_std": 0.3420943880081177,
548
+ "rewards/TranscriptCorrectionGrpoReward/mean": -8.940696716308593e-10,
549
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
550
+ "step": 1000
551
+ },
552
+ {
553
+ "eval_cer_subset": 0.01309296985685671,
554
+ "eval_cer_subset_edit_distance": 804,
555
+ "eval_cer_subset_groups": 250,
556
+ "eval_cer_subset_items": 250,
557
+ "eval_cer_subset_ref_chars": 61407,
558
+ "eval_loss": 0.03959455178967374,
559
+ "step": 1013
560
+ },
561
+ {
562
+ "clip_ratio/high_max": 0.0,
563
+ "clip_ratio/high_mean": 0.0,
564
+ "clip_ratio/low_mean": 0.0,
565
+ "clip_ratio/low_min": 0.0,
566
+ "clip_ratio/region_mean": 0.0,
567
+ "completion_length": 61.745,
568
+ "completions/clipped_ratio": 0.0,
569
+ "completions/max_length": 62.04,
570
+ "completions/max_terminated_length": 62.04,
571
+ "completions/mean_length": 61.745,
572
+ "completions/mean_terminated_length": 61.745,
573
+ "completions/min_length": 61.52,
574
+ "completions/min_terminated_length": 61.52,
575
+ "epoch": 0.001864975755315181,
576
+ "frac_reward_zero_std": 0.6,
577
+ "grad_norm": 2.53622838499723e-05,
578
+ "kl": 0.26631607685238123,
579
+ "learning_rate": 1.1644945716125309e-07,
580
+ "loss": 0.0013614676892757416,
581
+ "num_tokens": 993560.0,
582
+ "reward": 2.9802322387695313e-10,
583
+ "reward_std": 0.42761797904968263,
584
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.9802322387695313e-10,
585
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
586
+ "step": 1050
587
+ },
588
+ {
589
+ "clip_ratio/high_max": 0.0,
590
+ "clip_ratio/high_mean": 0.0,
591
+ "clip_ratio/low_mean": 0.0,
592
+ "clip_ratio/low_min": 0.0,
593
+ "clip_ratio/region_mean": 0.0,
594
+ "completion_length": 62.33,
595
+ "completions/clipped_ratio": 0.0,
596
+ "completions/max_length": 62.56,
597
+ "completions/max_terminated_length": 62.56,
598
+ "completions/mean_length": 62.33,
599
+ "completions/mean_terminated_length": 62.33,
600
+ "completions/min_length": 62.16,
601
+ "completions/min_terminated_length": 62.16,
602
+ "epoch": 0.0019537841246159036,
603
+ "frac_reward_zero_std": 0.6,
604
+ "grad_norm": 0.16987353563308716,
605
+ "kl": 0.25355677947402,
606
+ "learning_rate": 1.2199995559601254e-07,
607
+ "loss": 0.0013103175163269043,
608
+ "num_tokens": 1038210.0,
609
+ "reward": 1.7881393432617187e-09,
610
+ "reward_std": 0.42761797904968263,
611
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
612
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
613
+ "step": 1100
614
+ },
615
+ {
616
+ "clip_ratio/high_max": 0.0,
617
+ "clip_ratio/high_mean": 0.0,
618
+ "clip_ratio/low_mean": 0.0,
619
+ "clip_ratio/low_min": 0.0,
620
+ "clip_ratio/region_mean": 0.0,
621
+ "completion_length": 67.595,
622
+ "completions/clipped_ratio": 0.0,
623
+ "completions/max_length": 67.72,
624
+ "completions/max_terminated_length": 67.72,
625
+ "completions/mean_length": 67.595,
626
+ "completions/mean_terminated_length": 67.595,
627
+ "completions/min_length": 67.36,
628
+ "completions/min_terminated_length": 67.36,
629
+ "epoch": 0.0020425924939166265,
630
+ "frac_reward_zero_std": 0.56,
631
+ "grad_norm": 0.21912723779678345,
632
+ "kl": 0.2458942100405693,
633
+ "learning_rate": 1.2755045403077197e-07,
634
+ "loss": 0.0009975716471672057,
635
+ "num_tokens": 1086193.0,
636
+ "reward": 0.0,
637
+ "reward_std": 0.4703797769546509,
638
+ "rewards/TranscriptCorrectionGrpoReward/mean": 0.0,
639
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4703797769546509,
640
+ "step": 1150
641
+ },
642
+ {
643
+ "clip_ratio/high_max": 0.0,
644
+ "clip_ratio/high_mean": 0.0,
645
+ "clip_ratio/low_mean": 0.0,
646
+ "clip_ratio/low_min": 0.0,
647
+ "clip_ratio/region_mean": 0.0,
648
+ "completion_length": 67.81,
649
+ "completions/clipped_ratio": 0.0,
650
+ "completions/max_length": 67.96,
651
+ "completions/max_terminated_length": 67.96,
652
+ "completions/mean_length": 67.81,
653
+ "completions/mean_terminated_length": 67.81,
654
+ "completions/min_length": 67.68,
655
+ "completions/min_terminated_length": 67.68,
656
+ "epoch": 0.0021314008632173495,
657
+ "frac_reward_zero_std": 0.68,
658
+ "grad_norm": 0.00010790217493195087,
659
+ "kl": 0.24153089322149754,
660
+ "learning_rate": 1.331009524655314e-07,
661
+ "loss": 0.0017908763885498046,
662
+ "num_tokens": 1134467.0,
663
+ "reward": 3.2782554626464844e-09,
664
+ "reward_std": 0.3420943832397461,
665
+ "rewards/TranscriptCorrectionGrpoReward/mean": 3.2782554626464844e-09,
666
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
667
+ "step": 1200
668
+ },
669
+ {
670
+ "clip_ratio/high_max": 0.0,
671
+ "clip_ratio/high_mean": 0.0,
672
+ "clip_ratio/low_mean": 0.0,
673
+ "clip_ratio/low_min": 0.0,
674
+ "clip_ratio/region_mean": 0.0,
675
+ "completion_length": 68.42,
676
+ "completions/clipped_ratio": 0.0,
677
+ "completions/max_length": 68.8,
678
+ "completions/max_terminated_length": 68.8,
679
+ "completions/mean_length": 68.42,
680
+ "completions/mean_terminated_length": 68.42,
681
+ "completions/min_length": 68.0,
682
+ "completions/min_terminated_length": 68.0,
683
+ "epoch": 0.0022202092325180724,
684
+ "frac_reward_zero_std": 0.6,
685
+ "grad_norm": 0.2312355786561966,
686
+ "kl": 0.23632734417915344,
687
+ "learning_rate": 1.3865145090029086e-07,
688
+ "loss": 0.0006246243417263031,
689
+ "num_tokens": 1183007.0,
690
+ "reward": 0.0,
691
+ "reward_std": 0.4276179885864258,
692
+ "rewards/TranscriptCorrectionGrpoReward/mean": 0.0,
693
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
694
+ "step": 1250
695
+ },
696
+ {
697
+ "clip_ratio/high_max": 0.0,
698
+ "clip_ratio/high_mean": 0.0,
699
+ "clip_ratio/low_mean": 0.0,
700
+ "clip_ratio/low_min": 0.0,
701
+ "clip_ratio/region_mean": 0.0,
702
+ "completion_length": 64.655,
703
+ "completions/clipped_ratio": 0.0,
704
+ "completions/max_length": 64.8,
705
+ "completions/max_terminated_length": 64.8,
706
+ "completions/mean_length": 64.655,
707
+ "completions/mean_terminated_length": 64.655,
708
+ "completions/min_length": 64.48,
709
+ "completions/min_terminated_length": 64.48,
710
+ "epoch": 0.0023090176018187953,
711
+ "frac_reward_zero_std": 0.72,
712
+ "grad_norm": 1.9456435893516755e-06,
713
+ "kl": 0.2434770704805851,
714
+ "learning_rate": 1.442019493350503e-07,
715
+ "loss": 0.00031983185559511187,
716
+ "num_tokens": 1229834.0,
717
+ "reward": -2.9802322387695313e-10,
718
+ "reward_std": 0.29933258056640627,
719
+ "rewards/TranscriptCorrectionGrpoReward/mean": -2.9802322387695313e-10,
720
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.29933258533477786,
721
+ "step": 1300
722
+ },
723
+ {
724
+ "clip_ratio/high_max": 0.0,
725
+ "clip_ratio/high_mean": 0.0,
726
+ "clip_ratio/low_mean": 0.0,
727
+ "clip_ratio/low_min": 0.0,
728
+ "clip_ratio/region_mean": 0.0,
729
+ "completion_length": 62.805,
730
+ "completions/clipped_ratio": 0.0,
731
+ "completions/max_length": 63.04,
732
+ "completions/max_terminated_length": 63.04,
733
+ "completions/mean_length": 62.805,
734
+ "completions/mean_terminated_length": 62.805,
735
+ "completions/min_length": 62.68,
736
+ "completions/min_terminated_length": 62.68,
737
+ "epoch": 0.0023978259711195183,
738
+ "frac_reward_zero_std": 0.56,
739
+ "grad_norm": 7.737488340353593e-05,
740
+ "kl": 0.2773256845027208,
741
+ "learning_rate": 1.4975244776980973e-07,
742
+ "loss": 0.001342010796070099,
743
+ "num_tokens": 1274811.0,
744
+ "reward": -1.1920928955078125e-09,
745
+ "reward_std": 0.47037978172302247,
746
+ "rewards/TranscriptCorrectionGrpoReward/mean": -1.1920928955078125e-09,
747
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4703797769546509,
748
+ "step": 1350
749
+ },
750
+ {
751
+ "clip_ratio/high_max": 0.0,
752
+ "clip_ratio/high_mean": 0.0,
753
+ "clip_ratio/low_mean": 0.0002222222276031971,
754
+ "clip_ratio/low_min": 0.0002222222276031971,
755
+ "clip_ratio/region_mean": 0.0002222222276031971,
756
+ "completion_length": 62.545,
757
+ "completions/clipped_ratio": 0.0,
758
+ "completions/max_length": 62.64,
759
+ "completions/max_terminated_length": 62.64,
760
+ "completions/mean_length": 62.545,
761
+ "completions/mean_terminated_length": 62.545,
762
+ "completions/min_length": 62.36,
763
+ "completions/min_terminated_length": 62.36,
764
+ "epoch": 0.002486634340420241,
765
+ "frac_reward_zero_std": 0.56,
766
+ "grad_norm": 9.343001693196129e-06,
767
+ "kl": 0.2813856675475836,
768
+ "learning_rate": 1.553029462045692e-07,
769
+ "loss": 0.0014287029206752778,
770
+ "num_tokens": 1320056.0,
771
+ "reward": 3.5762786865234374e-09,
772
+ "reward_std": 0.4703797769546509,
773
+ "rewards/TranscriptCorrectionGrpoReward/mean": 3.5762786865234374e-09,
774
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4703797769546509,
775
+ "step": 1400
776
+ },
777
+ {
778
+ "clip_ratio/high_max": 0.0,
779
+ "clip_ratio/high_mean": 0.0,
780
+ "clip_ratio/low_mean": 0.0,
781
+ "clip_ratio/low_min": 0.0,
782
+ "clip_ratio/region_mean": 0.0,
783
+ "completion_length": 63.605,
784
+ "completions/clipped_ratio": 0.0,
785
+ "completions/max_length": 63.68,
786
+ "completions/max_terminated_length": 63.68,
787
+ "completions/mean_length": 63.605,
788
+ "completions/mean_terminated_length": 63.605,
789
+ "completions/min_length": 63.56,
790
+ "completions/min_terminated_length": 63.56,
791
+ "epoch": 0.002575442709720964,
792
+ "frac_reward_zero_std": 0.88,
793
+ "grad_norm": 8.816698937152978e-06,
794
+ "kl": 0.2244873683154583,
795
+ "learning_rate": 1.6085344463932862e-07,
796
+ "loss": 0.0012283537536859513,
797
+ "num_tokens": 1366185.0,
798
+ "reward": 1.1920928955078125e-09,
799
+ "reward_std": 0.12828539371490477,
800
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.1920928955078125e-09,
801
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.12828539371490477,
802
+ "step": 1450
803
+ },
804
+ {
805
+ "clip_ratio/high_max": 0.0,
806
+ "clip_ratio/high_mean": 0.0,
807
+ "clip_ratio/low_mean": 7.352941203862428e-05,
808
+ "clip_ratio/low_min": 7.352941203862428e-05,
809
+ "clip_ratio/region_mean": 7.352941203862428e-05,
810
+ "completion_length": 68.155,
811
+ "completions/clipped_ratio": 0.0,
812
+ "completions/max_length": 68.48,
813
+ "completions/max_terminated_length": 68.48,
814
+ "completions/mean_length": 68.155,
815
+ "completions/mean_terminated_length": 68.155,
816
+ "completions/min_length": 67.96,
817
+ "completions/min_terminated_length": 67.96,
818
+ "epoch": 0.002664251079021687,
819
+ "frac_reward_zero_std": 0.56,
820
+ "grad_norm": 0.28626617789268494,
821
+ "kl": 0.2416078744456172,
822
+ "learning_rate": 1.6640394307408808e-07,
823
+ "loss": 0.001232394576072693,
824
+ "num_tokens": 1415224.0,
825
+ "reward": 2.384185791015625e-09,
826
+ "reward_std": 0.4703797721862793,
827
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.384185791015625e-09,
828
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4703797769546509,
829
+ "step": 1500
830
+ },
831
+ {
832
+ "clip_ratio/high_max": 0.0,
833
+ "clip_ratio/high_mean": 0.0,
834
+ "clip_ratio/low_mean": 4.424778744578361e-05,
835
+ "clip_ratio/low_min": 4.424778744578361e-05,
836
+ "clip_ratio/region_mean": 4.424778744578361e-05,
837
+ "completion_length": 70.45,
838
+ "completions/clipped_ratio": 0.0,
839
+ "completions/max_length": 70.68,
840
+ "completions/max_terminated_length": 70.68,
841
+ "completions/mean_length": 70.45,
842
+ "completions/mean_terminated_length": 70.45,
843
+ "completions/min_length": 70.28,
844
+ "completions/min_terminated_length": 70.28,
845
+ "epoch": 0.00275305944832241,
846
+ "frac_reward_zero_std": 0.6,
847
+ "grad_norm": 0.18359848856925964,
848
+ "kl": 0.21603738449513912,
849
+ "learning_rate": 1.7195444150884752e-07,
850
+ "loss": 0.0005052408576011658,
851
+ "num_tokens": 1465586.0,
852
+ "reward": 2.384185791015625e-09,
853
+ "reward_std": 0.42761797904968263,
854
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.384185791015625e-09,
855
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
856
+ "step": 1550
857
+ },
858
+ {
859
+ "clip_ratio/high_max": 0.0,
860
+ "clip_ratio/high_mean": 0.0,
861
+ "clip_ratio/low_mean": 0.0,
862
+ "clip_ratio/low_min": 0.0,
863
+ "clip_ratio/region_mean": 0.0,
864
+ "completion_length": 65.97,
865
+ "completions/clipped_ratio": 0.0,
866
+ "completions/max_length": 66.28,
867
+ "completions/max_terminated_length": 66.28,
868
+ "completions/mean_length": 65.97,
869
+ "completions/mean_terminated_length": 65.97,
870
+ "completions/min_length": 65.72,
871
+ "completions/min_terminated_length": 65.72,
872
+ "epoch": 0.002841867817623133,
873
+ "frac_reward_zero_std": 0.6,
874
+ "grad_norm": 0.001628563622944057,
875
+ "kl": 0.2370056840032339,
876
+ "learning_rate": 1.7750493994360695e-07,
877
+ "loss": 0.0010255253314971923,
878
+ "num_tokens": 1513108.0,
879
+ "reward": 5.960464477539063e-10,
880
+ "reward_std": 0.42761797904968263,
881
+ "rewards/TranscriptCorrectionGrpoReward/mean": 5.960464477539063e-10,
882
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
883
+ "step": 1600
884
+ },
885
+ {
886
+ "clip_ratio/high_max": 0.0,
887
+ "clip_ratio/high_mean": 0.0,
888
+ "clip_ratio/low_mean": 0.00017536476254463197,
889
+ "clip_ratio/low_min": 0.00017536476254463197,
890
+ "clip_ratio/region_mean": 0.00017536476254463197,
891
+ "completion_length": 65.855,
892
+ "completions/clipped_ratio": 0.0,
893
+ "completions/max_length": 66.16,
894
+ "completions/max_terminated_length": 66.16,
895
+ "completions/mean_length": 65.855,
896
+ "completions/mean_terminated_length": 65.855,
897
+ "completions/min_length": 65.6,
898
+ "completions/min_terminated_length": 65.6,
899
+ "epoch": 0.002930676186923856,
900
+ "frac_reward_zero_std": 0.52,
901
+ "grad_norm": 0.053214434534311295,
902
+ "kl": 0.24284519530832768,
903
+ "learning_rate": 1.8305543837836638e-07,
904
+ "loss": 0.00212494432926178,
905
+ "num_tokens": 1559663.0,
906
+ "reward": 1.4901161193847657e-09,
907
+ "reward_std": 0.5131415796279907,
908
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.4901161193847657e-09,
909
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5131415748596191,
910
+ "step": 1650
911
+ },
912
+ {
913
+ "clip_ratio/high_max": 0.0,
914
+ "clip_ratio/high_mean": 0.0,
915
+ "clip_ratio/low_mean": 0.0,
916
+ "clip_ratio/low_min": 0.0,
917
+ "clip_ratio/region_mean": 0.0,
918
+ "completion_length": 67.06,
919
+ "completions/clipped_ratio": 0.0,
920
+ "completions/max_length": 67.2,
921
+ "completions/max_terminated_length": 67.2,
922
+ "completions/mean_length": 67.06,
923
+ "completions/mean_terminated_length": 67.06,
924
+ "completions/min_length": 66.8,
925
+ "completions/min_terminated_length": 66.8,
926
+ "epoch": 0.0030194845562245788,
927
+ "frac_reward_zero_std": 0.6,
928
+ "grad_norm": 4.732516117655905e-06,
929
+ "kl": 0.2397597748413682,
930
+ "learning_rate": 1.8860593681312582e-07,
931
+ "loss": 0.0009849418699741364,
932
+ "num_tokens": 1607275.0,
933
+ "reward": 8.940696716308593e-10,
934
+ "reward_std": 0.42761796951293946,
935
+ "rewards/TranscriptCorrectionGrpoReward/mean": 8.940696716308593e-10,
936
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
937
+ "step": 1700
938
+ },
939
+ {
940
+ "clip_ratio/high_max": 0.0,
941
+ "clip_ratio/high_mean": 0.0,
942
+ "clip_ratio/low_mean": 0.0,
943
+ "clip_ratio/low_min": 0.0,
944
+ "clip_ratio/region_mean": 0.0,
945
+ "completion_length": 59.24,
946
+ "completions/clipped_ratio": 0.0,
947
+ "completions/max_length": 59.44,
948
+ "completions/max_terminated_length": 59.44,
949
+ "completions/mean_length": 59.24,
950
+ "completions/mean_terminated_length": 59.24,
951
+ "completions/min_length": 58.64,
952
+ "completions/min_terminated_length": 58.64,
953
+ "epoch": 0.0031082929255253017,
954
+ "frac_reward_zero_std": 0.52,
955
+ "grad_norm": 0.0002946323947981,
956
+ "kl": 0.2524130155146122,
957
+ "learning_rate": 1.9415643524788528e-07,
958
+ "loss": -0.00041158072650432587,
959
+ "num_tokens": 1649715.0,
960
+ "reward": 1.7881393432617187e-09,
961
+ "reward_std": 0.5131415748596191,
962
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
963
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5131415748596191,
964
+ "step": 1750
965
+ },
966
+ {
967
+ "clip_ratio/high_max": 0.0,
968
+ "clip_ratio/high_mean": 0.0,
969
+ "clip_ratio/low_mean": 8.620689623057842e-05,
970
+ "clip_ratio/low_min": 8.620689623057842e-05,
971
+ "clip_ratio/region_mean": 8.620689623057842e-05,
972
+ "completion_length": 69.005,
973
+ "completions/clipped_ratio": 0.0,
974
+ "completions/max_length": 69.24,
975
+ "completions/max_terminated_length": 69.24,
976
+ "completions/mean_length": 69.005,
977
+ "completions/mean_terminated_length": 69.005,
978
+ "completions/min_length": 68.72,
979
+ "completions/min_terminated_length": 68.72,
980
+ "epoch": 0.003197101294826024,
981
+ "frac_reward_zero_std": 0.6,
982
+ "grad_norm": 1.7265272163058398e-06,
983
+ "kl": 0.2614383715391159,
984
+ "learning_rate": 1.997069336826447e-07,
985
+ "loss": 0.0002399120479822159,
986
+ "num_tokens": 1698684.0,
987
+ "reward": 2.9802322387695313e-10,
988
+ "reward_std": 0.42761797904968263,
989
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.9802322387695313e-10,
990
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
991
+ "step": 1800
992
+ },
993
+ {
994
+ "clip_ratio/high_max": 0.0,
995
+ "clip_ratio/high_mean": 0.0,
996
+ "clip_ratio/low_mean": 0.0,
997
+ "clip_ratio/low_min": 0.0,
998
+ "clip_ratio/region_mean": 0.0,
999
+ "completion_length": 72.4,
1000
+ "completions/clipped_ratio": 0.0,
1001
+ "completions/max_length": 72.76,
1002
+ "completions/max_terminated_length": 72.76,
1003
+ "completions/mean_length": 72.4,
1004
+ "completions/mean_terminated_length": 72.4,
1005
+ "completions/min_length": 72.08,
1006
+ "completions/min_terminated_length": 72.08,
1007
+ "epoch": 0.003285909664126747,
1008
+ "frac_reward_zero_std": 0.44,
1009
+ "grad_norm": 1.8452697077009361e-06,
1010
+ "kl": 0.2667967968434095,
1011
+ "learning_rate": 2.0525743211740417e-07,
1012
+ "loss": 0.0009418275952339173,
1013
+ "num_tokens": 1750300.0,
1014
+ "reward": 2.9802322387695314e-09,
1015
+ "reward_std": 0.5986651706695557,
1016
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.9802322387695314e-09,
1017
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5986651706695557,
1018
+ "step": 1850
1019
+ },
1020
+ {
1021
+ "clip_ratio/high_max": 0.0,
1022
+ "clip_ratio/high_mean": 0.0,
1023
+ "clip_ratio/low_mean": 0.0,
1024
+ "clip_ratio/low_min": 0.0,
1025
+ "clip_ratio/region_mean": 0.0,
1026
+ "completion_length": 61.86,
1027
+ "completions/clipped_ratio": 0.0,
1028
+ "completions/max_length": 62.0,
1029
+ "completions/max_terminated_length": 62.0,
1030
+ "completions/mean_length": 61.86,
1031
+ "completions/mean_terminated_length": 61.86,
1032
+ "completions/min_length": 61.68,
1033
+ "completions/min_terminated_length": 61.68,
1034
+ "epoch": 0.00337471803342747,
1035
+ "frac_reward_zero_std": 0.64,
1036
+ "grad_norm": 9.591031812306028e-06,
1037
+ "kl": 0.2360048858821392,
1038
+ "learning_rate": 2.1080793055216363e-07,
1039
+ "loss": 0.001219682991504669,
1040
+ "num_tokens": 1794576.0,
1041
+ "reward": 4.76837158203125e-09,
1042
+ "reward_std": 0.38485618114471437,
1043
+ "rewards/TranscriptCorrectionGrpoReward/mean": 4.76837158203125e-09,
1044
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.38485618114471437,
1045
+ "step": 1900
1046
+ },
1047
+ {
1048
+ "clip_ratio/high_max": 0.0,
1049
+ "clip_ratio/high_mean": 0.0,
1050
+ "clip_ratio/low_mean": 7.24637694656849e-05,
1051
+ "clip_ratio/low_min": 7.24637694656849e-05,
1052
+ "clip_ratio/region_mean": 7.24637694656849e-05,
1053
+ "completion_length": 67.59,
1054
+ "completions/clipped_ratio": 0.0,
1055
+ "completions/max_length": 67.76,
1056
+ "completions/max_terminated_length": 67.76,
1057
+ "completions/mean_length": 67.59,
1058
+ "completions/mean_terminated_length": 67.59,
1059
+ "completions/min_length": 67.48,
1060
+ "completions/min_terminated_length": 67.48,
1061
+ "epoch": 0.003463526402728193,
1062
+ "frac_reward_zero_std": 0.56,
1063
+ "grad_norm": 1.6016636436688714e-05,
1064
+ "kl": 0.23783834256231784,
1065
+ "learning_rate": 2.1635842898692306e-07,
1066
+ "loss": 0.0007591144740581512,
1067
+ "num_tokens": 1842646.0,
1068
+ "reward": 2.9802322387695314e-09,
1069
+ "reward_std": 0.4703797769546509,
1070
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.9802322387695314e-09,
1071
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4703797769546509,
1072
+ "step": 1950
1073
+ },
1074
+ {
1075
+ "clip_ratio/high_max": 0.0,
1076
+ "clip_ratio/high_mean": 0.0,
1077
+ "clip_ratio/low_mean": 0.00012345679104328156,
1078
+ "clip_ratio/low_min": 0.00012345679104328156,
1079
+ "clip_ratio/region_mean": 0.00012345679104328156,
1080
+ "completion_length": 61.395,
1081
+ "completions/clipped_ratio": 0.0,
1082
+ "completions/max_length": 61.64,
1083
+ "completions/max_terminated_length": 61.64,
1084
+ "completions/mean_length": 61.395,
1085
+ "completions/mean_terminated_length": 61.395,
1086
+ "completions/min_length": 61.16,
1087
+ "completions/min_terminated_length": 61.16,
1088
+ "epoch": 0.003552334772028916,
1089
+ "frac_reward_zero_std": 0.6,
1090
+ "grad_norm": 4.785554847330786e-05,
1091
+ "kl": 0.25009929470717907,
1092
+ "learning_rate": 2.219089274216825e-07,
1093
+ "loss": 0.0014793139696121216,
1094
+ "num_tokens": 1886781.0,
1095
+ "reward": 0.0,
1096
+ "reward_std": 0.4276179885864258,
1097
+ "rewards/TranscriptCorrectionGrpoReward/mean": 0.0,
1098
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
1099
+ "step": 2000
1100
+ },
1101
+ {
1102
+ "eval_cer_subset": 0.012620710993860635,
1103
+ "eval_cer_subset_edit_distance": 775,
1104
+ "eval_cer_subset_groups": 250,
1105
+ "eval_cer_subset_items": 250,
1106
+ "eval_cer_subset_ref_chars": 61407,
1107
+ "eval_loss": 0.03968957871396896,
1108
+ "step": 2026
1109
+ },
1110
+ {
1111
+ "clip_ratio/high_max": 0.0,
1112
+ "clip_ratio/high_mean": 0.0,
1113
+ "clip_ratio/low_mean": 0.0,
1114
+ "clip_ratio/low_min": 0.0,
1115
+ "clip_ratio/region_mean": 0.0,
1116
+ "completion_length": 70.95,
1117
+ "completions/clipped_ratio": 0.0,
1118
+ "completions/max_length": 71.12,
1119
+ "completions/max_terminated_length": 71.12,
1120
+ "completions/mean_length": 70.95,
1121
+ "completions/mean_terminated_length": 70.95,
1122
+ "completions/min_length": 70.8,
1123
+ "completions/min_terminated_length": 70.8,
1124
+ "epoch": 0.003641143141329639,
1125
+ "frac_reward_zero_std": 0.64,
1126
+ "grad_norm": 1.3850582035956904e-05,
1127
+ "kl": 0.24420862443745137,
1128
+ "learning_rate": 2.2745942585644193e-07,
1129
+ "loss": 0.0013575667142868041,
1130
+ "num_tokens": 1937379.0,
1131
+ "reward": 2.682209014892578e-09,
1132
+ "reward_std": 0.3848561763763428,
1133
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.682209014892578e-09,
1134
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.38485618114471437,
1135
+ "step": 2050
1136
+ },
1137
+ {
1138
+ "clip_ratio/high_max": 0.0001694915257394314,
1139
+ "clip_ratio/high_mean": 0.0001694915257394314,
1140
+ "clip_ratio/low_mean": 0.0001694915257394314,
1141
+ "clip_ratio/low_min": 0.0001694915257394314,
1142
+ "clip_ratio/region_mean": 0.0003389830514788628,
1143
+ "completion_length": 66.48,
1144
+ "completions/clipped_ratio": 0.0,
1145
+ "completions/max_length": 66.8,
1146
+ "completions/max_terminated_length": 66.8,
1147
+ "completions/mean_length": 66.48,
1148
+ "completions/mean_terminated_length": 66.48,
1149
+ "completions/min_length": 66.28,
1150
+ "completions/min_terminated_length": 66.28,
1151
+ "epoch": 0.003729951510630362,
1152
+ "frac_reward_zero_std": 0.44,
1153
+ "grad_norm": 9.167926691588946e-06,
1154
+ "kl": 0.255931094288826,
1155
+ "learning_rate": 2.3300992429120136e-07,
1156
+ "loss": 0.0010222384333610534,
1157
+ "num_tokens": 1984251.0,
1158
+ "reward": 1.7881393432617187e-09,
1159
+ "reward_std": 0.5986651706695557,
1160
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
1161
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5986651706695557,
1162
+ "step": 2100
1163
+ },
1164
+ {
1165
+ "clip_ratio/high_max": 0.0,
1166
+ "clip_ratio/high_mean": 0.0,
1167
+ "clip_ratio/low_mean": 5.882353056222201e-05,
1168
+ "clip_ratio/low_min": 5.882353056222201e-05,
1169
+ "clip_ratio/region_mean": 5.882353056222201e-05,
1170
+ "completion_length": 69.44,
1171
+ "completions/clipped_ratio": 0.0,
1172
+ "completions/max_length": 69.6,
1173
+ "completions/max_terminated_length": 69.6,
1174
+ "completions/mean_length": 69.44,
1175
+ "completions/mean_terminated_length": 69.44,
1176
+ "completions/min_length": 69.04,
1177
+ "completions/min_terminated_length": 69.04,
1178
+ "epoch": 0.0038187598799310847,
1179
+ "frac_reward_zero_std": 0.64,
1180
+ "grad_norm": 0.1219794824719429,
1181
+ "kl": 0.22736032783985138,
1182
+ "learning_rate": 2.385604227259608e-07,
1183
+ "loss": 0.00033545993268489836,
1184
+ "num_tokens": 2034003.0,
1185
+ "reward": 3.5762786865234374e-09,
1186
+ "reward_std": 0.38485618114471437,
1187
+ "rewards/TranscriptCorrectionGrpoReward/mean": 3.5762786865234374e-09,
1188
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.38485618114471437,
1189
+ "step": 2150
1190
+ },
1191
+ {
1192
+ "clip_ratio/high_max": 5.8139534667134284e-05,
1193
+ "clip_ratio/high_mean": 5.8139534667134284e-05,
1194
+ "clip_ratio/low_mean": 0.0,
1195
+ "clip_ratio/low_min": 0.0,
1196
+ "clip_ratio/region_mean": 5.8139534667134284e-05,
1197
+ "completion_length": 59.31,
1198
+ "completions/clipped_ratio": 0.0,
1199
+ "completions/max_length": 59.52,
1200
+ "completions/max_terminated_length": 59.52,
1201
+ "completions/mean_length": 59.31,
1202
+ "completions/mean_terminated_length": 59.31,
1203
+ "completions/min_length": 59.08,
1204
+ "completions/min_terminated_length": 59.08,
1205
+ "epoch": 0.003907568249231807,
1206
+ "frac_reward_zero_std": 0.6,
1207
+ "grad_norm": 2.484521610313095e-05,
1208
+ "kl": 0.2782662433385849,
1209
+ "learning_rate": 2.4411092116072025e-07,
1210
+ "loss": 0.0015113741159439088,
1211
+ "num_tokens": 2076841.0,
1212
+ "reward": 8.940696716308593e-10,
1213
+ "reward_std": 0.42761797428131104,
1214
+ "rewards/TranscriptCorrectionGrpoReward/mean": 8.940696716308593e-10,
1215
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
1216
+ "step": 2200
1217
+ },
1218
+ {
1219
+ "clip_ratio/high_max": 0.0,
1220
+ "clip_ratio/high_mean": 0.0,
1221
+ "clip_ratio/low_mean": 0.0,
1222
+ "clip_ratio/low_min": 0.0,
1223
+ "clip_ratio/region_mean": 0.0,
1224
+ "completion_length": 65.14,
1225
+ "completions/clipped_ratio": 0.0,
1226
+ "completions/max_length": 65.4,
1227
+ "completions/max_terminated_length": 65.4,
1228
+ "completions/mean_length": 65.14,
1229
+ "completions/mean_terminated_length": 65.14,
1230
+ "completions/min_length": 64.88,
1231
+ "completions/min_terminated_length": 64.88,
1232
+ "epoch": 0.00399637661853253,
1233
+ "frac_reward_zero_std": 0.6,
1234
+ "grad_norm": 4.3012671085307375e-05,
1235
+ "kl": 0.23348225951194762,
1236
+ "learning_rate": 2.4966141959547966e-07,
1237
+ "loss": 0.0008960606157779694,
1238
+ "num_tokens": 2122893.0,
1239
+ "reward": 0.0,
1240
+ "reward_std": 0.4276179838180542,
1241
+ "rewards/TranscriptCorrectionGrpoReward/mean": 0.0,
1242
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4276179838180542,
1243
+ "step": 2250
1244
+ },
1245
+ {
1246
+ "clip_ratio/high_max": 0.0,
1247
+ "clip_ratio/high_mean": 0.0,
1248
+ "clip_ratio/low_mean": 0.00014705882407724857,
1249
+ "clip_ratio/low_min": 0.00014705882407724857,
1250
+ "clip_ratio/region_mean": 0.00014705882407724857,
1251
+ "completion_length": 63.71,
1252
+ "completions/clipped_ratio": 0.0,
1253
+ "completions/max_length": 64.0,
1254
+ "completions/max_terminated_length": 64.0,
1255
+ "completions/mean_length": 63.71,
1256
+ "completions/mean_terminated_length": 63.71,
1257
+ "completions/min_length": 63.52,
1258
+ "completions/min_terminated_length": 63.52,
1259
+ "epoch": 0.004085184987833253,
1260
+ "frac_reward_zero_std": 0.68,
1261
+ "grad_norm": 0.15578855574131012,
1262
+ "kl": 0.24591257102787495,
1263
+ "learning_rate": 2.552119180302391e-07,
1264
+ "loss": 0.0016065806150436402,
1265
+ "num_tokens": 2168475.0,
1266
+ "reward": 1.7881393432617187e-09,
1267
+ "reward_std": 0.3420943832397461,
1268
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
1269
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
1270
+ "step": 2300
1271
+ },
1272
+ {
1273
+ "clip_ratio/high_max": 0.0,
1274
+ "clip_ratio/high_mean": 0.0,
1275
+ "clip_ratio/low_mean": 0.0002476652339100838,
1276
+ "clip_ratio/low_min": 0.0002476652339100838,
1277
+ "clip_ratio/region_mean": 0.0002476652339100838,
1278
+ "completion_length": 60.545,
1279
+ "completions/clipped_ratio": 0.0,
1280
+ "completions/max_length": 60.76,
1281
+ "completions/max_terminated_length": 60.76,
1282
+ "completions/mean_length": 60.545,
1283
+ "completions/mean_terminated_length": 60.545,
1284
+ "completions/min_length": 60.24,
1285
+ "completions/min_terminated_length": 60.24,
1286
+ "epoch": 0.004173993357133976,
1287
+ "frac_reward_zero_std": 0.64,
1288
+ "grad_norm": 0.07905205339193344,
1289
+ "kl": 0.23256765738129614,
1290
+ "learning_rate": 2.607624164649986e-07,
1291
+ "loss": 0.0005145055055618286,
1292
+ "num_tokens": 2212784.0,
1293
+ "reward": 2.9802322387695314e-09,
1294
+ "reward_std": 0.38485618591308596,
1295
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.9802322387695314e-09,
1296
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.38485618114471437,
1297
+ "step": 2350
1298
+ },
1299
+ {
1300
+ "clip_ratio/high_max": 0.0,
1301
+ "clip_ratio/high_mean": 0.0,
1302
+ "clip_ratio/low_mean": 4.9019609577953814e-05,
1303
+ "clip_ratio/low_min": 4.9019609577953814e-05,
1304
+ "clip_ratio/region_mean": 4.9019609577953814e-05,
1305
+ "completion_length": 73.87,
1306
+ "completions/clipped_ratio": 0.0,
1307
+ "completions/max_length": 74.0,
1308
+ "completions/max_terminated_length": 74.0,
1309
+ "completions/mean_length": 73.87,
1310
+ "completions/mean_terminated_length": 73.87,
1311
+ "completions/min_length": 73.64,
1312
+ "completions/min_terminated_length": 73.64,
1313
+ "epoch": 0.004262801726434699,
1314
+ "frac_reward_zero_std": 0.72,
1315
+ "grad_norm": 0.00022796839766670018,
1316
+ "kl": 0.22991662960499526,
1317
+ "learning_rate": 2.6631291489975804e-07,
1318
+ "loss": 0.0011378635466098785,
1319
+ "num_tokens": 2264686.0,
1320
+ "reward": 1.1920928955078125e-09,
1321
+ "reward_std": 0.2993325901031494,
1322
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.1920928955078125e-09,
1323
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.29933258533477786,
1324
+ "step": 2400
1325
+ },
1326
+ {
1327
+ "clip_ratio/high_max": 0.0,
1328
+ "clip_ratio/high_mean": 0.0,
1329
+ "clip_ratio/low_mean": 5.2083334885537625e-05,
1330
+ "clip_ratio/low_min": 5.2083334885537625e-05,
1331
+ "clip_ratio/region_mean": 5.2083334885537625e-05,
1332
+ "completion_length": 67.66,
1333
+ "completions/clipped_ratio": 0.0,
1334
+ "completions/max_length": 67.92,
1335
+ "completions/max_terminated_length": 67.92,
1336
+ "completions/mean_length": 67.66,
1337
+ "completions/mean_terminated_length": 67.66,
1338
+ "completions/min_length": 67.44,
1339
+ "completions/min_terminated_length": 67.44,
1340
+ "epoch": 0.004351610095735422,
1341
+ "frac_reward_zero_std": 0.36,
1342
+ "grad_norm": 0.20806169509887695,
1343
+ "kl": 0.23162569977343084,
1344
+ "learning_rate": 2.7186341333451744e-07,
1345
+ "loss": 0.0006114953756332398,
1346
+ "num_tokens": 2312594.0,
1347
+ "reward": 3.5762786865234374e-09,
1348
+ "reward_std": 0.6841887712478638,
1349
+ "rewards/TranscriptCorrectionGrpoReward/mean": 3.5762786865234374e-09,
1350
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.6841887664794922,
1351
+ "step": 2450
1352
+ },
1353
+ {
1354
+ "clip_ratio/high_max": 0.0,
1355
+ "clip_ratio/high_mean": 0.0,
1356
+ "clip_ratio/low_mean": 0.0001515151560306549,
1357
+ "clip_ratio/low_min": 0.0001515151560306549,
1358
+ "clip_ratio/region_mean": 0.0001515151560306549,
1359
+ "completion_length": 62.02,
1360
+ "completions/clipped_ratio": 0.0,
1361
+ "completions/max_length": 62.44,
1362
+ "completions/max_terminated_length": 62.44,
1363
+ "completions/mean_length": 62.02,
1364
+ "completions/mean_terminated_length": 62.02,
1365
+ "completions/min_length": 61.68,
1366
+ "completions/min_terminated_length": 61.68,
1367
+ "epoch": 0.004440418465036145,
1368
+ "frac_reward_zero_std": 0.44,
1369
+ "grad_norm": 0.03983508050441742,
1370
+ "kl": 0.24266708113253116,
1371
+ "learning_rate": 2.774139117692769e-07,
1372
+ "loss": 0.0009456196427345276,
1373
+ "num_tokens": 2357438.0,
1374
+ "reward": 2.086162567138672e-09,
1375
+ "reward_std": 0.5986651706695557,
1376
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.086162567138672e-09,
1377
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5986651706695557,
1378
+ "step": 2500
1379
+ },
1380
+ {
1381
+ "clip_ratio/high_max": 0.0,
1382
+ "clip_ratio/high_mean": 0.0,
1383
+ "clip_ratio/low_mean": 0.0,
1384
+ "clip_ratio/low_min": 0.0,
1385
+ "clip_ratio/region_mean": 0.0,
1386
+ "completion_length": 72.79,
1387
+ "completions/clipped_ratio": 0.0,
1388
+ "completions/max_length": 72.88,
1389
+ "completions/max_terminated_length": 72.88,
1390
+ "completions/mean_length": 72.79,
1391
+ "completions/mean_terminated_length": 72.79,
1392
+ "completions/min_length": 72.64,
1393
+ "completions/min_terminated_length": 72.64,
1394
+ "epoch": 0.004529226834336868,
1395
+ "frac_reward_zero_std": 0.6,
1396
+ "grad_norm": 1.2328011507634073e-05,
1397
+ "kl": 0.22370575830340386,
1398
+ "learning_rate": 2.8296441020403636e-07,
1399
+ "loss": 0.0008269050717353821,
1400
+ "num_tokens": 2408772.0,
1401
+ "reward": -2.9802322387695313e-10,
1402
+ "reward_std": 0.4276179838180542,
1403
+ "rewards/TranscriptCorrectionGrpoReward/mean": -2.9802322387695313e-10,
1404
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
1405
+ "step": 2550
1406
+ },
1407
+ {
1408
+ "clip_ratio/high_max": 0.0,
1409
+ "clip_ratio/high_mean": 0.0,
1410
+ "clip_ratio/low_mean": 9.09090880304575e-05,
1411
+ "clip_ratio/low_min": 9.09090880304575e-05,
1412
+ "clip_ratio/region_mean": 9.09090880304575e-05,
1413
+ "completion_length": 63.755,
1414
+ "completions/clipped_ratio": 0.0,
1415
+ "completions/max_length": 64.0,
1416
+ "completions/max_terminated_length": 64.0,
1417
+ "completions/mean_length": 63.755,
1418
+ "completions/mean_terminated_length": 63.755,
1419
+ "completions/min_length": 63.28,
1420
+ "completions/min_terminated_length": 63.28,
1421
+ "epoch": 0.004618035203637591,
1422
+ "frac_reward_zero_std": 0.48,
1423
+ "grad_norm": 1.6053049876063596e-06,
1424
+ "kl": 0.26070301596075296,
1425
+ "learning_rate": 2.8851490863879577e-07,
1426
+ "loss": 0.00038771554827690127,
1427
+ "num_tokens": 2454795.0,
1428
+ "reward": 4.172325134277344e-09,
1429
+ "reward_std": 0.5559033775329589,
1430
+ "rewards/TranscriptCorrectionGrpoReward/mean": 4.172325134277344e-09,
1431
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5559033727645875,
1432
+ "step": 2600
1433
+ },
1434
+ {
1435
+ "clip_ratio/high_max": 0.0,
1436
+ "clip_ratio/high_mean": 0.0,
1437
+ "clip_ratio/low_mean": 0.0,
1438
+ "clip_ratio/low_min": 0.0,
1439
+ "clip_ratio/region_mean": 0.0,
1440
+ "completion_length": 61.365,
1441
+ "completions/clipped_ratio": 0.0,
1442
+ "completions/max_length": 61.48,
1443
+ "completions/max_terminated_length": 61.48,
1444
+ "completions/mean_length": 61.365,
1445
+ "completions/mean_terminated_length": 61.365,
1446
+ "completions/min_length": 61.12,
1447
+ "completions/min_terminated_length": 61.12,
1448
+ "epoch": 0.004706843572938314,
1449
+ "frac_reward_zero_std": 0.68,
1450
+ "grad_norm": 1.4448678484768607e-05,
1451
+ "kl": 0.2749948535114527,
1452
+ "learning_rate": 2.9406540707355523e-07,
1453
+ "loss": 0.00047413244843482974,
1454
+ "num_tokens": 2498940.0,
1455
+ "reward": 2.384185791015625e-09,
1456
+ "reward_std": 0.3420943832397461,
1457
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.384185791015625e-09,
1458
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
1459
+ "step": 2650
1460
+ },
1461
+ {
1462
+ "clip_ratio/high_max": 0.0,
1463
+ "clip_ratio/high_mean": 0.0,
1464
+ "clip_ratio/low_mean": 0.0,
1465
+ "clip_ratio/low_min": 0.0,
1466
+ "clip_ratio/region_mean": 0.0,
1467
+ "completion_length": 58.905,
1468
+ "completions/clipped_ratio": 0.0,
1469
+ "completions/max_length": 59.12,
1470
+ "completions/max_terminated_length": 59.12,
1471
+ "completions/mean_length": 58.905,
1472
+ "completions/mean_terminated_length": 58.905,
1473
+ "completions/min_length": 58.6,
1474
+ "completions/min_terminated_length": 58.6,
1475
+ "epoch": 0.0047956519422390365,
1476
+ "frac_reward_zero_std": 0.6,
1477
+ "grad_norm": 4.0039867599261925e-07,
1478
+ "kl": 0.28236587412655356,
1479
+ "learning_rate": 2.9961590550831464e-07,
1480
+ "loss": 0.0008276945352554321,
1481
+ "num_tokens": 2541689.0,
1482
+ "reward": 1.1920928955078125e-09,
1483
+ "reward_std": 0.42761797904968263,
1484
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.1920928955078125e-09,
1485
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
1486
+ "step": 2700
1487
+ },
1488
+ {
1489
+ "clip_ratio/high_max": 0.0,
1490
+ "clip_ratio/high_mean": 0.0,
1491
+ "clip_ratio/low_mean": 0.00017241379246115684,
1492
+ "clip_ratio/low_min": 0.00017241379246115684,
1493
+ "clip_ratio/region_mean": 0.00017241379246115684,
1494
+ "completion_length": 70.22,
1495
+ "completions/clipped_ratio": 0.0,
1496
+ "completions/max_length": 70.6,
1497
+ "completions/max_terminated_length": 70.6,
1498
+ "completions/mean_length": 70.22,
1499
+ "completions/mean_terminated_length": 70.22,
1500
+ "completions/min_length": 69.96,
1501
+ "completions/min_terminated_length": 69.96,
1502
+ "epoch": 0.0048844603115397595,
1503
+ "frac_reward_zero_std": 0.6,
1504
+ "grad_norm": 0.19178590178489685,
1505
+ "kl": 0.2311771822348237,
1506
+ "learning_rate": 3.051664039430741e-07,
1507
+ "loss": 0.0014842292666435242,
1508
+ "num_tokens": 2591485.0,
1509
+ "reward": -1.1920928955078125e-09,
1510
+ "reward_std": 0.42761797904968263,
1511
+ "rewards/TranscriptCorrectionGrpoReward/mean": -1.1920928955078125e-09,
1512
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
1513
+ "step": 2750
1514
+ },
1515
+ {
1516
+ "clip_ratio/high_max": 0.0,
1517
+ "clip_ratio/high_mean": 0.0,
1518
+ "clip_ratio/low_mean": 0.0,
1519
+ "clip_ratio/low_min": 0.0,
1520
+ "clip_ratio/region_mean": 0.0,
1521
+ "completion_length": 56.515,
1522
+ "completions/clipped_ratio": 0.0,
1523
+ "completions/max_length": 56.76,
1524
+ "completions/max_terminated_length": 56.76,
1525
+ "completions/mean_length": 56.515,
1526
+ "completions/mean_terminated_length": 56.515,
1527
+ "completions/min_length": 56.04,
1528
+ "completions/min_terminated_length": 56.04,
1529
+ "epoch": 0.004973268680840482,
1530
+ "frac_reward_zero_std": 0.36,
1531
+ "grad_norm": 0.5189715623855591,
1532
+ "kl": 0.34065449707210066,
1533
+ "learning_rate": 3.1071690237783356e-07,
1534
+ "loss": 0.0011833453178405763,
1535
+ "num_tokens": 2632556.0,
1536
+ "reward": 4.76837158203125e-09,
1537
+ "reward_std": 0.6841887712478638,
1538
+ "rewards/TranscriptCorrectionGrpoReward/mean": 4.76837158203125e-09,
1539
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.6841887664794922,
1540
+ "step": 2800
1541
+ },
1542
+ {
1543
+ "clip_ratio/high_max": 0.0,
1544
+ "clip_ratio/high_mean": 0.0,
1545
+ "clip_ratio/low_mean": 0.0,
1546
+ "clip_ratio/low_min": 0.0,
1547
+ "clip_ratio/region_mean": 0.0,
1548
+ "completion_length": 62.935,
1549
+ "completions/clipped_ratio": 0.0,
1550
+ "completions/max_length": 63.08,
1551
+ "completions/max_terminated_length": 63.08,
1552
+ "completions/mean_length": 62.935,
1553
+ "completions/mean_terminated_length": 62.935,
1554
+ "completions/min_length": 62.8,
1555
+ "completions/min_terminated_length": 62.8,
1556
+ "epoch": 0.005062077050141205,
1557
+ "frac_reward_zero_std": 0.6,
1558
+ "grad_norm": 0.24930885434150696,
1559
+ "kl": 0.26265339344739913,
1560
+ "learning_rate": 3.16267400812593e-07,
1561
+ "loss": 0.0013879780471324921,
1562
+ "num_tokens": 2678095.0,
1563
+ "reward": 4.172325134277344e-09,
1564
+ "reward_std": 0.42761797904968263,
1565
+ "rewards/TranscriptCorrectionGrpoReward/mean": 4.172325134277344e-09,
1566
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4276179838180542,
1567
+ "step": 2850
1568
+ },
1569
+ {
1570
+ "clip_ratio/high_max": 0.0,
1571
+ "clip_ratio/high_mean": 0.0,
1572
+ "clip_ratio/low_mean": 0.0001515151560306549,
1573
+ "clip_ratio/low_min": 0.0001515151560306549,
1574
+ "clip_ratio/region_mean": 0.0001515151560306549,
1575
+ "completion_length": 60.645,
1576
+ "completions/clipped_ratio": 0.0,
1577
+ "completions/max_length": 60.8,
1578
+ "completions/max_terminated_length": 60.8,
1579
+ "completions/mean_length": 60.645,
1580
+ "completions/mean_terminated_length": 60.645,
1581
+ "completions/min_length": 60.56,
1582
+ "completions/min_terminated_length": 60.56,
1583
+ "epoch": 0.005150885419441928,
1584
+ "frac_reward_zero_std": 0.64,
1585
+ "grad_norm": 0.0002522142603993416,
1586
+ "kl": 0.2475396691262722,
1587
+ "learning_rate": 3.218178992473524e-07,
1588
+ "loss": 0.0014786873757839203,
1589
+ "num_tokens": 2721896.0,
1590
+ "reward": -5.960464477539063e-10,
1591
+ "reward_std": 0.3848561763763428,
1592
+ "rewards/TranscriptCorrectionGrpoReward/mean": -5.960464477539063e-10,
1593
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.38485618114471437,
1594
+ "step": 2900
1595
+ },
1596
+ {
1597
+ "clip_ratio/high_max": 0.0,
1598
+ "clip_ratio/high_mean": 0.0,
1599
+ "clip_ratio/low_mean": 0.0,
1600
+ "clip_ratio/low_min": 0.0,
1601
+ "clip_ratio/region_mean": 0.0,
1602
+ "completion_length": 71.345,
1603
+ "completions/clipped_ratio": 0.0,
1604
+ "completions/max_length": 71.52,
1605
+ "completions/max_terminated_length": 71.52,
1606
+ "completions/mean_length": 71.345,
1607
+ "completions/mean_terminated_length": 71.345,
1608
+ "completions/min_length": 71.0,
1609
+ "completions/min_terminated_length": 71.0,
1610
+ "epoch": 0.005239693788742651,
1611
+ "frac_reward_zero_std": 0.52,
1612
+ "grad_norm": 3.1222178222378716e-05,
1613
+ "kl": 0.23209657415747642,
1614
+ "learning_rate": 3.273683976821119e-07,
1615
+ "loss": 0.0005750331282615662,
1616
+ "num_tokens": 2772885.0,
1617
+ "reward": 2.086162567138672e-09,
1618
+ "reward_std": 0.5131415748596191,
1619
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.086162567138672e-09,
1620
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5131415748596191,
1621
+ "step": 2950
1622
+ },
1623
+ {
1624
+ "clip_ratio/high_max": 0.0,
1625
+ "clip_ratio/high_mean": 0.0,
1626
+ "clip_ratio/low_mean": 6.329114083200693e-05,
1627
+ "clip_ratio/low_min": 6.329114083200693e-05,
1628
+ "clip_ratio/region_mean": 6.329114083200693e-05,
1629
+ "completion_length": 59.215,
1630
+ "completions/clipped_ratio": 0.0,
1631
+ "completions/max_length": 59.4,
1632
+ "completions/max_terminated_length": 59.4,
1633
+ "completions/mean_length": 59.215,
1634
+ "completions/mean_terminated_length": 59.215,
1635
+ "completions/min_length": 58.92,
1636
+ "completions/min_terminated_length": 58.92,
1637
+ "epoch": 0.005328502158043374,
1638
+ "frac_reward_zero_std": 0.56,
1639
+ "grad_norm": 0.23252332210540771,
1640
+ "kl": 0.2617241670191288,
1641
+ "learning_rate": 3.3291889611687134e-07,
1642
+ "loss": 0.0009682238847017288,
1643
+ "num_tokens": 2815864.0,
1644
+ "reward": 4.470348358154297e-09,
1645
+ "reward_std": 0.4703797721862793,
1646
+ "rewards/TranscriptCorrectionGrpoReward/mean": 4.470348358154297e-09,
1647
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4703797769546509,
1648
+ "step": 3000
1649
+ },
1650
+ {
1651
+ "eval_cer_subset": 0.01273470451251486,
1652
+ "eval_cer_subset_edit_distance": 782,
1653
+ "eval_cer_subset_groups": 250,
1654
+ "eval_cer_subset_items": 250,
1655
+ "eval_cer_subset_ref_chars": 61407,
1656
+ "eval_loss": 0.03991130820399113,
1657
+ "step": 3039
1658
+ },
1659
+ {
1660
+ "clip_ratio/high_max": 0.0,
1661
+ "clip_ratio/high_mean": 0.0,
1662
+ "clip_ratio/low_mean": 0.0,
1663
+ "clip_ratio/low_min": 0.0,
1664
+ "clip_ratio/region_mean": 0.0,
1665
+ "completion_length": 72.55,
1666
+ "completions/clipped_ratio": 0.0,
1667
+ "completions/max_length": 72.72,
1668
+ "completions/max_terminated_length": 72.72,
1669
+ "completions/mean_length": 72.55,
1670
+ "completions/mean_terminated_length": 72.55,
1671
+ "completions/min_length": 72.32,
1672
+ "completions/min_terminated_length": 72.32,
1673
+ "epoch": 0.005417310527344097,
1674
+ "frac_reward_zero_std": 0.68,
1675
+ "grad_norm": 0.29253867268562317,
1676
+ "kl": 0.21061068180948495,
1677
+ "learning_rate": 3.3846939455163075e-07,
1678
+ "loss": 0.0007596138119697571,
1679
+ "num_tokens": 2867654.0,
1680
+ "reward": -5.960464477539063e-10,
1681
+ "reward_std": 0.3420943880081177,
1682
+ "rewards/TranscriptCorrectionGrpoReward/mean": -5.960464477539063e-10,
1683
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
1684
+ "step": 3050
1685
+ },
1686
+ {
1687
+ "clip_ratio/high_max": 0.0,
1688
+ "clip_ratio/high_mean": 0.0,
1689
+ "clip_ratio/low_mean": 0.0,
1690
+ "clip_ratio/low_min": 0.0,
1691
+ "clip_ratio/region_mean": 0.0,
1692
+ "completion_length": 73.175,
1693
+ "completions/clipped_ratio": 0.0,
1694
+ "completions/max_length": 73.28,
1695
+ "completions/max_terminated_length": 73.28,
1696
+ "completions/mean_length": 73.175,
1697
+ "completions/mean_terminated_length": 73.175,
1698
+ "completions/min_length": 72.92,
1699
+ "completions/min_terminated_length": 72.92,
1700
+ "epoch": 0.00550611889664482,
1701
+ "frac_reward_zero_std": 0.52,
1702
+ "grad_norm": 0.11588963121175766,
1703
+ "kl": 0.28017708033323285,
1704
+ "learning_rate": 3.440198929863902e-07,
1705
+ "loss": 0.0013754051923751832,
1706
+ "num_tokens": 2919545.0,
1707
+ "reward": 1.7881393432617187e-09,
1708
+ "reward_std": 0.5131415796279907,
1709
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
1710
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5131415748596191,
1711
+ "step": 3100
1712
+ },
1713
+ {
1714
+ "clip_ratio/high_max": 0.0,
1715
+ "clip_ratio/high_mean": 0.0,
1716
+ "clip_ratio/low_mean": 7.692307699471712e-05,
1717
+ "clip_ratio/low_min": 7.692307699471712e-05,
1718
+ "clip_ratio/region_mean": 7.692307699471712e-05,
1719
+ "completion_length": 63.095,
1720
+ "completions/clipped_ratio": 0.0,
1721
+ "completions/max_length": 63.48,
1722
+ "completions/max_terminated_length": 63.48,
1723
+ "completions/mean_length": 63.095,
1724
+ "completions/mean_terminated_length": 63.095,
1725
+ "completions/min_length": 62.64,
1726
+ "completions/min_terminated_length": 62.64,
1727
+ "epoch": 0.005594927265945543,
1728
+ "frac_reward_zero_std": 0.48,
1729
+ "grad_norm": 0.33691588044166565,
1730
+ "kl": 0.26949979782104494,
1731
+ "learning_rate": 3.495703914211496e-07,
1732
+ "loss": -0.00045975685119628904,
1733
+ "num_tokens": 2964788.0,
1734
+ "reward": -2.9802322387695313e-10,
1735
+ "reward_std": 0.5559033775329589,
1736
+ "rewards/TranscriptCorrectionGrpoReward/mean": -2.9802322387695313e-10,
1737
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5559033727645875,
1738
+ "step": 3150
1739
+ },
1740
+ {
1741
+ "clip_ratio/high_max": 0.0,
1742
+ "clip_ratio/high_mean": 0.0,
1743
+ "clip_ratio/low_mean": 0.0,
1744
+ "clip_ratio/low_min": 0.0,
1745
+ "clip_ratio/region_mean": 0.0,
1746
+ "completion_length": 66.54,
1747
+ "completions/clipped_ratio": 0.0,
1748
+ "completions/max_length": 66.68,
1749
+ "completions/max_terminated_length": 66.68,
1750
+ "completions/mean_length": 66.54,
1751
+ "completions/mean_terminated_length": 66.54,
1752
+ "completions/min_length": 66.48,
1753
+ "completions/min_terminated_length": 66.48,
1754
+ "epoch": 0.005683735635246266,
1755
+ "frac_reward_zero_std": 0.68,
1756
+ "grad_norm": 0.3239286541938782,
1757
+ "kl": 0.2465494102984667,
1758
+ "learning_rate": 3.5512088985590907e-07,
1759
+ "loss": 0.0014468005299568176,
1760
+ "num_tokens": 3012112.0,
1761
+ "reward": 1.7881393432617187e-09,
1762
+ "reward_std": 0.3420943832397461,
1763
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
1764
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
1765
+ "step": 3200
1766
+ },
1767
+ {
1768
+ "clip_ratio/high_max": 0.0,
1769
+ "clip_ratio/high_mean": 0.0,
1770
+ "clip_ratio/low_mean": 0.00011363636702299118,
1771
+ "clip_ratio/low_min": 0.00011363636702299118,
1772
+ "clip_ratio/region_mean": 0.00011363636702299118,
1773
+ "completion_length": 73.105,
1774
+ "completions/clipped_ratio": 0.0,
1775
+ "completions/max_length": 73.16,
1776
+ "completions/max_terminated_length": 73.16,
1777
+ "completions/mean_length": 73.105,
1778
+ "completions/mean_terminated_length": 73.105,
1779
+ "completions/min_length": 73.0,
1780
+ "completions/min_terminated_length": 73.0,
1781
+ "epoch": 0.005772544004546989,
1782
+ "frac_reward_zero_std": 0.6,
1783
+ "grad_norm": 0.041704874485731125,
1784
+ "kl": 0.2259557291865349,
1785
+ "learning_rate": 3.6067138829066853e-07,
1786
+ "loss": 0.0008075982332229614,
1787
+ "num_tokens": 3063941.0,
1788
+ "reward": 8.940696716308593e-10,
1789
+ "reward_std": 0.4276179838180542,
1790
+ "rewards/TranscriptCorrectionGrpoReward/mean": 8.940696716308593e-10,
1791
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
1792
+ "step": 3250
1793
+ },
1794
+ {
1795
+ "clip_ratio/high_max": 0.0,
1796
+ "clip_ratio/high_mean": 0.0,
1797
+ "clip_ratio/low_mean": 0.0,
1798
+ "clip_ratio/low_min": 0.0,
1799
+ "clip_ratio/region_mean": 0.0,
1800
+ "completion_length": 65.525,
1801
+ "completions/clipped_ratio": 0.0,
1802
+ "completions/max_length": 65.68,
1803
+ "completions/max_terminated_length": 65.68,
1804
+ "completions/mean_length": 65.525,
1805
+ "completions/mean_terminated_length": 65.525,
1806
+ "completions/min_length": 65.24,
1807
+ "completions/min_terminated_length": 65.24,
1808
+ "epoch": 0.005861352373847712,
1809
+ "frac_reward_zero_std": 0.68,
1810
+ "grad_norm": 2.9454427931341343e-05,
1811
+ "kl": 0.2666619791090488,
1812
+ "learning_rate": 3.6622188672542794e-07,
1813
+ "loss": 0.0019758230447769164,
1814
+ "num_tokens": 3110286.0,
1815
+ "reward": -5.960464477539063e-10,
1816
+ "reward_std": 0.3420943832397461,
1817
+ "rewards/TranscriptCorrectionGrpoReward/mean": -5.960464477539063e-10,
1818
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
1819
+ "step": 3300
1820
+ },
1821
+ {
1822
+ "clip_ratio/high_max": 0.0,
1823
+ "clip_ratio/high_mean": 0.0,
1824
+ "clip_ratio/low_mean": 0.0,
1825
+ "clip_ratio/low_min": 0.0,
1826
+ "clip_ratio/region_mean": 0.0,
1827
+ "completion_length": 61.93,
1828
+ "completions/clipped_ratio": 0.0,
1829
+ "completions/max_length": 62.12,
1830
+ "completions/max_terminated_length": 62.12,
1831
+ "completions/mean_length": 61.93,
1832
+ "completions/mean_terminated_length": 61.93,
1833
+ "completions/min_length": 61.76,
1834
+ "completions/min_terminated_length": 61.76,
1835
+ "epoch": 0.005950160743148435,
1836
+ "frac_reward_zero_std": 0.6,
1837
+ "grad_norm": 2.3984543986443896e-06,
1838
+ "kl": 0.25049989838153125,
1839
+ "learning_rate": 3.717723851601874e-07,
1840
+ "loss": 0.0016641992330551148,
1841
+ "num_tokens": 3154800.0,
1842
+ "reward": -1.4901161193847657e-09,
1843
+ "reward_std": 0.42761797904968263,
1844
+ "rewards/TranscriptCorrectionGrpoReward/mean": -1.4901161193847657e-09,
1845
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
1846
+ "step": 3350
1847
+ },
1848
+ {
1849
+ "clip_ratio/high_max": 0.0,
1850
+ "clip_ratio/high_mean": 0.0,
1851
+ "clip_ratio/low_mean": 7.042253389954567e-05,
1852
+ "clip_ratio/low_min": 7.042253389954567e-05,
1853
+ "clip_ratio/region_mean": 7.042253389954567e-05,
1854
+ "completion_length": 56.81,
1855
+ "completions/clipped_ratio": 0.0,
1856
+ "completions/max_length": 57.0,
1857
+ "completions/max_terminated_length": 57.0,
1858
+ "completions/mean_length": 56.81,
1859
+ "completions/mean_terminated_length": 56.81,
1860
+ "completions/min_length": 56.6,
1861
+ "completions/min_terminated_length": 56.6,
1862
+ "epoch": 0.0060389691124491576,
1863
+ "frac_reward_zero_std": 0.52,
1864
+ "grad_norm": 0.4836759865283966,
1865
+ "kl": 0.3084017189592123,
1866
+ "learning_rate": 3.773228835949468e-07,
1867
+ "loss": 0.001545657515525818,
1868
+ "num_tokens": 3195802.0,
1869
+ "reward": 1.7881393432617187e-09,
1870
+ "reward_std": 0.5131415748596191,
1871
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
1872
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5131415748596191,
1873
+ "step": 3400
1874
+ },
1875
+ {
1876
+ "clip_ratio/high_max": 0.0,
1877
+ "clip_ratio/high_mean": 0.0,
1878
+ "clip_ratio/low_mean": 8.333333767950535e-05,
1879
+ "clip_ratio/low_min": 8.333333767950535e-05,
1880
+ "clip_ratio/region_mean": 8.333333767950535e-05,
1881
+ "completion_length": 66.025,
1882
+ "completions/clipped_ratio": 0.0,
1883
+ "completions/max_length": 66.28,
1884
+ "completions/max_terminated_length": 66.28,
1885
+ "completions/mean_length": 66.025,
1886
+ "completions/mean_terminated_length": 66.025,
1887
+ "completions/min_length": 65.84,
1888
+ "completions/min_terminated_length": 65.84,
1889
+ "epoch": 0.0061277774817498805,
1890
+ "frac_reward_zero_std": 0.56,
1891
+ "grad_norm": 0.003919742535799742,
1892
+ "kl": 0.23774921588599682,
1893
+ "learning_rate": 3.8287338202970627e-07,
1894
+ "loss": 0.0006137275695800781,
1895
+ "num_tokens": 3243103.0,
1896
+ "reward": 1.7881393432617187e-09,
1897
+ "reward_std": 0.4703797769546509,
1898
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
1899
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4703797769546509,
1900
+ "step": 3450
1901
+ },
1902
+ {
1903
+ "clip_ratio/high_max": 0.0,
1904
+ "clip_ratio/high_mean": 0.0,
1905
+ "clip_ratio/low_mean": 0.0,
1906
+ "clip_ratio/low_min": 0.0,
1907
+ "clip_ratio/region_mean": 0.0,
1908
+ "completion_length": 67.315,
1909
+ "completions/clipped_ratio": 0.0,
1910
+ "completions/max_length": 67.52,
1911
+ "completions/max_terminated_length": 67.52,
1912
+ "completions/mean_length": 67.315,
1913
+ "completions/mean_terminated_length": 67.315,
1914
+ "completions/min_length": 67.04,
1915
+ "completions/min_terminated_length": 67.04,
1916
+ "epoch": 0.006216585851050603,
1917
+ "frac_reward_zero_std": 0.64,
1918
+ "grad_norm": 2.6315128707210533e-06,
1919
+ "kl": 0.2313031293451786,
1920
+ "learning_rate": 3.884238804644658e-07,
1921
+ "loss": 0.0008652891218662262,
1922
+ "num_tokens": 3290574.0,
1923
+ "reward": 1.1920928955078125e-09,
1924
+ "reward_std": 0.3848561763763428,
1925
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.1920928955078125e-09,
1926
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.38485618591308596,
1927
+ "step": 3500
1928
+ },
1929
+ {
1930
+ "clip_ratio/high_max": 0.0,
1931
+ "clip_ratio/high_mean": 0.0,
1932
+ "clip_ratio/low_mean": 0.0,
1933
+ "clip_ratio/low_min": 0.0,
1934
+ "clip_ratio/region_mean": 0.0,
1935
+ "completion_length": 68.54,
1936
+ "completions/clipped_ratio": 0.0,
1937
+ "completions/max_length": 68.96,
1938
+ "completions/max_terminated_length": 68.96,
1939
+ "completions/mean_length": 68.54,
1940
+ "completions/mean_terminated_length": 68.54,
1941
+ "completions/min_length": 67.84,
1942
+ "completions/min_terminated_length": 67.84,
1943
+ "epoch": 0.0063053942203513255,
1944
+ "frac_reward_zero_std": 0.4,
1945
+ "grad_norm": 0.47953200340270996,
1946
+ "kl": 0.24469035990536214,
1947
+ "learning_rate": 3.9397437889922524e-07,
1948
+ "loss": -0.0004951600730419159,
1949
+ "num_tokens": 3338986.0,
1950
+ "reward": 3.2782554626464844e-09,
1951
+ "reward_std": 0.6414269638061524,
1952
+ "rewards/TranscriptCorrectionGrpoReward/mean": 3.2782554626464844e-09,
1953
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.641426968574524,
1954
+ "step": 3550
1955
+ },
1956
+ {
1957
+ "clip_ratio/high_max": 8.196720853447914e-05,
1958
+ "clip_ratio/high_mean": 8.196720853447914e-05,
1959
+ "clip_ratio/low_mean": 0.0,
1960
+ "clip_ratio/low_min": 0.0,
1961
+ "clip_ratio/region_mean": 8.196720853447914e-05,
1962
+ "completion_length": 63.63,
1963
+ "completions/clipped_ratio": 0.0,
1964
+ "completions/max_length": 63.96,
1965
+ "completions/max_terminated_length": 63.96,
1966
+ "completions/mean_length": 63.63,
1967
+ "completions/mean_terminated_length": 63.63,
1968
+ "completions/min_length": 63.32,
1969
+ "completions/min_terminated_length": 63.32,
1970
+ "epoch": 0.006394202589652048,
1971
+ "frac_reward_zero_std": 0.44,
1972
+ "grad_norm": 7.881725650804583e-06,
1973
+ "kl": 0.23511427730321885,
1974
+ "learning_rate": 3.9952487733398464e-07,
1975
+ "loss": 0.0018363992869853973,
1976
+ "num_tokens": 3384712.0,
1977
+ "reward": 2.9802322387695313e-10,
1978
+ "reward_std": 0.5986651659011841,
1979
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.9802322387695313e-10,
1980
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5986651706695557,
1981
+ "step": 3600
1982
+ },
1983
+ {
1984
+ "clip_ratio/high_max": 0.0,
1985
+ "clip_ratio/high_mean": 0.0,
1986
+ "clip_ratio/low_mean": 0.0,
1987
+ "clip_ratio/low_min": 0.0,
1988
+ "clip_ratio/region_mean": 0.0,
1989
+ "completion_length": 57.22,
1990
+ "completions/clipped_ratio": 0.0,
1991
+ "completions/max_length": 57.48,
1992
+ "completions/max_terminated_length": 57.48,
1993
+ "completions/mean_length": 57.22,
1994
+ "completions/mean_terminated_length": 57.22,
1995
+ "completions/min_length": 56.96,
1996
+ "completions/min_terminated_length": 56.96,
1997
+ "epoch": 0.006483010958952771,
1998
+ "frac_reward_zero_std": 0.6,
1999
+ "grad_norm": 0.00020814557501580566,
2000
+ "kl": 0.2519885949790478,
2001
+ "learning_rate": 4.050753757687441e-07,
2002
+ "loss": 0.0011325660347938538,
2003
+ "num_tokens": 3425972.0,
2004
+ "reward": 1.7881393432617187e-09,
2005
+ "reward_std": 0.42761797904968263,
2006
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.7881393432617187e-09,
2007
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.42761797904968263,
2008
+ "step": 3650
2009
+ },
2010
+ {
2011
+ "clip_ratio/high_max": 0.0,
2012
+ "clip_ratio/high_mean": 0.0,
2013
+ "clip_ratio/low_mean": 0.0,
2014
+ "clip_ratio/low_min": 0.0,
2015
+ "clip_ratio/region_mean": 0.0,
2016
+ "completion_length": 67.325,
2017
+ "completions/clipped_ratio": 0.0,
2018
+ "completions/max_length": 67.64,
2019
+ "completions/max_terminated_length": 67.64,
2020
+ "completions/mean_length": 67.325,
2021
+ "completions/mean_terminated_length": 67.325,
2022
+ "completions/min_length": 66.84,
2023
+ "completions/min_terminated_length": 66.84,
2024
+ "epoch": 0.006571819328253494,
2025
+ "frac_reward_zero_std": 0.52,
2026
+ "grad_norm": 0.20728643238544464,
2027
+ "kl": 0.24612667731940746,
2028
+ "learning_rate": 4.106258742035035e-07,
2029
+ "loss": 0.0013442845642566681,
2030
+ "num_tokens": 3473653.0,
2031
+ "reward": 3.5762786865234374e-09,
2032
+ "reward_std": 0.5131415748596191,
2033
+ "rewards/TranscriptCorrectionGrpoReward/mean": 3.5762786865234374e-09,
2034
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5131415748596191,
2035
+ "step": 3700
2036
+ },
2037
+ {
2038
+ "clip_ratio/high_max": 0.0,
2039
+ "clip_ratio/high_mean": 0.0,
2040
+ "clip_ratio/low_mean": 0.00011111111380159854,
2041
+ "clip_ratio/low_min": 0.00011111111380159854,
2042
+ "clip_ratio/region_mean": 0.00011111111380159854,
2043
+ "completion_length": 71.095,
2044
+ "completions/clipped_ratio": 0.0,
2045
+ "completions/max_length": 71.28,
2046
+ "completions/max_terminated_length": 71.28,
2047
+ "completions/mean_length": 71.095,
2048
+ "completions/mean_terminated_length": 71.095,
2049
+ "completions/min_length": 70.8,
2050
+ "completions/min_terminated_length": 70.8,
2051
+ "epoch": 0.006660627697554217,
2052
+ "frac_reward_zero_std": 0.48,
2053
+ "grad_norm": 2.862835935957264e-05,
2054
+ "kl": 0.2331434202194214,
2055
+ "learning_rate": 4.1617637263826297e-07,
2056
+ "loss": 0.0005902900546789169,
2057
+ "num_tokens": 3523592.0,
2058
+ "reward": 1.1920928955078125e-09,
2059
+ "reward_std": 0.5559033679962159,
2060
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.1920928955078125e-09,
2061
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5559033727645875,
2062
+ "step": 3750
2063
+ },
2064
+ {
2065
+ "clip_ratio/high_max": 0.0,
2066
+ "clip_ratio/high_mean": 0.0,
2067
+ "clip_ratio/low_mean": 4.7619049437344076e-05,
2068
+ "clip_ratio/low_min": 4.7619049437344076e-05,
2069
+ "clip_ratio/region_mean": 4.7619049437344076e-05,
2070
+ "completion_length": 67.965,
2071
+ "completions/clipped_ratio": 0.0,
2072
+ "completions/max_length": 68.08,
2073
+ "completions/max_terminated_length": 68.08,
2074
+ "completions/mean_length": 67.965,
2075
+ "completions/mean_terminated_length": 67.965,
2076
+ "completions/min_length": 67.84,
2077
+ "completions/min_terminated_length": 67.84,
2078
+ "epoch": 0.00674943606685494,
2079
+ "frac_reward_zero_std": 0.68,
2080
+ "grad_norm": 3.3257359177696344e-07,
2081
+ "kl": 0.23384407460689544,
2082
+ "learning_rate": 4.2172687107302243e-07,
2083
+ "loss": 0.0007511702179908752,
2084
+ "num_tokens": 3571849.0,
2085
+ "reward": -1.1920928955078125e-09,
2086
+ "reward_std": 0.3420943784713745,
2087
+ "rewards/TranscriptCorrectionGrpoReward/mean": -1.1920928955078125e-09,
2088
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
2089
+ "step": 3800
2090
+ },
2091
+ {
2092
+ "clip_ratio/high_max": 0.0,
2093
+ "clip_ratio/high_mean": 0.0,
2094
+ "clip_ratio/low_mean": 6.75675692036748e-05,
2095
+ "clip_ratio/low_min": 6.75675692036748e-05,
2096
+ "clip_ratio/region_mean": 6.75675692036748e-05,
2097
+ "completion_length": 67.4,
2098
+ "completions/clipped_ratio": 0.0,
2099
+ "completions/max_length": 67.6,
2100
+ "completions/max_terminated_length": 67.6,
2101
+ "completions/mean_length": 67.4,
2102
+ "completions/mean_terminated_length": 67.4,
2103
+ "completions/min_length": 67.12,
2104
+ "completions/min_terminated_length": 67.12,
2105
+ "epoch": 0.006838244436155663,
2106
+ "frac_reward_zero_std": 0.68,
2107
+ "grad_norm": 4.252907456248067e-05,
2108
+ "kl": 0.24010798826813698,
2109
+ "learning_rate": 4.2727736950778184e-07,
2110
+ "loss": 0.0015544074773788453,
2111
+ "num_tokens": 3620329.0,
2112
+ "reward": 2.682209014892578e-09,
2113
+ "reward_std": 0.3420943880081177,
2114
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.682209014892578e-09,
2115
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
2116
+ "step": 3850
2117
+ },
2118
+ {
2119
+ "clip_ratio/high_max": 0.0,
2120
+ "clip_ratio/high_mean": 0.0,
2121
+ "clip_ratio/low_mean": 0.0,
2122
+ "clip_ratio/low_min": 0.0,
2123
+ "clip_ratio/region_mean": 0.0,
2124
+ "completion_length": 67.125,
2125
+ "completions/clipped_ratio": 0.0,
2126
+ "completions/max_length": 67.36,
2127
+ "completions/max_terminated_length": 67.36,
2128
+ "completions/mean_length": 67.125,
2129
+ "completions/mean_terminated_length": 67.125,
2130
+ "completions/min_length": 66.96,
2131
+ "completions/min_terminated_length": 66.96,
2132
+ "epoch": 0.006927052805456386,
2133
+ "frac_reward_zero_std": 0.48,
2134
+ "grad_norm": 1.9069068002863787e-05,
2135
+ "kl": 0.25949006535112856,
2136
+ "learning_rate": 4.328278679425413e-07,
2137
+ "loss": 0.0014364664256572724,
2138
+ "num_tokens": 3667970.0,
2139
+ "reward": 8.940696716308593e-10,
2140
+ "reward_std": 0.5559033727645875,
2141
+ "rewards/TranscriptCorrectionGrpoReward/mean": 8.940696716308593e-10,
2142
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.5559033727645875,
2143
+ "step": 3900
2144
+ },
2145
+ {
2146
+ "clip_ratio/high_max": 0.0,
2147
+ "clip_ratio/high_mean": 0.0,
2148
+ "clip_ratio/low_mean": 0.0,
2149
+ "clip_ratio/low_min": 0.0,
2150
+ "clip_ratio/region_mean": 0.0,
2151
+ "completion_length": 63.395,
2152
+ "completions/clipped_ratio": 0.0,
2153
+ "completions/max_length": 63.56,
2154
+ "completions/max_terminated_length": 63.56,
2155
+ "completions/mean_length": 63.395,
2156
+ "completions/mean_terminated_length": 63.395,
2157
+ "completions/min_length": 63.2,
2158
+ "completions/min_terminated_length": 63.2,
2159
+ "epoch": 0.007015861174757109,
2160
+ "frac_reward_zero_std": 0.76,
2161
+ "grad_norm": 0.0010645861038938165,
2162
+ "kl": 0.2909430082514882,
2163
+ "learning_rate": 4.383783663773007e-07,
2164
+ "loss": 0.0015060535073280334,
2165
+ "num_tokens": 3713369.0,
2166
+ "reward": 2.2351741790771484e-09,
2167
+ "reward_std": 0.25657079219818113,
2168
+ "rewards/TranscriptCorrectionGrpoReward/mean": 2.2351741790771484e-09,
2169
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.25657078742980954,
2170
+ "step": 3950
2171
+ },
2172
+ {
2173
+ "clip_ratio/high_max": 0.0,
2174
+ "clip_ratio/high_mean": 0.0,
2175
+ "clip_ratio/low_mean": 9.615384973585606e-05,
2176
+ "clip_ratio/low_min": 9.615384973585606e-05,
2177
+ "clip_ratio/region_mean": 9.615384973585606e-05,
2178
+ "completion_length": 64.495,
2179
+ "completions/clipped_ratio": 0.0,
2180
+ "completions/max_length": 64.68,
2181
+ "completions/max_terminated_length": 64.68,
2182
+ "completions/mean_length": 64.495,
2183
+ "completions/mean_terminated_length": 64.495,
2184
+ "completions/min_length": 64.4,
2185
+ "completions/min_terminated_length": 64.4,
2186
+ "epoch": 0.007104669544057832,
2187
+ "frac_reward_zero_std": 0.68,
2188
+ "grad_norm": 0.0001376425352646038,
2189
+ "kl": 0.2843581172823906,
2190
+ "learning_rate": 4.4392886481206016e-07,
2191
+ "loss": 0.0012637190520763397,
2192
+ "num_tokens": 3759740.0,
2193
+ "reward": 1.4901161193847657e-09,
2194
+ "reward_std": 0.3420943832397461,
2195
+ "rewards/TranscriptCorrectionGrpoReward/mean": 1.4901161193847657e-09,
2196
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.3420943832397461,
2197
+ "step": 4000
2198
+ },
2199
+ {
2200
+ "clip_ratio/high_max": 0.0,
2201
+ "clip_ratio/high_mean": 0.0,
2202
+ "clip_ratio/low_mean": 7.936508394777775e-05,
2203
+ "clip_ratio/low_min": 7.936508394777775e-05,
2204
+ "clip_ratio/region_mean": 7.936508394777775e-05,
2205
+ "completion_length": 68.475,
2206
+ "completions/clipped_ratio": 0.0,
2207
+ "completions/max_length": 68.68,
2208
+ "completions/max_terminated_length": 68.68,
2209
+ "completions/mean_length": 68.475,
2210
+ "completions/mean_terminated_length": 68.475,
2211
+ "completions/min_length": 68.28,
2212
+ "completions/min_terminated_length": 68.28,
2213
+ "epoch": 0.007193477913358555,
2214
+ "frac_reward_zero_std": 0.6,
2215
+ "grad_norm": 0.048516687005758286,
2216
+ "kl": 0.2665792652219534,
2217
+ "learning_rate": 4.4947936324681957e-07,
2218
+ "loss": 0.0013603484630584717,
2219
+ "num_tokens": 3808163.0,
2220
+ "reward": -8.940696716308593e-10,
2221
+ "reward_std": 0.42761797904968263,
2222
+ "rewards/TranscriptCorrectionGrpoReward/mean": -8.940696716308593e-10,
2223
+ "rewards/TranscriptCorrectionGrpoReward/std": 0.4276179838180542,
2224
+ "step": 4050
2225
+ },
2226
+ {
2227
+ "eval_cer_subset": 0.012702134935756509,
2228
+ "eval_cer_subset_edit_distance": 780,
2229
+ "eval_cer_subset_groups": 250,
2230
+ "eval_cer_subset_items": 250,
2231
+ "eval_cer_subset_ref_chars": 61407,
2232
+ "eval_loss": 0.04000633512828635,
2233
+ "step": 4052
2234
+ }
2235
+ ],
2236
+ "logging_steps": 50,
2237
+ "max_steps": 2252040,
2238
+ "num_input_tokens_seen": 3809971,
2239
+ "num_train_epochs": 4,
2240
+ "save_steps": 1013,
2241
+ "stateful_callbacks": {
2242
+ "TrainerControl": {
2243
+ "args": {
2244
+ "should_epoch_stop": false,
2245
+ "should_evaluate": false,
2246
+ "should_log": false,
2247
+ "should_save": true,
2248
+ "should_training_stop": false
2249
+ },
2250
+ "attributes": {}
2251
+ }
2252
+ },
2253
+ "total_flos": 0.0,
2254
+ "train_batch_size": 1,
2255
+ "trial_name": null,
2256
+ "trial_params": null
2257
+ }
meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:142af57cbc15b9beb117b5972d297aef1f92226b896a513335c81b0f2c19ecf7
3
+ size 6737
meta-llama__llama-3.2-1b/grpo/run_manifest.json CHANGED
@@ -9,7 +9,7 @@
9
  "best_metric_value": 0.012620710993860635,
10
  "best_step": 2026,
11
  "best_checkpoint_step": 2026,
12
- "latest_checkpoint_step": 3039,
13
  "checkpoints": [
14
  {
15
  "step": 1013,
@@ -28,10 +28,16 @@
28
  "epoch": 0.0053977726860979375,
29
  "repo_path": "meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-3039",
30
  "timestamp": "2026-04-09T04:44:46.465728+00:00"
 
 
 
 
 
 
31
  }
32
  ],
33
  "epoch_artifacts": [],
34
  "resume_history": [],
35
  "created_at": "2026-04-09T04:05:57.752019+00:00",
36
- "updated_at": "2026-04-09T04:44:46.465745+00:00"
37
  }
 
9
  "best_metric_value": 0.012620710993860635,
10
  "best_step": 2026,
11
  "best_checkpoint_step": 2026,
12
+ "latest_checkpoint_step": 4052,
13
  "checkpoints": [
14
  {
15
  "step": 1013,
 
28
  "epoch": 0.0053977726860979375,
29
  "repo_path": "meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-3039",
30
  "timestamp": "2026-04-09T04:44:46.465728+00:00"
31
+ },
32
+ {
33
+ "step": 4052,
34
+ "epoch": 0.007197030248130584,
35
+ "repo_path": "meta-llama__llama-3.2-1b/grpo/checkpoints/checkpoint-4052",
36
+ "timestamp": "2026-04-09T04:57:36.308853+00:00"
37
  }
38
  ],
39
  "epoch_artifacts": [],
40
  "resume_history": [],
41
  "created_at": "2026-04-09T04:05:57.752019+00:00",
42
+ "updated_at": "2026-04-09T04:57:36.308869+00:00"
43
  }