Upload checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins

checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/wandb/offline-run-20260129_223311-checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins-run0/files/config.yaml

@@ -0,0 +1,459 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    python_version: 3.11.10
+    cli_version: 0.23.1
+    framework: huggingface
+    huggingface_version: 4.49.0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    start_time: 1769725992
+    t:
+      1:
+      - 1
+      - 5
+      - 11
+      - 41
+      - 49
+      - 53
+      - 71
+      - 105
+      2:
+      - 1
+      - 5
+      - 11
+      - 41
+      - 49
+      - 53
+      - 71
+      - 105
+      3:
+      - 2
+      - 4
+      - 13
+      - 14
+      - 37
+      - 42
+      - 61
+      4: 3.11.10
+      5: 0.23.1
+      6: 4.49.0
+      13: linux-x86_64
+    e:
+      wfvoftkdc4um9761m8ix7oiachlmdvuv:
+        os: Linux-6.6.93+-x86_64-with-glibc2.35
+        python: CPython 3.11.10
+        started_at: '2026-01-29T22:33:11.859047Z'
+        args:
+        - --dataset_config_file
+        - ./data/configs/vlm_gym_reference_dot_train_celoss_no_mse.yaml
+        - --eval_dataset_config_file
+        - ./data/configs/vlm_gym_reference_dot_eval_celoss_no_mse.yaml
+        - --viz_dataset_config_file
+        - ./data/configs/vlm_gym_reference_dot_eval_celoss_no_mse.yaml
+        - --train_data_dir
+        - /home/clouduser/Code/data/gym/reference_dot_v5/train/
+        - --train_jsonl_path
+        - /home/clouduser/Code/data/gym/reference_dot_v5/train/
+        - --eval_data_dir
+        - /home/clouduser/Code/data/gym/reference_dot_v5/val/
+        - --eval_jsonl_path
+        - /home/clouduser/Code/data/gym/reference_dot_v5/val/
+        - --inference_hash_file
+        - /home/clouduser/Code/Github/launch_new/hashes_test_set_v10.json
+        - --task_name
+        - reference_dot_v5
+        - --instructions_dir
+        - ./data/instructions
+        - --model_path
+        - /home/clouduser/Code/Models/BAGEL-7B-MoT
+        - --layer_module
+        - Qwen2MoTDecoderLayer
+        - --max_latent_size
+        - '64'
+        - --resume-from
+        - /home/clouduser/Code/Models/BAGEL-7B-MoT
+        - --finetune_from_hf
+        - 'True'
+        - --auto_resume
+        - 'False'
+        - --resume-model-only
+        - 'True'
+        - --finetune-from-ema
+        - 'True'
+        - --log_every
+        - '1'
+        - --lr
+        - 2e-5
+        - --warmup_steps
+        - '300'
+        - --lr_scheduler
+        - cosine
+        - --num_worker
+        - '1'
+        - --expected_num_tokens
+        - '30000'
+        - --max_num_tokens
+        - '30000'
+        - --max_num_tokens_per_sample
+        - '30000'
+        - --visual_und
+        - 'True'
+        - --visual_gen
+        - 'False'
+        - --save_every
+        - '5000'
+        - --total_steps
+        - '5000'
+        - --text_cond_dropout_prob
+        - '0.0'
+        - --vae_cond_dropout_prob
+        - '0.3'
+        - --vit_cond_dropout_prob
+        - '0.0'
+        - --ema
+        - '0.993'
+        - --checkpoint_dir
+        - /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins
+        - --wandb_project
+        - bagel
+        - --wandb_name
+        - checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins
+        - --wandb_dir
+        - /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins
+        - --wandb_offline
+        - 'True'
+        program: /home/clouduser/Code/Github/unified_world_model/train/pretrain_unified_navit.py
+        code_path: train/pretrain_unified_navit.py
+        code_path_local: train/pretrain_unified_navit.py
+        git:
+          remote_url: https://github.com/para-lost/unified_world_model
+          commit: 8d7b26b7e552fc87b592cf3be94d93be7aeca2a9
+        root: /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins
+        host: junyizhang-launch-new-226785958-1-0
+        executable: /opt/conda/bin/python3.11
+        cpu_count: 48
+        cpu_count_logical: 96
+        gpu_type: NVIDIA A100-SXM4-80GB
+        gpu_count: 8
+        disk:
+          /:
+            total: '1052461830144'
+            used: '164471345152'
+        memory:
+          total: '1437332606976'
+        gpu_nvidia:
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-27013fed-9784-d445-a1eb-01629cf403cc
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-c4922cf6-bc87-9458-c12f-23210cb43686
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-1af9405a-c062-486e-383f-7ea6c6ef5158
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-793b7211-7436-7429-8bd7-cc05be70cc75
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-5eb44009-8d7d-911d-0730-f219cb50498c
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-62c85054-47c8-b915-18e9-e4433fc0f9bb
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-c3b59f2c-b6b6-7730-54ff-8cf5fee4ea9c
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-e988baaf-6bc5-3bb9-91fb-ab2cb214233d
+        cuda_version: '12.2'
+        writer_id: wfvoftkdc4um9761m8ix7oiachlmdvuv
+visual_gen:
+  desc: null
+  value: false
+visual_und:
+  desc: null
+  value: true
+results_dir:
+  desc: null
+  value: results
+checkpoint_dir:
+  desc: null
+  value: /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins
+wandb_project:
+  desc: null
+  value: bagel
+wandb_name:
+  desc: null
+  value: checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins
+wandb_runid:
+  desc: null
+  value: '0'
+wandb_resume:
+  desc: null
+  value: allow
+wandb_offline:
+  desc: null
+  value: true
+wandb_dir:
+  desc: null
+  value: /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins
+global_seed:
+  desc: null
+  value: 4396
+auto_resume:
+  desc: null
+  value: false
+resume_from:
+  desc: null
+  value: /home/clouduser/Code/Models/BAGEL-7B-MoT
+resume_model_only:
+  desc: null
+  value: true
+finetune_from_ema:
+  desc: null
+  value: true
+finetune_from_hf:
+  desc: null
+  value: true
+log_every:
+  desc: null
+  value: 1
+save_every:
+  desc: null
+  value: 5000
+total_steps:
+  desc: null
+  value: 5000
+warmup_steps:
+  desc: null
+  value: 300
+lr_scheduler:
+  desc: null
+  value: cosine
+lr:
+  desc: null
+  value: 2.0e-05
+min_lr:
+  desc: null
+  value: 1.0e-07
+beta1:
+  desc: null
+  value: 0.9
+beta2:
+  desc: null
+  value: 0.95
+eps:
+  desc: null
+  value: 1.0e-15
+ema:
+  desc: null
+  value: 0.993
+max_grad_norm:
+  desc: null
+  value: 1.0
+timestep_shift:
+  desc: null
+  value: 1.0
+mse_weight:
+  desc: null
+  value: 1.0
+ce_weight:
+  desc: null
+  value: 1.0
+ce_loss_reweighting:
+  desc: null
+  value: false
+expected_num_tokens:
+  desc: null
+  value: 30000
+num_replicate:
+  desc: null
+  value: 1
+num_shard:
+  desc: null
+  value: 8
+sharding_strategy:
+  desc: null
+  value: HYBRID_SHARD
+backward_prefetch:
+  desc: null
+  value: BACKWARD_PRE
+cpu_offload:
+  desc: null
+  value: false
+freeze_llm:
+  desc: null
+  value: false
+freeze_vit:
+  desc: null
+  value: false
+freeze_vae:
+  desc: null
+  value: true
+freeze_und:
+  desc: null
+  value: false
+copy_init_moe:
+  desc: null
+  value: true
+use_flex:
+  desc: null
+  value: false
+eval_every:
+  desc: null
+  value: 500
+num_eval_batches:
+  desc: null
+  value: 20
+use_ema_for_eval:
+  desc: null
+  value: true
+eval_log_dir:
+  desc: null
+  value: null
+eval_run_tag:
+  desc: null
+  value: ''
+viz_every:
+  desc: null
+  value: 500
+viz_n:
+  desc: null
+  value: 8
+viz_outdir:
+  desc: null
+  value: results/viz
+eval_dataset_config_file:
+  desc: null
+  value: ./data/configs/vlm_gym_reference_dot_eval_celoss_no_mse.yaml
+viz_dataset_config_file:
+  desc: null
+  value: ./data/configs/vlm_gym_reference_dot_eval_celoss_no_mse.yaml
+eval_print_n:
+  desc: null
+  value: 3
+save_ema_only:
+  desc: null
+  value: true
+save_optimizer:
+  desc: null
+  value: false
+model_path:
+  desc: null
+  value: /home/clouduser/Code/Models/BAGEL-7B-MoT
+llm_path:
+  desc: null
+  value: hf/Qwen2.5-0.5B-Instruct/
+llm_qk_norm:
+  desc: null
+  value: true
+tie_word_embeddings:
+  desc: null
+  value: false
+layer_module:
+  desc: null
+  value: Qwen2MoTDecoderLayer
+vae_path:
+  desc: null
+  value: flux/vae/ae.safetensors
+vit_path:
+  desc: null
+  value: hf/siglip-so400m-14-980-flash-attn2-navit/
+max_latent_size:
+  desc: null
+  value: 64
+latent_patch_size:
+  desc: null
+  value: 2
+vit_patch_size:
+  desc: null
+  value: 14
+vit_max_num_patch_per_side:
+  desc: null
+  value: 70
+connector_act:
+  desc: null
+  value: gelu_pytorch_tanh
+interpolate_pos:
+  desc: null
+  value: false
+vit_select_layer:
+  desc: null
+  value: -2
+vit_rope:
+  desc: null
+  value: false
+text_cond_dropout_prob:
+  desc: null
+  value: 0.0
+vae_cond_dropout_prob:
+  desc: null
+  value: 0.3
+vit_cond_dropout_prob:
+  desc: null
+  value: 0.0
+dataset_config_file:
+  desc: null
+  value: ./data/configs/vlm_gym_reference_dot_train_celoss_no_mse.yaml
+train_data_dir:
+  desc: null
+  value: /home/clouduser/Code/data/gym/reference_dot_v5/train/
+train_jsonl_path:
+  desc: null
+  value: /home/clouduser/Code/data/gym/reference_dot_v5/train/
+eval_data_dir:
+  desc: null
+  value: /home/clouduser/Code/data/gym/reference_dot_v5/val/
+eval_jsonl_path:
+  desc: null
+  value: /home/clouduser/Code/data/gym/reference_dot_v5/val/
+inference_hash_file:
+  desc: null
+  value: /home/clouduser/Code/Github/launch_new/hashes_test_set_v10.json
+task_name:
+  desc: null
+  value: reference_dot_v5
+instructions_dir:
+  desc: null
+  value: ./data/instructions
+prefetch_factor:
+  desc: null
+  value: 2
+num_workers:
+  desc: null
+  value: 1
+max_num_tokens_per_sample:
+  desc: null
+  value: 30000
+max_num_tokens:
+  desc: null
+  value: 30000
+prefer_buffer_before:
+  desc: null
+  value: 16384
+max_buffer_size:
+  desc: null
+  value: 50
+data_seed:
+  desc: null
+  value: 42

checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/wandb/offline-run-20260129_223311-checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins-run0/files/output.log

@@ -1,180 +1,3 @@
-FullyShardedDataParallel(
-  (_fsdp_wrapped_module): Bagel(
-    (language_model): Qwen2ForCausalLM(
-      (model): Qwen2Model(
-        (embed_tokens): Embedding(152064, 3584)
-        (layers): ModuleList(
-          (0-27): 28 x FullyShardedDataParallel(
-            (_fsdp_wrapped_module): CheckpointWrapper(
-              (_checkpoint_wrapped_module): Qwen2MoTDecoderLayer(
-                (self_attn): PackedAttentionMoT(
-                  (q_proj): Linear(in_features=3584, out_features=3584, bias=True)
-                  (k_proj): Linear(in_features=3584, out_features=512, bias=True)
-                  (v_proj): Linear(in_features=3584, out_features=512, bias=True)
-                  (o_proj): Linear(in_features=3584, out_features=3584, bias=False)
-                  (q_norm): Qwen2RMSNorm((128,), eps=1e-06)
-                  (k_norm): Qwen2RMSNorm((128,), eps=1e-06)
-                  (q_norm_moe_gen): Qwen2RMSNorm((128,), eps=1e-06)
-                  (k_norm_moe_gen): Qwen2RMSNorm((128,), eps=1e-06)
-                  (q_proj_moe_gen): Linear(in_features=3584, out_features=3584, bias=True)
-                  (k_proj_moe_gen): Linear(in_features=3584, out_features=512, bias=True)
-                  (v_proj_moe_gen): Linear(in_features=3584, out_features=512, bias=True)
-                  (o_proj_moe_gen): Linear(in_features=3584, out_features=3584, bias=False)
-                )
-                (mlp): Qwen2MLP(
-                  (gate_proj): Linear(in_features=3584, out_features=18944, bias=False)
-                  (up_proj): Linear(in_features=3584, out_features=18944, bias=False)
-                  (down_proj): Linear(in_features=18944, out_features=3584, bias=False)
-                  (act_fn): SiLU()
-                )
-                (mlp_moe_gen): Qwen2MLP(
-                  (gate_proj): Linear(in_features=3584, out_features=18944, bias=False)
-                  (up_proj): Linear(in_features=3584, out_features=18944, bias=False)
-                  (down_proj): Linear(in_features=18944, out_features=3584, bias=False)
-                  (act_fn): SiLU()
-                )
-                (input_layernorm): Qwen2RMSNorm((3584,), eps=1e-06)
-                (input_layernorm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
-                (post_attention_layernorm): Qwen2RMSNorm((3584,), eps=1e-06)
-                (post_attention_layernorm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
-              )
-            )
-          )
-        )
-        (norm): Qwen2RMSNorm((3584,), eps=1e-06)
-        (norm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
-        (rotary_emb): Qwen2RotaryEmbedding()
-      )
-      (lm_head): Linear(in_features=3584, out_features=152064, bias=False)
-    )
-    (vit_model): SiglipVisionModel(
-      (vision_model): FullyShardedDataParallel(
-        (_fsdp_wrapped_module): SiglipVisionTransformer(
-          (embeddings): SiglipVisionEmbeddings(
-            (position_embedding): Embedding(4900, 1152)
-            (patch_embedding): Linear(in_features=588, out_features=1152, bias=True)
-          )
-          (encoder): SiglipEncoder(
-            (layers): ModuleList(
-              (0-25): 26 x FullyShardedDataParallel(
-                (_fsdp_wrapped_module): CheckpointWrapper(
-                  (_checkpoint_wrapped_module): SiglipEncoderLayer(
-                    (self_attn): SiglipFlashAttention2(
-                      (k_proj): Linear(in_features=1152, out_features=1152, bias=True)
-                      (v_proj): Linear(in_features=1152, out_features=1152, bias=True)
-                      (q_proj): Linear(in_features=1152, out_features=1152, bias=True)
-                      (out_proj): Linear(in_features=1152, out_features=1152, bias=True)
-                    )
-                    (layer_norm1): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
-                    (mlp): SiglipMLP(
-                      (activation_fn): PytorchGELUTanh()
-                      (fc1): Linear(in_features=1152, out_features=4304, bias=True)
-                      (fc2): Linear(in_features=4304, out_features=1152, bias=True)
-                    )
-                    (layer_norm2): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
-                  )
-                )
-              )
-            )
-          )
-          (post_layernorm): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
-        )
-      )
-    )
-    (connector): FullyShardedDataParallel(
-      (_fsdp_wrapped_module): CheckpointWrapper(
-        (_checkpoint_wrapped_module): MLPconnector(
-          (activation_fn): PytorchGELUTanh()
-          (fc1): Linear(in_features=1152, out_features=3584, bias=True)
-          (fc2): Linear(in_features=3584, out_features=3584, bias=True)
-        )
-      )
-    )
-    (vit_pos_embed): FullyShardedDataParallel(
-      (_fsdp_wrapped_module): PositionEmbedding()
-    )
-  )
-)
-_flat_param True
-language_model.model.layers.0._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.1._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.2._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.3._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.4._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.5._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.6._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.7._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.8._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.9._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.10._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.11._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.12._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.13._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.14._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.15._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.16._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.17._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.18._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.19._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.20._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.21._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.22._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.23._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.24._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.25._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.26._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.27._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.0._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.1._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.2._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.3._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.4._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.5._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.6._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.7._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.8._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.9._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.10._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.11._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.12._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.13._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.14._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.15._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.16._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.17._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.18._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.19._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.20._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.21._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.22._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.23._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.24._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.25._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-connector._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_pos_embed._fsdp_wrapped_module._flat_param False
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse/vlm_gym_reference_dot_train
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step0
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-ce_avg: 1.1923778057098389, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step500
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-ce_avg: 0.3785431981086731, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step1000
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-ce_avg: 0.40663474798202515, mse_avg: 0.0
 wandb: Detected [huggingface_hub.inference] in use.
 wandb: Use W&B Weave for improved LLM call tracing. Install Weave with `pip install weave` then add `import weave` to the top of your script.
 wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
@@ -1137,6 +960,197 @@ wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
 [2026-01-29 22:58:42] (step=0000949) Train Loss mse: 0.0000, Train Loss ce: 0.3724, Train Steps/Sec: 0.90,
 [2026-01-29 22:58:44] (step=0000950) Train Loss mse: 0.0000, Train Loss ce: 0.3805, Train Steps/Sec: 0.89,
 [2026-01-29 22:58:45] (step=0000951) Train Loss mse: 0.0000, Train Loss ce: 0.3712, Train Steps/Sec: 0.84,
+FullyShardedDataParallel(
+  (_fsdp_wrapped_module): Bagel(
+    (language_model): Qwen2ForCausalLM(
+      (model): Qwen2Model(
+        (embed_tokens): Embedding(152064, 3584)
+        (layers): ModuleList(
+          (0-27): 28 x FullyShardedDataParallel(
+            (_fsdp_wrapped_module): CheckpointWrapper(
+              (_checkpoint_wrapped_module): Qwen2MoTDecoderLayer(
+                (self_attn): PackedAttentionMoT(
+                  (q_proj): Linear(in_features=3584, out_features=3584, bias=True)
+                  (k_proj): Linear(in_features=3584, out_features=512, bias=True)
+                  (v_proj): Linear(in_features=3584, out_features=512, bias=True)
+                  (o_proj): Linear(in_features=3584, out_features=3584, bias=False)
+                  (q_norm): Qwen2RMSNorm((128,), eps=1e-06)
+                  (k_norm): Qwen2RMSNorm((128,), eps=1e-06)
+                  (q_norm_moe_gen): Qwen2RMSNorm((128,), eps=1e-06)
+                  (k_norm_moe_gen): Qwen2RMSNorm((128,), eps=1e-06)
+                  (q_proj_moe_gen): Linear(in_features=3584, out_features=3584, bias=True)
+                  (k_proj_moe_gen): Linear(in_features=3584, out_features=512, bias=True)
+                  (v_proj_moe_gen): Linear(in_features=3584, out_features=512, bias=True)
+                  (o_proj_moe_gen): Linear(in_features=3584, out_features=3584, bias=False)
+                )
+                (mlp): Qwen2MLP(
+                  (gate_proj): Linear(in_features=3584, out_features=18944, bias=False)
+                  (up_proj): Linear(in_features=3584, out_features=18944, bias=False)
+                  (down_proj): Linear(in_features=18944, out_features=3584, bias=False)
+                  (act_fn): SiLU()
+                )
+                (mlp_moe_gen): Qwen2MLP(
+                  (gate_proj): Linear(in_features=3584, out_features=18944, bias=False)
+                  (up_proj): Linear(in_features=3584, out_features=18944, bias=False)
+                  (down_proj): Linear(in_features=18944, out_features=3584, bias=False)
+                  (act_fn): SiLU()
+                )
+                (input_layernorm): Qwen2RMSNorm((3584,), eps=1e-06)
+                (input_layernorm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
+                (post_attention_layernorm): Qwen2RMSNorm((3584,), eps=1e-06)
+                (post_attention_layernorm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
+              )
+            )
+          )
+        )
+        (norm): Qwen2RMSNorm((3584,), eps=1e-06)
+        (norm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
+        (rotary_emb): Qwen2RotaryEmbedding()
+      )
+      (lm_head): Linear(in_features=3584, out_features=152064, bias=False)
+    )
+    (vit_model): SiglipVisionModel(
+      (vision_model): FullyShardedDataParallel(
+        (_fsdp_wrapped_module): SiglipVisionTransformer(
+          (embeddings): SiglipVisionEmbeddings(
+            (position_embedding): Embedding(4900, 1152)
+            (patch_embedding): Linear(in_features=588, out_features=1152, bias=True)
+          )
+          (encoder): SiglipEncoder(
+            (layers): ModuleList(
+              (0-25): 26 x FullyShardedDataParallel(
+                (_fsdp_wrapped_module): CheckpointWrapper(
+                  (_checkpoint_wrapped_module): SiglipEncoderLayer(
+                    (self_attn): SiglipFlashAttention2(
+                      (k_proj): Linear(in_features=1152, out_features=1152, bias=True)
+                      (v_proj): Linear(in_features=1152, out_features=1152, bias=True)
+                      (q_proj): Linear(in_features=1152, out_features=1152, bias=True)
+                      (out_proj): Linear(in_features=1152, out_features=1152, bias=True)
+                    )
+                    (layer_norm1): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
+                    (mlp): SiglipMLP(
+                      (activation_fn): PytorchGELUTanh()
+                      (fc1): Linear(in_features=1152, out_features=4304, bias=True)
+                      (fc2): Linear(in_features=4304, out_features=1152, bias=True)
+                    )
+                    (layer_norm2): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
+                  )
+                )
+              )
+            )
+          )
+          (post_layernorm): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
+        )
+      )
+    )
+    (connector): FullyShardedDataParallel(
+      (_fsdp_wrapped_module): CheckpointWrapper(
+        (_checkpoint_wrapped_module): MLPconnector(
+          (activation_fn): PytorchGELUTanh()
+          (fc1): Linear(in_features=1152, out_features=3584, bias=True)
+          (fc2): Linear(in_features=3584, out_features=3584, bias=True)
+        )
+      )
+    )
+    (vit_pos_embed): FullyShardedDataParallel(
+      (_fsdp_wrapped_module): PositionEmbedding()
+    )
+  )
+)
+_flat_param True
+language_model.model.layers.0._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.1._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.2._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.3._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.4._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.5._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.6._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.7._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.8._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.9._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.10._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.11._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.12._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.13._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.14._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.15._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.16._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.17._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.18._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.19._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.20._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.21._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.22._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.23._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.24._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.25._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.26._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.27._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.0._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.1._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.2._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.3._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.4._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.5._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.6._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.7._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.8._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.9._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.10._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.11._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.12._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.13._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.14._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.15._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.16._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.17._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.18._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.19._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.20._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.21._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.22._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.23._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.24._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.25._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+connector._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_pos_embed._fsdp_wrapped_module._flat_param False
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse/vlm_gym_reference_dot_train
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step0
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 1.1923778057098389, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step500
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 0.3785431981086731, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step1000
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 0.40663474798202515, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step1500
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 0.5510271191596985, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step2000
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 0.5553194880485535, mse_avg: 0.0
 [2026-01-29 22:58:46] (step=0000952) Train Loss mse: 0.0000, Train Loss ce: 0.3810, Train Steps/Sec: 0.75,
 [2026-01-29 22:58:47] (step=0000953) Train Loss mse: 0.0000, Train Loss ce: 0.3781, Train Steps/Sec: 0.84,
 [2026-01-29 22:58:48] (step=0000954) Train Loss mse: 0.0000, Train Loss ce: 0.3897, Train Steps/Sec: 0.87,
@@ -1390,20 +1404,6 @@ wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
 [2026-01-29 23:04:00] (step=0001202) Train Loss mse: 0.0000, Train Loss ce: 0.3769, Train Steps/Sec: 0.80,
 [2026-01-29 23:04:01] (step=0001203) Train Loss mse: 0.0000, Train Loss ce: 0.3707, Train Steps/Sec: 0.89,
 [2026-01-29 23:04:03] (step=0001204) Train Loss mse: 0.0000, Train Loss ce: 0.3861, Train Steps/Sec: 0.57,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step1500
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-ce_avg: 0.5510271191596985, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step2000
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-ce_avg: 0.5553194880485535, mse_avg: 0.0
 [2026-01-29 23:04:04] (step=0001205) Train Loss mse: 0.0000, Train Loss ce: 0.3762, Train Steps/Sec: 0.71,
 [2026-01-29 23:04:06] (step=0001206) Train Loss mse: 0.0000, Train Loss ce: 0.3861, Train Steps/Sec: 0.84,
 [2026-01-29 23:04:07] (step=0001207) Train Loss mse: 0.0000, Train Loss ce: 0.3744, Train Steps/Sec: 0.90,
@@ -2477,6 +2477,27 @@ ce_avg: 0.5553194880485535, mse_avg: 0.0
 [2026-01-29 23:26:13] (step=0002275) Train Loss mse: 0.0000, Train Loss ce: 0.3819, Train Steps/Sec: 0.70,
 [2026-01-29 23:26:14] (step=0002276) Train Loss mse: 0.0000, Train Loss ce: 0.3862, Train Steps/Sec: 0.85,
 [2026-01-29 23:26:15] (step=0002277) Train Loss mse: 0.0000, Train Loss ce: 0.3689, Train Steps/Sec: 0.86,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step2500
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 0.9706690311431885, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step3000
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 1.6775857210159302, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step3500
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 2.2602760791778564, mse_avg: 0.0
 [2026-01-29 23:26:17] (step=0002278) Train Loss mse: 0.0000, Train Loss ce: 0.3637, Train Steps/Sec: 0.88,
 [2026-01-29 23:26:18] (step=0002279) Train Loss mse: 0.0000, Train Loss ce: 0.3588, Train Steps/Sec: 0.84,
 [2026-01-29 23:26:19] (step=0002280) Train Loss mse: 0.0000, Train Loss ce: 0.3704, Train Steps/Sec: 0.70,
@@ -2915,27 +2936,6 @@ ce_avg: 0.5553194880485535, mse_avg: 0.0
 [2026-01-29 23:35:26] (step=0002713) Train Loss mse: 0.0000, Train Loss ce: 0.3485, Train Steps/Sec: 0.84,
 [2026-01-29 23:35:27] (step=0002714) Train Loss mse: 0.0000, Train Loss ce: 0.3643, Train Steps/Sec: 0.90,
 [2026-01-29 23:35:29] (step=0002715) Train Loss mse: 0.0000, Train Loss ce: 0.3573, Train Steps/Sec: 0.87,
-[2026-01-29 23:35:30
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step2500
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-ce_avg: 0.9706690311431885, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step3000
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-ce_avg: 1.6775857210159302, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step3500
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
 [2026-01-29 23:35:30] (step=0002716) Train Loss mse: 0.0000, Train Loss ce: 0.3701, Train Steps/Sec: 0.85,
 [2026-01-29 23:35:31] (step=0002717) Train Loss mse: 0.0000, Train Loss ce: 0.3544, Train Steps/Sec: 0.86,
 [2026-01-29 23:35:32] (step=0002718) Train Loss mse: 0.0000, Train Loss ce: 0.3516, Train Steps/Sec: 0.89,
@@ -3522,6 +3522,20 @@ Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference
 [2026-01-29 23:47:48] (step=0003299) Train Loss mse: 0.0000, Train Loss ce: 0.3641, Train Steps/Sec: 0.67,
 [2026-01-29 23:47:49] (step=0003300) Train Loss mse: 0.0000, Train Loss ce: 0.3764, Train Steps/Sec: 0.89,
 [2026-01-29 23:47:51] (step=0003301) Train Loss mse: 0.0000, Train Loss ce: 0.3725, Train Steps/Sec: 0.72,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step4000
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 2.886735677719116, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step4500
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 3.8481361865997314, mse_avg: 0.0
 [2026-01-29 23:47:52] (step=0003302) Train Loss mse: 0.0000, Train Loss ce: 0.3544, Train Steps/Sec: 0.90,
 [2026-01-29 23:47:53] (step=0003303) Train Loss mse: 0.0000, Train Loss ce: 0.3668, Train Steps/Sec: 0.84,
 [2026-01-29 23:47:54] (step=0003304) Train Loss mse: 0.0000, Train Loss ce: 0.3708, Train Steps/Sec: 0.71,
@@ -4278,27 +4292,105 @@ Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference
 [2026-01-30 00:03:28] (step=0004055) Train Loss mse: 0.0000, Train Loss ce: 0.3722, Train Steps/Sec: 0.89,
 [2026-01-30 00:03:29] (step=0004056) Train Loss mse: 0.0000, Train Loss ce: 0.3767, Train Steps/Sec: 0.84,
 [2026-01-30 00:03:31] (step=0004057) Train Loss mse: 0.0000, Train Loss ce: 0.3793, Train Steps/Sec: 0.67,
-[2026-01-29 23:35:30] (step=0002716) Train Loss mse: 0.0000, Train Loss ce: 0.3701, Train Steps/Sec: 0.85,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step4000
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-ce_avg: 2.886735677719116, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step4500
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-ce_avg: 3.8481361865997314, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step5000
-Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+[2026-01-30 00:03:32] (step=0004058) Train Loss mse: 0.0000, Train Loss ce: 0.3543, Train Steps/Sec: 0.90,
+[2026-01-30 00:03:34] (step=0004059) Train Loss mse: 0.0000, Train Loss ce: 0.3605, Train Steps/Sec: 0.56,
+[2026-01-30 00:03:35] (step=0004060) Train Loss mse: 0.0000, Train Loss ce: 0.3574, Train Steps/Sec: 0.85,
+[2026-01-30 00:03:36] (step=0004061) Train Loss mse: 0.0000, Train Loss ce: 0.3738, Train Steps/Sec: 0.74,
+[2026-01-30 00:03:38] (step=0004062) Train Loss mse: 0.0000, Train Loss ce: 0.3642, Train Steps/Sec: 0.68,
+[2026-01-30 00:03:39] (step=0004063) Train Loss mse: 0.0000, Train Loss ce: 0.3514, Train Steps/Sec: 0.85,
+[2026-01-30 00:03:40] (step=0004064) Train Loss mse: 0.0000, Train Loss ce: 0.3550, Train Steps/Sec: 0.90,
+[2026-01-30 00:03:41] (step=0004065) Train Loss mse: 0.0000, Train Loss ce: 0.3646, Train Steps/Sec: 0.90,
+[2026-01-30 00:03:43] (step=0004066) Train Loss mse: 0.0000, Train Loss ce: 0.3483, Train Steps/Sec: 0.72,
+[2026-01-30 00:03:44] (step=0004067) Train Loss mse: 0.0000, Train Loss ce: 0.3649, Train Steps/Sec: 0.67,
+[2026-01-30 00:03:45] (step=0004068) Train Loss mse: 0.0000, Train Loss ce: 0.3743, Train Steps/Sec: 0.69,
+[2026-01-30 00:03:47] (step=0004069) Train Loss mse: 0.0000, Train Loss ce: 0.3579, Train Steps/Sec: 0.90,
+[2026-01-30 00:03:48] (step=0004070) Train Loss mse: 0.0000, Train Loss ce: 0.3509, Train Steps/Sec: 0.75,
+[2026-01-30 00:03:49] (step=0004071) Train Loss mse: 0.0000, Train Loss ce: 0.3552, Train Steps/Sec: 0.69,
+[2026-01-30 00:03:50] (step=0004072) Train Loss mse: 0.0000, Train Loss ce: 0.3556, Train Steps/Sec: 0.89,
+[2026-01-30 00:03:52] (step=0004073) Train Loss mse: 0.0000, Train Loss ce: 0.3654, Train Steps/Sec: 0.90,
+[2026-01-30 00:03:53] (step=0004074) Train Loss mse: 0.0000, Train Loss ce: 0.3773, Train Steps/Sec: 0.71,
+[2026-01-30 00:03:54] (step=0004075) Train Loss mse: 0.0000, Train Loss ce: 0.3589, Train Steps/Sec: 0.69,
+[2026-01-30 00:03:56] (step=0004076) Train Loss mse: 0.0000, Train Loss ce: 0.3530, Train Steps/Sec: 0.72,
+[2026-01-30 00:03:57] (step=0004077) Train Loss mse: 0.0000, Train Loss ce: 0.3523, Train Steps/Sec: 0.90,
+[2026-01-30 00:03:58] (step=0004078) Train Loss mse: 0.0000, Train Loss ce: 0.3796, Train Steps/Sec: 0.72,
+[2026-01-30 00:03:59] (step=0004079) Train Loss mse: 0.0000, Train Loss ce: 0.3589, Train Steps/Sec: 0.90,
+[2026-01-30 00:04:01] (step=0004080) Train Loss mse: 0.0000, Train Loss ce: 0.3491, Train Steps/Sec: 0.69,
+[2026-01-30 00:04:02] (step=0004081) Train Loss mse: 0.0000, Train Loss ce: 0.3521, Train Steps/Sec: 0.84,
+[2026-01-30 00:04:04] (step=0004082) Train Loss mse: 0.0000, Train Loss ce: 0.3533, Train Steps/Sec: 0.68,
+[2026-01-30 00:04:05] (step=0004083) Train Loss mse: 0.0000, Train Loss ce: 0.3622, Train Steps/Sec: 0.71,
+[2026-01-30 00:04:06] (step=0004084) Train Loss mse: 0.0000, Train Loss ce: 0.3492, Train Steps/Sec: 0.91,
+[2026-01-30 00:04:07] (step=0004085) Train Loss mse: 0.0000, Train Loss ce: 0.3543, Train Steps/Sec: 0.90,
+[2026-01-30 00:04:09] (step=0004086) Train Loss mse: 0.0000, Train Loss ce: 0.3772, Train Steps/Sec: 0.72,
+[2026-01-30 00:04:10] (step=0004087) Train Loss mse: 0.0000, Train Loss ce: 0.3568, Train Steps/Sec: 0.88,
+[2026-01-30 00:04:11] (step=0004088) Train Loss mse: 0.0000, Train Loss ce: 0.3771, Train Steps/Sec: 0.70,
+[2026-01-30 00:04:12] (step=0004089) Train Loss mse: 0.0000, Train Loss ce: 0.3694, Train Steps/Sec: 0.89,
+[2026-01-30 00:04:14] (step=0004090) Train Loss mse: 0.0000, Train Loss ce: 0.3473, Train Steps/Sec: 0.69,
+[2026-01-30 00:04:15] (step=0004091) Train Loss mse: 0.0000, Train Loss ce: 0.3516, Train Steps/Sec: 0.71,
+[2026-01-30 00:04:16] (step=0004092) Train Loss mse: 0.0000, Train Loss ce: 0.3711, Train Steps/Sec: 0.90,
+[2026-01-30 00:04:17] (step=0004093) Train Loss mse: 0.0000, Train Loss ce: 0.3651, Train Steps/Sec: 0.90,
+[2026-01-30 00:04:19] (step=0004094) Train Loss mse: 0.0000, Train Loss ce: 0.3730, Train Steps/Sec: 0.75,
+[2026-01-30 00:04:20] (step=0004095) Train Loss mse: 0.0000, Train Loss ce: 0.3569, Train Steps/Sec: 0.86,
+[2026-01-30 00:04:21] (step=0004096) Train Loss mse: 0.0000, Train Loss ce: 0.3653, Train Steps/Sec: 0.66,
+[2026-01-30 00:04:23] (step=0004097) Train Loss mse: 0.0000, Train Loss ce: 0.3717, Train Steps/Sec: 0.89,
+[2026-01-30 00:04:24] (step=0004098) Train Loss mse: 0.0000, Train Loss ce: 0.3593, Train Steps/Sec: 0.57,
+[2026-01-30 00:04:25] (step=0004099) Train Loss mse: 0.0000, Train Loss ce: 0.3588, Train Steps/Sec: 0.85,
+[2026-01-30 00:04:27] (step=0004100) Train Loss mse: 0.0000, Train Loss ce: 0.3626, Train Steps/Sec: 0.91,
+[2026-01-30 00:04:28] (step=0004101) Train Loss mse: 0.0000, Train Loss ce: 0.3545, Train Steps/Sec: 0.85,
+[2026-01-30 00:04:29] (step=0004102) Train Loss mse: 0.0000, Train Loss ce: 0.3589, Train Steps/Sec: 0.76,
+[2026-01-30 00:04:30] (step=0004103) Train Loss mse: 0.0000, Train Loss ce: 0.3637, Train Steps/Sec: 0.90,
+[2026-01-30 00:04:32] (step=0004104) Train Loss mse: 0.0000, Train Loss ce: 0.3636, Train Steps/Sec: 0.70,
+[2026-01-30 00:04:33] (step=0004105) Train Loss mse: 0.0000, Train Loss ce: 0.3388, Train Steps/Sec: 0.71,
+[2026-01-30 00:04:34] (step=0004106) Train Loss mse: 0.0000, Train Loss ce: 0.3552, Train Steps/Sec: 0.72,
+[2026-01-30 00:04:36] (step=0004107) Train Loss mse: 0.0000, Train Loss ce: 0.3586, Train Steps/Sec: 0.88,
+[2026-01-30 00:04:37] (step=0004108) Train Loss mse: 0.0000, Train Loss ce: 0.3677, Train Steps/Sec: 0.86,
+[2026-01-30 00:04:38] (step=0004109) Train Loss mse: 0.0000, Train Loss ce: 0.3573, Train Steps/Sec: 0.76,
+[2026-01-30 00:04:39] (step=0004110) Train Loss mse: 0.0000, Train Loss ce: 0.3628, Train Steps/Sec: 0.90,
+[2026-01-30 00:04:40] (step=0004111) Train Loss mse: 0.0000, Train Loss ce: 0.3559, Train Steps/Sec: 0.89,
+[2026-01-30 00:04:42] (step=0004112) Train Loss mse: 0.0000, Train Loss ce: 0.3957, Train Steps/Sec: 0.67,
+[2026-01-30 00:04:43] (step=0004113) Train Loss mse: 0.0000, Train Loss ce: 0.3690, Train Steps/Sec: 0.69,
+[2026-01-30 00:04:44] (step=0004114) Train Loss mse: 0.0000, Train Loss ce: 0.3488, Train Steps/Sec: 0.84,
+[2026-01-30 00:04:46] (step=0004115) Train Loss mse: 0.0000, Train Loss ce: 0.3580, Train Steps/Sec: 0.90,
+[2026-01-30 00:04:47] (step=0004116) Train Loss mse: 0.0000, Train Loss ce: 0.3638, Train Steps/Sec: 0.84,
+[2026-01-30 00:04:48] (step=0004117) Train Loss mse: 0.0000, Train Loss ce: 0.3688, Train Steps/Sec: 0.76,
+[2026-01-30 00:04:49] (step=0004118) Train Loss mse: 0.0000, Train Loss ce: 0.3791, Train Steps/Sec: 0.84,
+[2026-01-30 00:04:51] (step=0004119) Train Loss mse: 0.0000, Train Loss ce: 0.3575, Train Steps/Sec: 0.69,
+[2026-01-30 00:04:52] (step=0004120) Train Loss mse: 0.0000, Train Loss ce: 0.3538, Train Steps/Sec: 0.71,
+[2026-01-30 00:04:53] (step=0004121) Train Loss mse: 0.0000, Train Loss ce: 0.3558, Train Steps/Sec: 0.84,
+[2026-01-30 00:04:55] (step=0004122) Train Loss mse: 0.0000, Train Loss ce: 0.3825, Train Steps/Sec: 0.69,
+[2026-01-30 00:04:56] (step=0004123) Train Loss mse: 0.0000, Train Loss ce: 0.3577, Train Steps/Sec: 0.90,
+[2026-01-30 00:04:57] (step=0004124) Train Loss mse: 0.0000, Train Loss ce: 0.3586, Train Steps/Sec: 0.86,
+[2026-01-30 00:04:58] (step=0004125) Train Loss mse: 0.0000, Train Loss ce: 0.3512, Train Steps/Sec: 0.84,
+[2026-01-30 00:04:59] (step=0004126) Train Loss mse: 0.0000, Train Loss ce: 0.3589, Train Steps/Sec: 0.91,
+[2026-01-30 00:05:01] (step=0004127) Train Loss mse: 0.0000, Train Loss ce: 0.3581, Train Steps/Sec: 0.68,
+[2026-01-30 00:05:02] (step=0004128) Train Loss mse: 0.0000, Train Loss ce: 0.3660, Train Steps/Sec: 0.91,
+[2026-01-30 00:05:03] (step=0004129) Train Loss mse: 0.0000, Train Loss ce: 0.3729, Train Steps/Sec: 0.67,
+[2026-01-30 00:05:05] (step=0004130) Train Loss mse: 0.0000, Train Loss ce: 0.3638, Train Steps/Sec: 0.88,
+[2026-01-30 00:05:06] (step=0004131) Train Loss mse: 0.0000, Train Loss ce: 0.3535, Train Steps/Sec: 0.84,
+[2026-01-30 00:05:07] (step=0004132) Train Loss mse: 0.0000, Train Loss ce: 0.3656, Train Steps/Sec: 0.71,
+[2026-01-30 00:05:08] (step=0004133) Train Loss mse: 0.0000, Train Loss ce: 0.3856, Train Steps/Sec: 0.91,
+[2026-01-30 00:05:10] (step=0004134) Train Loss mse: 0.0000, Train Loss ce: 0.3409, Train Steps/Sec: 0.70,
+[2026-01-30 00:05:11] (step=0004135) Train Loss mse: 0.0000, Train Loss ce: 0.3629, Train Steps/Sec: 0.84,
+[2026-01-30 00:05:12] (step=0004136) Train Loss mse: 0.0000, Train Loss ce: 0.3530, Train Steps/Sec: 0.76,
+[2026-01-30 00:05:13] (step=0004137) Train Loss mse: 0.0000, Train Loss ce: 0.3742, Train Steps/Sec: 0.90,
+[2026-01-30 00:05:14] (step=0004138) Train Loss mse: 0.0000, Train Loss ce: 0.3588, Train Steps/Sec: 0.90,
+[2026-01-30 00:05:16] (step=0004139) Train Loss mse: 0.0000, Train Loss ce: 0.3755, Train Steps/Sec: 0.71,
+[2026-01-30 00:05:17] (step=0004140) Train Loss mse: 0.0000, Train Loss ce: 0.3742, Train Steps/Sec: 0.86,
+[2026-01-30 00:05:18] (step=0004141) Train Loss mse: 0.0000, Train Loss ce: 0.3802, Train Steps/Sec: 0.70,
+[2026-01-30 00:05:20] (step=0004142) Train Loss mse: 0.0000, Train Loss ce: 0.3632, Train Steps/Sec: 0.84,
+[2026-01-30 00:05:21] (step=0004143) Train Loss mse: 0.0000, Train Loss ce: 0.3615, Train Steps/Sec: 0.70,
+[2026-01-30 00:05:22] (step=0004144) Train Loss mse: 0.0000, Train Loss ce: 0.3584, Train Steps/Sec: 0.87,
+[2026-01-30 00:05:23] (step=0004145) Train Loss mse: 0.0000, Train Loss ce: 0.3450, Train Steps/Sec: 0.90,
+[2026-01-30 00:05:25] (step=0004146) Train Loss mse: 0.0000, Train Loss ce: 0.3440, Train Steps/Sec: 0.71,
+[2026-01-30 00:05:26] (step=0004147) Train Loss mse: 0.0000, Train Loss ce: 0.3615, Train Steps/Sec: 0.83,
+[2026-01-30 00:05:27] (step=0004148) Train Loss mse: 0.0000, Train Loss ce: 0.3581, Train Steps/Sec: 0.71,
+[2026-01-30 00:05:29] (step=0004149) Train Loss mse: 0.0000, Train Loss ce: 0.3668, Train Steps/Sec: 0.84,
+[2026-01-30 00:05:30] (step=0004150) Train Loss mse: 0.0000, Train Loss ce: 0.3482, Train Steps/Sec: 0.69,
+[2026-01-30 00:05:31] (step=0004151) Train Loss mse: 0.0000, Train Loss ce: 0.3644, Train Steps/Sec: 0.86,
+[2026-01-30 00:05:32] (step=0004152) Train Loss mse: 0.0000, Train Loss ce: 0.3465, Train Steps/Sec: 0.83,
+[2026-01-30 00:05:34] (step=0004153) Train Loss mse: 0.0000, Train Loss ce: 0.3505, Train Steps/Sec: 0.70,
+[2026-01-30 00:05:35] (step=0004154) Train Loss mse: 0.0000, Train Loss ce: 0.3737, Train Steps/Sec: 0.83,
+[2026-01-30 00:05:36] (step=0004155) Train Loss mse: 0.0000, Train Loss ce: 0.3745, Train Steps/Sec: 0.73,
+[2026-01-30 00:05:37] (step=0004156) Train Loss mse: 0.0000, Train Loss ce: 0.3861, Train Steps/Sec: 0.90,
 [2026-01-30 00:05:39] (step=0004157) Train Loss mse: 0.0000, Train Loss ce: 0.3557, Train Steps/Sec: 0.71,
 [2026-01-30 00:05:40] (step=0004158) Train Loss mse: 0.0000, Train Loss ce: 0.3751, Train Steps/Sec: 0.68,
 [2026-01-30 00:05:41] (step=0004159) Train Loss mse: 0.0000, Train Loss ce: 0.3665, Train Steps/Sec: 0.91,
@@ -4874,6 +4966,13 @@ Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference
 [2026-01-30 00:17:44] (step=0004729) Train Loss mse: 0.0000, Train Loss ce: 0.3545, Train Steps/Sec: 0.71,
 [2026-01-30 00:17:45] (step=0004730) Train Loss mse: 0.0000, Train Loss ce: 0.3697, Train Steps/Sec: 0.83,
 [2026-01-30 00:17:47] (step=0004731) Train Loss mse: 0.0000, Train Loss ce: 0.3464, Train Steps/Sec: 0.58,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins_step5000
+Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference_dot_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_no_mse_evalonce'}]
+ce_avg: 6.3412766456604, mse_avg: 0.0
 [2026-01-30 00:17:48] (step=0004732) Train Loss mse: 0.0000, Train Loss ce: 0.3669, Train Steps/Sec: 0.84,
 [2026-01-30 00:17:49] (step=0004733) Train Loss mse: 0.0000, Train Loss ce: 0.3566, Train Steps/Sec: 0.89,
 [2026-01-30 00:17:50] (step=0004734) Train Loss mse: 0.0000, Train Loss ce: 0.3595, Train Steps/Sec: 0.84,
@@ -5146,5 +5245,4 @@ Preparing Dataset vlm_gym_reference_dot_celoss_no_mse_evalonce/vlm_gym_reference
 [2026-01-30 00:23:24] Saving checkpoint to /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_no_mse_ins/0005000.
 /opt/conda/lib/python3.11/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:690: FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP. API doc: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict.get_state_dict .Tutorial: https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html .
   warnings.warn(
-[2026-01-30 00:25:59] Done!
-[2026-01-30 00:05:39] (step=0004157) Train Loss mse: 0.0000, Train Loss ce: 0.3557, Train Steps/Sec: 0.71,
+[2026-01-30 00:25:59] Done!