Upload checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins

checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/wandb/offline-run-20260126_213949-checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins-run0/files/output.log

@@ -1,3 +1,196 @@
+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)
+    )
+    (time_embedder): FullyShardedDataParallel(
+      (_fsdp_wrapped_module): TimestepEmbedder(
+        (mlp): Sequential(
+          (0): Linear(in_features=256, out_features=3584, bias=True)
+          (1): SiLU()
+          (2): Linear(in_features=3584, out_features=3584, bias=True)
+        )
+      )
+    )
+    (vae2llm): Linear(in_features=64, out_features=3584, bias=True)
+    (llm2vae): Linear(in_features=3584, out_features=64, bias=True)
+    (latent_pos_embed): FullyShardedDataParallel(
+      (_fsdp_wrapped_module): PositionEmbedding()
+    )
+    (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
+time_embedder._fsdp_wrapped_module._flat_param True
+latent_pos_embed._fsdp_wrapped_module._flat_param False
+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_match_equation_sos_celoss/vlm_gym_match_equation_sos_train
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step0
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 1.8181179761886597, mse_avg: 0.6639004349708557
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step500
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.06269639730453491, mse_avg: 0.0149351442232728
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step1000
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.06731808930635452, mse_avg: 0.011200404725968838
 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/
@@ -1018,6 +1211,20 @@ wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
 [2026-01-26 22:12:48] (step=0001007) Train Loss mse: 0.0116, Train Loss ce: 0.0712, Train Steps/Sec: 0.58,
 [2026-01-26 22:12:50] (step=0001008) Train Loss mse: 0.0130, Train Loss ce: 0.0336, Train Steps/Sec: 0.68,
 [2026-01-26 22:12:51] (step=0001009) Train Loss mse: 0.0141, Train Loss ce: 0.0598, Train Steps/Sec: 0.68,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step1500
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.1969330757856369, mse_avg: 0.010982552543282509
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step2000
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.24203190207481384, mse_avg: 0.011826912872493267
 [2026-01-26 22:12:53] (step=0001010) Train Loss mse: 0.0162, Train Loss ce: 0.0454, Train Steps/Sec: 0.68,
 [2026-01-26 22:12:54] (step=0001011) Train Loss mse: 0.0109, Train Loss ce: 0.0345, Train Steps/Sec: 0.68,
 [2026-01-26 22:12:56] (step=0001012) Train Loss mse: 0.0095, Train Loss ce: 0.0584, Train Steps/Sec: 0.68,
@@ -1065,220 +1272,6 @@ wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
 [2026-01-26 22:14:01] (step=0001054) Train Loss mse: 0.0175, Train Loss ce: 0.0570, Train Steps/Sec: 0.68,
 [2026-01-26 22:14:03] (step=0001055) Train Loss mse: 0.0202, Train Loss ce: 0.0187, Train Steps/Sec: 0.68,
 [2026-01-26 22:14:04] (step=0001056) Train Loss mse: 0.0158, Train Loss ce: 0.0599, Train Steps/Sec: 0.68,
-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)
-    )
-    (time_embedder): FullyShardedDataParallel(
-      (_fsdp_wrapped_module): TimestepEmbedder(
-        (mlp): Sequential(
-          (0): Linear(in_features=256, out_features=3584, bias=True)
-          (1): SiLU()
-          (2): Linear(in_features=3584, out_features=3584, bias=True)
-        )
-      )
-    )
-    (vae2llm): Linear(in_features=64, out_features=3584, bias=True)
-    (llm2vae): Linear(in_features=3584, out_features=64, bias=True)
-    (latent_pos_embed): FullyShardedDataParallel(
-      (_fsdp_wrapped_module): PositionEmbedding()
-    )
-    (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
-time_embedder._fsdp_wrapped_module._flat_param True
-latent_pos_embed._fsdp_wrapped_module._flat_param False
-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_match_equation_sos_celoss/vlm_gym_match_equation_sos_train
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step0
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 1.8181179761886597, mse_avg: 0.6639004349708557
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step500
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.06269639730453491, mse_avg: 0.0149351442232728
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step1000
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.06731808930635452, mse_avg: 0.011200404725968838
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step1500
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.1969330757856369, mse_avg: 0.010982552543282509
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step2000
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.24203190207481384, mse_avg: 0.011826912872493267
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step2500
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.20800399780273438, mse_avg: 0.011797062121331692
 [2026-01-26 22:14:06] (step=0001057) Train Loss mse: 0.0142, Train Loss ce: 0.0901, Train Steps/Sec: 0.68,
 [2026-01-26 22:14:07] (step=0001058) Train Loss mse: 0.0084, Train Loss ce: 0.0419, Train Steps/Sec: 0.67,
 [2026-01-26 22:14:09] (step=0001059) Train Loss mse: 0.0127, Train Loss ce: 0.0692, Train Steps/Sec: 0.56,
@@ -2707,6 +2700,27 @@ ce_avg: 0.20800399780273438, mse_avg: 0.011797062121331692
 [2026-01-26 22:52:32] (step=0002482) Train Loss mse: 0.0074, Train Loss ce: 0.0610, Train Steps/Sec: 0.58,
 [2026-01-26 22:52:33] (step=0002483) Train Loss mse: 0.0053, Train Loss ce: 0.0485, Train Steps/Sec: 0.68,
 [2026-01-26 22:52:35] (step=0002484) Train Loss mse: 0.0062, Train Loss ce: 0.0460, Train Steps/Sec: 0.57,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step2500
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.20800399780273438, mse_avg: 0.011797062121331692
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step3000
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.04329414293169975, mse_avg: 0.006353132426738739
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step3500
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.04275057464838028, mse_avg: 0.0056571937166154385
 [2026-01-26 22:52:37] (step=0002485) Train Loss mse: 0.0051, Train Loss ce: 0.0446, Train Steps/Sec: 0.69,
 [2026-01-26 22:52:38] (step=0002486) Train Loss mse: 0.0086, Train Loss ce: 0.0438, Train Steps/Sec: 0.57,
 [2026-01-26 22:52:41] (step=0002487) Train Loss mse: 0.0067, Train Loss ce: 0.0641, Train Steps/Sec: 0.44,
@@ -2732,20 +2746,6 @@ ce_avg: 0.20800399780273438, mse_avg: 0.011797062121331692
 [2026-01-26 22:55:50] (step=0002504) Train Loss mse: 0.0089, Train Loss ce: 0.0733, Train Steps/Sec: 0.68,
 [2026-01-26 22:55:52] (step=0002505) Train Loss mse: 0.0061, Train Loss ce: 0.0662, Train Steps/Sec: 0.68,
 [2026-01-26 22:55:53] (step=0002506) Train Loss mse: 0.0068, Train Loss ce: 0.0412, Train Steps/Sec: 0.68,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step3000
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.04329414293169975, mse_avg: 0.006353132426738739
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step3500
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.04275057464838028, mse_avg: 0.0056571937166154385
 [2026-01-26 22:55:55] (step=0002507) Train Loss mse: 0.0085, Train Loss ce: 0.0475, Train Steps/Sec: 0.49,
 [2026-01-26 22:55:57] (step=0002508) Train Loss mse: 0.0056, Train Loss ce: 0.0384, Train Steps/Sec: 0.69,
 [2026-01-26 22:55:58] (step=0002509) Train Loss mse: 0.0106, Train Loss ce: 0.0332, Train Steps/Sec: 0.68,
@@ -3773,6 +3773,20 @@ ce_avg: 0.04275057464838028, mse_avg: 0.0056571937166154385
 [2026-01-26 23:23:15] (step=0003531) Train Loss mse: 0.0057, Train Loss ce: 0.0184, Train Steps/Sec: 0.68,
 [2026-01-26 23:23:16] (step=0003532) Train Loss mse: 0.0089, Train Loss ce: 0.0406, Train Steps/Sec: 0.68,
 [2026-01-26 23:23:18] (step=0003533) Train Loss mse: 0.0058, Train Loss ce: 0.0525, Train Steps/Sec: 0.55,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step4000
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.03896063566207886, mse_avg: 0.005920059513300657
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step4500
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.03593315929174423, mse_avg: 0.005641990341246128
 [2026-01-26 23:23:20] (step=0003534) Train Loss mse: 0.0057, Train Loss ce: 0.0243, Train Steps/Sec: 0.59,
 [2026-01-26 23:23:21] (step=0003535) Train Loss mse: 0.0072, Train Loss ce: 0.0283, Train Steps/Sec: 0.68,
 [2026-01-26 23:23:23] (step=0003536) Train Loss mse: 0.0100, Train Loss ce: 0.0241, Train Steps/Sec: 0.68,
@@ -3792,27 +3806,6 @@ ce_avg: 0.04275057464838028, mse_avg: 0.0056571937166154385
 [2026-01-26 23:23:44] (step=0003550) Train Loss mse: 0.0066, Train Loss ce: 0.0326, Train Steps/Sec: 0.68,
 [2026-01-26 23:23:46] (step=0003551) Train Loss mse: 0.0086, Train Loss ce: 0.0259, Train Steps/Sec: 0.48,
 [2026-01-26 23:23:48] (step=0003552) Train Loss mse: 0.0122, Train Loss ce: 0.0249, Train Steps/Sec: 0.59,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step4000
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.03896063566207886, mse_avg: 0.005920059513300657
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step4500
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.03593315929174423, mse_avg: 0.005641990341246128
-base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step5000
-Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
-ce_avg: 0.03604895621538162, mse_avg: 0.005358231253921986
 [2026-01-26 23:23:49] (step=0003553) Train Loss mse: 0.0082, Train Loss ce: 0.0280, Train Steps/Sec: 0.68,
 [2026-01-26 23:23:51] (step=0003554) Train Loss mse: 0.0045, Train Loss ce: 0.0642, Train Steps/Sec: 0.68,
 [2026-01-26 23:23:52] (step=0003555) Train Loss mse: 0.0075, Train Loss ce: 0.0468, Train Steps/Sec: 0.68,
@@ -5262,4 +5255,11 @@ ce_avg: 0.03604895621538162, mse_avg: 0.005358231253921986
 [2026-01-27 00:02:27] (step=0004999) Train Loss mse: 0.0058, Train Loss ce: 0.0432, Train Steps/Sec: 0.57,
 [2026-01-27 00:02:33] (step=0005000) Train Loss mse: 0.0040, Train Loss ce: 0.0410, Train Steps/Sec: 0.15,
 [2026-01-27 00:02:33] Saving checkpoint to /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/0005000.
-[2026-01-27 00:05:11] Done!
+[2026-01-27 00:05:11] Done!
+base_dir is /dev/shm/models/checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_match_equation_sos_one_image_lr2e_5_ce_ins_step5000
+Preparing Dataset vlm_gym_match_equation_sos_celoss_evalonce/vlm_gym_match_equation_sos_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_match_equation_sos_celoss_evalonce'}]
+ce_avg: 0.03604895621538162, mse_avg: 0.005358231253921986