Upload checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins

checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/0005000/ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:504ca814489e807517cb32fa3c8b8022540650eb89f469efddd34f7029f09ed0
+size 58429204680

checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/wandb/offline-run-20260129_223432-checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins-run0/files/config.yaml

@@ -0,0 +1,185 @@
+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: 1769726072
+    t:
+      1:
+      - 1
+      - 5
+      - 11
+      - 41
+      - 49
+      - 53
+      - 71
+      - 105
+      2:
+      - 1
+      - 5
+      - 11
+      - 41
+      - 49
+      - 53
+      - 71
+      - 105
+      3:
+      - 4
+      - 13
+      - 14
+      - 37
+      - 42
+      4: 3.11.10
+      5: 0.23.1
+      6: 4.49.0
+      13: linux-x86_64
+    e:
+      6tk8s0gf5rmih37ltdvrvulzn60qwyq6:
+        os: Linux-6.6.93+-x86_64-with-glibc2.35
+        python: CPython 3.11.10
+        started_at: '2026-01-29T22:34:32.380519Z'
+        args:
+        - --dataset_config_file
+        - ./data/configs/vlm_gym_reference_dot_train_celoss.yaml
+        - --eval_dataset_config_file
+        - ./data/configs/vlm_gym_reference_dot_eval_celoss.yaml
+        - --viz_dataset_config_file
+        - ./data/configs/vlm_gym_reference_dot_eval_celoss.yaml
+        - --inference_hash_file
+        - /home/clouduser/Code/Github/launch_new/hashes_test_set_v10.json
+        - --task_name
+        - reference_dot_v5
+        - --instructions_dir
+        - ./data/instructions
+        - --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/
+        - --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'
+        - --save_every
+        - '2500'
+        - --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_ins
+        - --wandb_project
+        - bagel
+        - --wandb_name
+        - checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins
+        - --wandb_dir
+        - /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_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_ins
+        host: junyizhang-launch-new-226786192-1-0
+        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: '164464467968'
+        memory:
+          total: '1437332606976'
+        gpu_nvidia:
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-718e070b-29b1-a9c6-6ba9-0b308e8e14b6
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-860c1016-b60a-e80f-ec59-7d7a02edcb79
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-2d63d08c-e8a0-154b-d0ef-8fda2ae1bf3f
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-6da90038-6dcb-16f8-9d1a-7512461977f7
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-1978a88e-d773-51d3-9c6e-16af2fa02d54
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-dc134b53-25d6-69ec-c7d8-a6539915eed3
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-32301ac6-38ad-f648-7d53-f81113c91cd0
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-d9e60c8f-df0c-485f-3d26-a38485ea847a
+        cuda_version: '12.2'
+        writer_id: 6tk8s0gf5rmih37ltdvrvulzn60qwyq6

checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/wandb/offline-run-20260130_175019-checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins-run0/files/config.yaml

@@ -0,0 +1,457 @@
+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: 1769795419
+    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:
+      oxr547070rpilwmqbpq2pd9ofzvd2cmh:
+        os: Linux-6.6.93+-x86_64-with-glibc2.35
+        python: CPython 3.11.10
+        started_at: '2026-01-30T17:50:19.134291Z'
+        args:
+        - --dataset_config_file
+        - ./data/configs/vlm_gym_reference_dot_train_celoss.yaml
+        - --eval_dataset_config_file
+        - ./data/configs/vlm_gym_reference_dot_eval_celoss.yaml
+        - --viz_dataset_config_file
+        - ./data/configs/vlm_gym_reference_dot_eval_celoss.yaml
+        - --inference_hash_file
+        - /home/clouduser/Code/Github/launch_new/hashes_test_set_v10.json
+        - --task_name
+        - reference_dot_v5
+        - --instructions_dir
+        - ./data/instructions
+        - --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/
+        - --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
+        - '40000'
+        - --max_num_tokens
+        - '40000'
+        - --max_num_tokens_per_sample
+        - '40000'
+        - --visual_und
+        - 'True'
+        - --save_every
+        - '2500'
+        - --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_ins
+        - --wandb_project
+        - bagel
+        - --wandb_name
+        - checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins
+        - --wandb_dir
+        - /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_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: 20694bc0d8bf3c48fe817e31af63a14d6e1e3619
+        root: /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins
+        host: junyizhang-launch-new-226991772-1-0
+        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: '164465041408'
+        memory:
+          total: '1437332606976'
+        gpu_nvidia:
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-cd28017c-3793-b451-76cc-3904ca15388e
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-e6224942-e179-ef85-abeb-afb4424bb934
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-21c6cca4-bc02-fbdc-e654-6039d84c8646
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-5f3fa3a7-ddb3-6751-9af7-4b4d38f7a411
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-4e7da01a-ba1b-e632-6334-7edc29e0de7f
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-2604cd40-2d05-63d5-0db4-27495b2a908c
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-8b72b62b-9a63-80eb-c8f8-bf16ad127f33
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-1ff104d8-f2ed-7a12-c3e9-3b5b61be1a4f
+        cuda_version: '12.2'
+        writer_id: oxr547070rpilwmqbpq2pd9ofzvd2cmh
+visual_gen:
+  desc: null
+  value: true
+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_ins
+wandb_project:
+  desc: null
+  value: bagel
+wandb_name:
+  desc: null
+  value: checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_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_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: 2500
+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: 40000
+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.yaml
+viz_dataset_config_file:
+  desc: null
+  value: ./data/configs/vlm_gym_reference_dot_eval_celoss.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.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: 40000
+max_num_tokens:
+  desc: null
+  value: 40000
+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_ins/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/wandb/offline-run-20260130_175019-checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins-run0/files/output.log

@@ -1153,27 +1153,6 @@ wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
 [2026-01-30 19:17:34] (step=0000956) Train Loss mse: 0.0260, Train Loss ce: 0.3673, Train Steps/Sec: 0.22,
 [2026-01-30 19:17:39] (step=0000957) Train Loss mse: 0.1040, Train Loss ce: 0.3833, Train Steps/Sec: 0.22,
 [2026-01-30 19:17:45] (step=0000958) Train Loss mse: 0.0684, Train Loss ce: 0.3811, Train Steps/Sec: 0.16,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step1000
-Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-ce_avg: 0.38083067536354065, mse_avg: 0.0056973714381456375
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step1500
-Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-ce_avg: 0.5134702324867249, mse_avg: 0.00566928181797266
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step2000
-Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-ce_avg: 0.6963174939155579, mse_avg: 0.005763474386185408
 [2026-01-30 19:17:50] (step=0000959) Train Loss mse: 0.0628, Train Loss ce: 0.3660, Train Steps/Sec: 0.20,
 [2026-01-30 19:17:55] (step=0000960) Train Loss mse: 0.0377, Train Loss ce: 0.3751, Train Steps/Sec: 0.19,
 [2026-01-30 19:18:00] (step=0000961) Train Loss mse: 0.0594, Train Loss ce: 0.3755, Train Steps/Sec: 0.20,
@@ -1272,6 +1251,27 @@ ce_avg: 0.6963174939155579, mse_avg: 0.005763474386185408
 [2026-01-30 19:26:15] (step=0001054) Train Loss mse: 0.0187, Train Loss ce: 0.3781, Train Steps/Sec: 0.22,
 [2026-01-30 19:26:19] (step=0001055) Train Loss mse: 0.0218, Train Loss ce: 0.3959, Train Steps/Sec: 0.21,
 [2026-01-30 19:26:24] (step=0001056) Train Loss mse: 0.0648, Train Loss ce: 0.3802, Train Steps/Sec: 0.23,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step1000
+Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+ce_avg: 0.38083067536354065, mse_avg: 0.0056973714381456375
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step1500
+Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+ce_avg: 0.5134702324867249, mse_avg: 0.00566928181797266
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step2000
+Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+ce_avg: 0.6963174939155579, mse_avg: 0.005763474386185408
 [2026-01-30 19:26:30] (step=0001057) Train Loss mse: 0.1090, Train Loss ce: 0.3662, Train Steps/Sec: 0.17,
 [2026-01-30 19:26:35] (step=0001058) Train Loss mse: 0.0627, Train Loss ce: 0.3800, Train Steps/Sec: 0.20,
 [2026-01-30 19:26:39] (step=0001059) Train Loss mse: 0.0413, Train Loss ce: 0.3784, Train Steps/Sec: 0.24,
@@ -2565,20 +2565,6 @@ ce_avg: 0.6963174939155579, mse_avg: 0.005763474386185408
 [2026-01-30 21:16:52] (step=0002347) Train Loss mse: 0.0233, Train Loss ce: 0.3611, Train Steps/Sec: 0.20,
 [2026-01-30 21:16:57] (step=0002348) Train Loss mse: 0.0407, Train Loss ce: 0.3794, Train Steps/Sec: 0.21,
 [2026-01-30 21:17:03] (step=0002349) Train Loss mse: 0.0832, Train Loss ce: 0.3697, Train Steps/Sec: 0.16,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step2500
-Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-ce_avg: 1.2680165767669678, mse_avg: 0.005764862522482872
-base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step3000
-Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
-ce_avg: 0.35950416326522827, mse_avg: 0.005150754004716873
 [2026-01-30 21:17:08] (step=0002350) Train Loss mse: 0.0151, Train Loss ce: 0.3606, Train Steps/Sec: 0.21,
 [2026-01-30 21:17:14] (step=0002351) Train Loss mse: 0.0258, Train Loss ce: 0.3678, Train Steps/Sec: 0.18,
 [2026-01-30 21:17:20] (step=0002352) Train Loss mse: 0.0565, Train Loss ce: 0.3763, Train Steps/Sec: 0.16,
@@ -2598,6 +2584,20 @@ ce_avg: 0.35950416326522827, mse_avg: 0.005150754004716873
 [2026-01-30 21:18:29] (step=0002366) Train Loss mse: 0.0676, Train Loss ce: 0.3779, Train Steps/Sec: 0.21,
 [2026-01-30 21:18:33] (step=0002367) Train Loss mse: 0.0689, Train Loss ce: 0.3695, Train Steps/Sec: 0.22,
 [2026-01-30 21:18:39] (step=0002368) Train Loss mse: 0.0200, Train Loss ce: 0.3593, Train Steps/Sec: 0.18,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step2500
+Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+ce_avg: 1.2680165767669678, mse_avg: 0.005764862522482872
+base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step3000
+Preparing Dataset vlm_gym_reference_dot_celoss_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_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
+ce_avg: 0.35950416326522827, mse_avg: 0.005150754004716873
 [2026-01-30 21:18:44] (step=0002369) Train Loss mse: 0.0564, Train Loss ce: 0.3732, Train Steps/Sec: 0.21,
 [2026-01-30 21:18:48] (step=0002370) Train Loss mse: 0.0158, Train Loss ce: 0.3686, Train Steps/Sec: 0.21,
 [2026-01-30 21:18:53] (step=0002371) Train Loss mse: 0.0460, Train Loss ce: 0.3620, Train Steps/Sec: 0.21,
@@ -3577,6 +3577,14 @@ ce_avg: 0.35950416326522827, mse_avg: 0.005150754004716873
 [2026-01-30 22:44:57] (step=0003342) Train Loss mse: 0.0204, Train Loss ce: 0.3625, Train Steps/Sec: 0.21,
 [2026-01-30 22:45:02] (step=0003343) Train Loss mse: 0.0397, Train Loss ce: 0.3534, Train Steps/Sec: 0.21,
 [2026-01-30 22:45:06] (step=0003344) Train Loss mse: 0.0457, Train Loss ce: 0.3492, Train Steps/Sec: 0.24,
+[2026-01-30 22:45:11] (step=0003345) Train Loss mse: 0.0685, Train Loss ce: 0.3638, Train Steps/Sec: 0.22,
+[2026-01-30 22:45:15] (step=0003346) Train Loss mse: 0.0287, Train Loss ce: 0.3681, Train Steps/Sec: 0.24,
+[2026-01-30 22:45:20] (step=0003347) Train Loss mse: 0.0241, Train Loss ce: 0.3600, Train Steps/Sec: 0.22,
+[2026-01-30 22:45:24] (step=0003348) Train Loss mse: 0.0670, Train Loss ce: 0.3696, Train Steps/Sec: 0.22,
+[2026-01-30 22:45:29] (step=0003349) Train Loss mse: 0.0325, Train Loss ce: 0.3693, Train Steps/Sec: 0.21,
+[2026-01-30 22:45:36] (step=0003350) Train Loss mse: 0.0331, Train Loss ce: 0.3672, Train Steps/Sec: 0.16,
+[2026-01-30 22:45:40] (step=0003351) Train Loss mse: 0.0047, Train Loss ce: 0.3599, Train Steps/Sec: 0.21,
+[2026-01-30 22:45:47] (step=0003352) Train Loss mse: 0.0047, Train Loss ce: 0.3617, Train Steps/Sec: 0.16,
 base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step3500
 Preparing Dataset vlm_gym_reference_dot_celoss_evalonce/vlm_gym_reference_dot_val
 [eval debug] first 3 batch fingerprints:
@@ -3598,14 +3606,6 @@ Preparing Dataset vlm_gym_reference_dot_celoss_evalonce/vlm_gym_reference_dot_va
   fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
   fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
 ce_avg: 0.8022381663322449, mse_avg: 0.006051613483577967
-[2026-01-30 22:45:11] (step=0003345) Train Loss mse: 0.0685, Train Loss ce: 0.3638, Train Steps/Sec: 0.22,
-[2026-01-30 22:45:15] (step=0003346) Train Loss mse: 0.0287, Train Loss ce: 0.3681, Train Steps/Sec: 0.24,
-[2026-01-30 22:45:20] (step=0003347) Train Loss mse: 0.0241, Train Loss ce: 0.3600, Train Steps/Sec: 0.22,
-[2026-01-30 22:45:24] (step=0003348) Train Loss mse: 0.0670, Train Loss ce: 0.3696, Train Steps/Sec: 0.22,
-[2026-01-30 22:45:29] (step=0003349) Train Loss mse: 0.0325, Train Loss ce: 0.3693, Train Steps/Sec: 0.21,
-[2026-01-30 22:45:36] (step=0003350) Train Loss mse: 0.0331, Train Loss ce: 0.3672, Train Steps/Sec: 0.16,
-[2026-01-30 22:45:40] (step=0003351) Train Loss mse: 0.0047, Train Loss ce: 0.3599, Train Steps/Sec: 0.21,
-[2026-01-30 22:45:47] (step=0003352) Train Loss mse: 0.0047, Train Loss ce: 0.3617, Train Steps/Sec: 0.16,
 [2026-01-30 22:45:52] (step=0003353) Train Loss mse: 0.0185, Train Loss ce: 0.3710, Train Steps/Sec: 0.19,
 [2026-01-30 22:45:57] (step=0003354) Train Loss mse: 0.0042, Train Loss ce: 0.3512, Train Steps/Sec: 0.21,
 [2026-01-30 22:46:02] (step=0003355) Train Loss mse: 0.0047, Train Loss ce: 0.3710, Train Steps/Sec: 0.18,
@@ -5079,11 +5079,6 @@ ce_avg: 0.8022381663322449, mse_avg: 0.006051613483577967
 [2026-01-31 00:51:59] (step=0004823) Train Loss mse: 0.0137, Train Loss ce: 0.3659, Train Steps/Sec: 0.21,
 [2026-01-31 00:52:04] (step=0004824) Train Loss mse: 0.0307, Train Loss ce: 0.3498, Train Steps/Sec: 0.20,
 [2026-01-31 00:52:09] (step=0004825) Train Loss mse: 0.0056, Train Loss ce: 0.3591, Train Steps/Sec: 0.22,
-[2026-01-31 00:52:13] (step=0004826) Train Loss mse: 0.0402, Train Loss ce: 0.3701, Train Steps/Sec: 0.23,
-[2026-01-31 00:52:18] (step=0004827) Train Loss mse: 0.0730, Train Loss ce: 0.3659, Train Steps/Sec: 0.20,
-[2026-01-31 00:52:22] (step=0004828) Train Loss mse: 0.0319, Train Loss ce: 0.3802, Train Steps/Sec: 0.22,
-[2026-01-31 00:52:28] (step=0004829) Train Loss mse: 0.0325, Train Loss ce: 0.3511, Train Steps/Sec: 0.18,
-[2026-01-31 00:52:34] (step=0004830) Train Loss mse: 0.0518, Train Loss ce: 0.3576, Train Steps/Sec: 0.18,
 base_dir is /dev/shm/models/checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_reference_dot_one_image_lr2e_5_ce_ins_step5000
 Preparing Dataset vlm_gym_reference_dot_celoss_evalonce/vlm_gym_reference_dot_val
 [eval debug] first 3 batch fingerprints:
@@ -5091,6 +5086,11 @@ Preparing Dataset vlm_gym_reference_dot_celoss_evalonce/vlm_gym_reference_dot_va
   fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
   fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_reference_dot_celoss_evalonce'}]
 ce_avg: 1.4548426866531372, mse_avg: 0.005192228127270937
+[2026-01-31 00:52:13] (step=0004826) Train Loss mse: 0.0402, Train Loss ce: 0.3701, Train Steps/Sec: 0.23,
+[2026-01-31 00:52:18] (step=0004827) Train Loss mse: 0.0730, Train Loss ce: 0.3659, Train Steps/Sec: 0.20,
+[2026-01-31 00:52:22] (step=0004828) Train Loss mse: 0.0319, Train Loss ce: 0.3802, Train Steps/Sec: 0.22,
+[2026-01-31 00:52:28] (step=0004829) Train Loss mse: 0.0325, Train Loss ce: 0.3511, Train Steps/Sec: 0.18,
+[2026-01-31 00:52:34] (step=0004830) Train Loss mse: 0.0518, Train Loss ce: 0.3576, Train Steps/Sec: 0.18,
 [2026-01-31 00:52:40] (step=0004831) Train Loss mse: 0.0038, Train Loss ce: 0.3513, Train Steps/Sec: 0.16,
 [2026-01-31 00:52:45] (step=0004832) Train Loss mse: 0.0511, Train Loss ce: 0.3626, Train Steps/Sec: 0.19,
 [2026-01-31 00:52:50] (step=0004833) Train Loss mse: 0.0457, Train Loss ce: 0.3571, Train Steps/Sec: 0.20,