Add files using upload-large-folder tool

checkpoints/steps_30000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6716ac2a700554e32edd37354a4fafdb0fc4fc0433e825900bd9bf8ed99d36a9
+size 9653648952

config.full.yaml

@@ -0,0 +1,64 @@
+run_id: stage1_full_4gpu
+run_root_dir: /workspace/tingting/3dvla-checkpoints
+seed: 42
+wandb_entity: tingtingdu06-uw-madison
+wandb_project: 3dvlm
+is_debug: false
+version_id: '0.21'
+framework:
+  name: QwenFast
+  qwenvl:
+    base_vlm: Qwen/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+  action_model:
+    action_model_type: FAST
+    action_dim: 7
+    future_action_window_size: 15
+    past_action_window_size: 0
+    action_horizon: 16
+  reduce_in_full_precision: true
+datasets:
+  vlm_data:
+    dataset_py: vlm_datasets_3d
+    dataset_use: materialized_mixture
+    annotation_path: /workspace/tingting/3dvla-data/sft/materialized/full_tokcal.jsonl
+    model_type: qwen3vl
+    model_max_length: 8192
+    max_pixels: 1254400
+    min_pixels: 784
+    video_min_pixels: 784
+    video_max_pixels: 401408
+    video_min_frames: 4
+    video_max_frames: 32
+    video_fps: 1
+    data_flatten: false
+    per_device_batch_size: 2
+    num_workers: 4
+trainer:
+  max_train_steps: 125000
+  num_warmup_steps: 2000
+  save_interval: 30000
+  eval_interval: 1000000000
+  logging_frequency: 10
+  learning_rate:
+    base: 1.0e-05
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  loss_scale:
+    vlm: 1.0
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 1
+  save_format: safetensors
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+config_yaml: /workspace/tingting/3dvla-stage0/train/stage1.yaml
+output_dir: /workspace/tingting/3dvla-checkpoints/stage1_full_4gpu

config.yaml

@@ -0,0 +1,43 @@
+datasets:
+  vlm_data:
+    dataset_py: vlm_datasets_3d
+    dataset_use: materialized_mixture
+    model_max_length: 8192
+    num_workers: 4
+    per_device_batch_size: 2
+framework:
+  action_model:
+    action_dim: 7
+    action_horizon: 16
+  name: QwenFast
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: Qwen/Qwen3-VL-4B-Instruct
+output_dir: /workspace/tingting/3dvla-checkpoints/stage1_full_4gpu
+run_id: stage1_full_4gpu
+run_root_dir: /workspace/tingting/3dvla-checkpoints
+seed: 42
+trainer:
+  eval_interval: 1000000000
+  freeze_modules: ''
+  gradient_clipping: 1.0
+  learning_rate:
+    base: 1.0e-05
+  logging_frequency: 10
+  loss_scale:
+    vlm: 1.0
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 125000
+  num_warmup_steps: 2000
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+  save_format: safetensors
+  save_interval: 30000
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: tingtingdu06-uw-madison
+wandb_project: 3dvlm

summary.jsonl

@@ -0,0 +1 @@
+{"steps": 30000}