config.json

{
  "dataset": {
    "pose_dir": "data/kinetics_processed",
    "desc_dir": "data/kinetics_full_output/descriptions",
    "video_dir": "data/kinetics-dataset/k700-2020",
    "val_split": 0.02,
    "max_samples_per_class": null,
    "num_workers": 8,
    "image_size": 224,
    "normalize_pose": true,
    "use_joint_angles": true,
    "sample_stride": 16,
    "include_temporal_context": false,
    "action_focus_prompt": false,
    "video_fps": 10.0,
    "augment_flip": true,
    "seed": 42
  },
  "model_config": {
    "qwen_model_name": "/root/.cache/huggingface/hub/models--Qwen--Qwen3-VL-4B-Instruct/snapshots/ebb281ec70b05090aa6165b016eac8ec08e71b17",
    "qwen_hidden_size": 2560,
    "use_intermediate_hidden": true,
    "hidden_layer_fraction": 0.5,
    "hidden_layer_index": 18,
    "use_early_exit": true,
    "use_deepstack_features": true,
    "use_flash_attention": true,
    "projection_dim": 1024,
    "action_dim": 44,
    "diffusion_hidden_dim": 1536,
    "num_diffusion_layers": 24,
    "num_diffusion_heads": 24,
    "num_future_tokens": 4,
    "action_horizon": 16,
    "num_frames": 4,
    "use_lora": true,
    "lora_rank": 128,
    "lora_alpha": 128,
    "lora_dropout": 0.05,
    "freeze_vision_encoder": true,
    "freeze_qwen_layers": 0,
    "use_thinking_mode": false,
    "diffusion_steps": 2,
    "init_from_current_pose": false
  },
  "learning_rate": 1e-05,
  "llm_learning_rate": 1e-06,
  "weight_decay": 0.01,
  "batch_size": 8,
  "num_epochs": 4,
  "unfreeze_pct": 0.5,
  "gradient_accumulation_steps": 16,
  "gradient_clip": 1.0,
  "use_amp": true,
  "gradient_checkpointing": false,
  "log_every_n_steps": 1,
  "save_every_n_steps": 750,
  "val_max_batches": 100,
  "perf_log_every": 100,
  "max_checkpoints": 5,
  "log_dir": "logs/kinetics_vla",
  "checkpoint_dir": "checkpoints/kinetics_vla"
}