# eval3_sanitycheck

SmolVLA sanity-check overfit policy for Eval 3 coke-can placement.

- Base VLM: `HuggingFaceTB/SmolVLM2-500M-Video-Instruct`
- Policy type: `smolvla`
- Dataset: `robot-learning-group47/eval3_overfit20`
- Local dataset root: `/data/lerobot_datasets/robot-learning-group47/eval3_overfit20`
- Source episodes: `0,1,6,7,12,13,18,19,24,25,30,31,36,37,42,43,48,49,54,55`
- Steps: `3000`
- Batch size: `8`
- Image augmentation: disabled
- Trainable params: expert/state/action heads only, vision/VLM frozen
- Final training loss: approximately `0.182`

Important inference settings are stored in `config.json`:

- `chunk_size=50`
- `n_action_steps=50`
- `num_steps=10`
- `attention_mode=self_attn`
- `num_vlm_layers=8`
- `num_expert_layers=4`
- `resize_imgs_with_padding=[512,512]`