# eval2_90

SmolVLA policy for Eval 2.

Dataset:
- `robot-learning-group47/eval2_90_chengming_promptaug_v1`

Training setup:
- Base policy: `lerobot/smolvla_base`
- One real camera: `observation.images.front`
- Camera mapping: `observation.images.front` → `observation.images.camera1`
- `empty_cameras=2`
- Dynamic prompt sampling: random prompt sampled within the same layout-target bucket during training
- `n_obs_steps=1`
- `chunk_size=50`
- `n_action_steps=50`
- Batch size: 64
- Steps: 30,000
- Save frequency: 2,500
- `train_expert_only=true`
- `freeze_vision_encoder=true`
- `num_vlm_layers=16`
- Learning rate: `1e-4`
- Weight decay: `1e-10`

The repo root contains the latest checkpoint for direct loading.

All intermediate checkpoints are stored under:

`checkpoints/`