# SigLIP2 So400m temporal action-chunk robotics VLA

Self-contained policy produced by `train_siglip2_action_chunk_vla.py`.

- Frozen SigLIP2 So400m 384/14 visual encoder
- Official global pooler token + pooled 8x8 spatial patch tokens (no CLS)
- Frozen SigLIP2 text encoder and normalized 1152D instruction embeddings
- Lowercased SentencePiece instructions with a 64-token context
- Raw-RGB spatial color tokens
- Four-step visual/proprio/phase history
- Three-head, eight-step 7D action chunks
- Overlapping temporal ensemble and gripper hysteresis at inference
- Parameters: `1,138,041,878`
- Selected/refit epochs: `31`

Upload every file in this directory. The adapter loads `model.safetensors`
strictly and performs no model-hub or network access.

The SigLIP2 checkpoint is distributed under Apache License 2.0. See `LICENSE`
and `ATTRIBUTION.md`.