mcobzarenco/bijou-checkpoints

Bijou training checkpoints

Raw training checkpoints for Bijou, a vision-language-action model for SO-100/SO-101 arms: a frozen, truncated Gemma 4 E2B prefix encoder (layers 0–14; K/V of the global-attention layers exported as cross-attention streams) driving a flow-matching action expert that denoises 50-step action chunks. Code: mcobzarenco/flow-matching (bijou/).

These are not packaged inference artifacts — each <run>/<step>/ directory is a raw training checkpoint:

• expert.safetensors — action-expert weights (the only trained part)
• optimizer.pt — AdamW moments + LR-schedule state (--resume support)
• bijou_config.json — expert config, backbone id, per-dataset normalization stats table (+ count-weighted aggregate fallback), train args, step

Actions/states are MEAN_STD-normalized per dataset; inference must normalize with the deployment rig's stats (see the stats table in the config). Training data: the community_dataset_v1_v3 / community_dataset_v2_v3 LeRobot v3 conversions of the Hugging Face SO-100 community collections.

To load: build the model with bijou.loading.from_backbone using the expert config from bijou_config.json, then load expert.safetensors into model.expert (see bijou/train.py --init-from / --resume).