# Artifact sources

The residual adapter is a BAS-VLA project training artifact from the OpenVLA / LIBERO-10 offline action-calibration setting. Project contributions use the MIT License in `LICENSE`.

The 400 action triplets, original agent-view observations, sampled simulator states, demonstration initial states, and two BDDL scenes are derived from [LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) and its [demonstration dataset](https://huggingface.co/datasets/yifengzhu-hf/LIBERO-datasets). The exact source demonstration filename, demo ID and frame index are retained in each record. LIBERO's copyright and MIT notice are preserved in `licenses/LIBERO-MIT.txt`.

The base carrier is [openvla/openvla-7b](https://huggingface.co/openvla/openvla-7b). The bundled normalization statistics come from [moojink/openvla-7b-oft-finetuned-libero-10](https://huggingface.co/moojink/openvla-7b-oft-finetuned-libero-10); their source is recorded in `normalization/source.json`, with the OpenVLA-OFT MIT notice in `licenses/OpenVLA-OFT-MIT.txt`.

The source implementation tokenizes instructions with lowercase `[a-z0-9]+` and uses binary presence features. The released weight file preserves every original tensor and is loadable by the public `ResidualAdapter` class. `example.py` applies it to one released matched record on CPU.