# Real-world fine-tuning and inference code

The complete implementation is now included in [JAM-realrobot](https://github.com/lixuan27/JAM-realrobot/tree/f94dbc8913f5809746a98b42c1b4ce0e86b814dc/baselines/fastwam), on both `main` and `inference-gello-client`. Use the [current inference branch](https://github.com/lixuan27/JAM-realrobot/tree/inference-gello-client/baselines/fastwam) when incorporating later changes, or the pinned version above to keep an experiment fixed.

```bash
git clone --branch inference-gello-client https://github.com/lixuan27/JAM-realrobot.git
cd JAM-realrobot/baselines/fastwam
```

For an existing checkout on that branch, use `git pull --ff-only`. Record the Git revision and bundle revision together when deploying.

- [Complete fine-tuning workflow](https://github.com/lixuan27/JAM-realrobot/blob/f94dbc8913f5809746a98b42c1b4ce0e86b814dc/baselines/fastwam/FINETUNE.md): pinned raw demonstrations, conversion, task-local splits and target checks, normalization, native visual/text preparation, training, validation, full-state resume and export.
- [Inference setup](https://github.com/lixuan27/JAM-realrobot/blob/f94dbc8913f5809746a98b42c1b4ce0e86b814dc/baselines/fastwam/INFERENCE.md): environment, original base assets, published or newly exported task bundles, GPU serving and the robot-side client.
- [HTTP interface and client integration](https://github.com/lixuan27/JAM-realrobot/blob/f94dbc8913f5809746a98b42c1b4ce0e86b814dc/baselines/INFERENCE_INTERFACE.md): exact request fields, pose10 conventions, model-specific observation requirements and denormalized outputs.

The code copied into GitHub matches the published source identified by its `SOURCE_ORIGIN.json`. Checkpoints remain in this Hugging Face repository. Keep each adapter, normalization, frozen prompt and identity together; use the loader for the selected model. The same source remains available under this repository's `finetune/` and `inference/` directories.

Fast-WAM consumes the current image and measured state. LingBot-VA requires continuous real observed image history and its companion history client. Both return absolute flange pose10 targets; do not denormalize their HTTP output again. Follow the selected model's guide before adapting a custom client.