---
license: apache-2.0
language:
  - en
  - ja
tags:
  - robotics
  - so101
  - mimicanno
  - gemma-3
  - vlm-annotation
pretty_name: SO101 annotations (Gemma-4 26B / MimicAnno)
---

# SO101 annotations — Gemma-4 26B (QLoRA) MimicAnno run

20 episodes from the **SO101** dataset, automatically segmented and
phase-labeled by the [MimicAnno](https://github.com/Gayagaya/MimicAnno)
pipeline using the **26B Gemma-4 QLoRA adapter**
(<https://huggingface.co/Gayagaya/gem4_26B_adapter>).

## Contents (per episode dir)

| file              | description                                |
| ----------------- | ------------------------------------------ |
| `video.mp4`       | source episode video                       |
| `signals.json`    | per-frame hand/object signals (schema v3)  |
| `boundaries.json` | segment boundaries                         |
| `annotation.json` | VLM phase labels + tools                   |
| `tracks.json`     | SAM3 object tracks                         |
| `manifest.json`   | run metadata                               |
| `_masks/`         | SAM3 mask exports                          |

## How this was produced

Reproduce with the MimicAnno repo:

```bash
GPU=0 bash scripts/run_26B_gem4.sh so101
```

See the MimicAnno README §"26B QLoRA adapter" for full env setup
(`unsloth_env` conda env, adapter fetch via `hf download`, SAM3 +
HaMeR backends).