# SO-101 ICL pairs (368)

## Processed robot data with corrected instructions

The [processed seed42 dataset](so101_hippovla_splits_seed42/README.md) contains 350 training episodes and 18 validation episodes in LeRobot v2.1 format. Its 22 English instructions specify cube colors, the left or right cup, and placement order. The original train/validation membership is preserved.

Use the files under [`so101_hippovla_splits_seed42/`](so101_hippovla_splits_seed42) for the updated task labels. See its README for download instructions and the rel16 action convention. The source data below retains the original two generic instructions.

## Original paired source data

Each line of `pairs.jsonl` is one successful teleop episode paired with one human video. The paths are relative to the
dataset root, and the layout is the same as `Zero-WAM/data/SO101/raw`:

| robot_index | instruction | robot episodes | human video |
|---|---|---|---|
| 1–240 | `Put the cube into the cup.` | 2026-09-22/23 teleop | #1–80 real (`data/<human episode>/external.mp4`), #81–240 MiniMax-H3 (`outputs/h3_batch200_20260924/.../review.mp4`) |
| 241–368 | `Put three cubes into the cup.` | 2026-09-24 teleop | all MiniMax-H3 (`outputs/h3_task2_batch128_20260928/.../review.mp4`) |

- The first 240 rows are `datasets/so101_icl/pairs.jsonl` unchanged, plus an `instruction` field.
- For rows 241–368, `task` is `<order of the three colours>_<cup>`. The rows come from
  `outputs/h3_task2_batch128_20260928/manifest.json`: 16 orders × 4 layouts × 2 samples. Each robot episode appears once.
- Nobody has reviewed the content of the 128 H3 videos (`human_review: null`). They were checked only for
  decodability and SHA256.
- The `instruction` comes from each episode's `frames.jsonl`. `Zero-WAM/script/prepare_so101.py` asserts a single
  language across all pairs, so it rejects the combined set as written.

Remote copy: `H100-2:/mnt/data/QingLan/datasets/so101_icl_368/`. The files it shares with the old upload are
hardlinks to `Zero-WAM/data/SO101/raw`.