---
license: apache-2.0
authors:
- Corneille Marechal
task_categories:
- robotics
tags:
- tsfile
- timeseries
- tabular
- robotics
- lerobot
- so101
- manipulation
- tea-making
modality:
- timeseries
- tabular
pretty_name: Cornito SO101 Tea2 TsFile
configs:
- config_name: default
  data_files:
  - split: train
    path: data/cornito_so101_tea2.tsfile
size_categories:
- 10K<n<100K
---

# Cornito SO101 Tea2 TsFile

This dataset is an Apache TsFile conversion of
[`Cornito/so101_tea2`](https://huggingface.co/datasets/Cornito/so101_tea2), a LeRobot v2.1 SO101 robot-manipulation dataset
containing 101 demonstrations for the task **Make tea**.

The converted repository contains numeric robot states, actions, frame timing,
and episode/task tags. The two camera streams remain in the original Hugging
Face dataset and are linked below.

## Source Dataset and Provenance

- Original dataset: [`Cornito/so101_tea2`](https://huggingface.co/datasets/Cornito/so101_tea2)
- Pinned source revision: [`3da2ed9f4d512bf7b5e06821a7a9c70d9c416c99`](https://huggingface.co/datasets/Cornito/so101_tea2/tree/3da2ed9f4d512bf7b5e06821a7a9c70d9c416c99)
- Original repository creator and uploader: [Corneille Marechal (`Cornito`)](https://huggingface.co/Cornito)
- License: Apache-2.0
- Robot type: `so101`
- LeRobot codebase version: `v2.1`
- Task: `Make tea` (`task_index = 0`)
- Split: `train`
- Sampling rate: 30 fps
- Scale: 101 episodes, 78,833 frame rows, 1 task, 101 source Parquet files, 202 source videos
- Source frame layout: `data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet`
- Source video layout: `videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4`

The source dataset card provides no paper or completed citation.

## Converted Files

- TsFile: `data/cornito_so101_tea2.tsfile`
- Table: `cornito_so101_tea2`
- Rows: 78,833
- Episodes/devices: 101
- TsFile size: 0.86 MiB
- Numeric source Parquet size: 3.31 MiB
- TsFile/Parquet size ratio: 0.2615
- Time precision: milliseconds
- Metadata: `meta/` is mirrored from the source, with `meta/info.json` rewritten for the TsFile artifact.

## TsFile Schema

`Time` is an INT64 millisecond timestamp computed as
`round(timestamp * 1000)` and restarts at zero for each episode.

TAG columns, stored through the TsFile table-model device/tag mechanism:

- `episode_index`
- `task_index`

FIELD columns:

- `frame_index`
- `sample_index`
- `action_0`
- `action_1`
- `action_2`
- `action_3`
- `action_4`
- `action_5`
- `observation_state_0`
- `observation_state_1`
- `observation_state_2`
- `observation_state_3`
- `observation_state_4`
- `observation_state_5`

Flattened vector groups:

- `action` -> `action_0` ... `action_5` (6 FLOAT fields)
- `observation.state` -> `observation_state_0` ... `observation_state_5` (6 FLOAT fields)

## Conversion Notes

- The train split is merged into one table-model TsFile. Filter by
  `episode_index` and `task_index` to select an episode or task.
- Storage profile: Time uses `TS_2DIFF + LZ4`; FLOAT/DOUBLE use
  `GORILLA + ZSTD`; INT32/INT64 use `TS_2DIFF + ZSTD`; BOOLEAN, if present,
  uses `RLE + LZ4`. Physical codecs were checked from the generated TsFile.
- `episode_index` and `task_index` are TsFile TAG/device columns.
- `action[6]` and `observation.state[6]` are flattened to scalar FLOAT fields;
  the full source prefix is retained and `.` is replaced with `_`.
- The source `timestamp` column is dropped after `Time` synthesis because it is
  redundant with `Time / 1000` seconds.
- The source `index` column is retained as `sample_index`; `frame_index` is retained unchanged.
- No rows or numeric trajectory dimensions are dropped.
- Source video features are intentionally omitted from the TsFile because they are external MP4 assets.

## Videos

Videos are not duplicated in this converted repository. The pinned source
contains two frame-aligned camera streams:

- [`observation.images.phone`](https://huggingface.co/datasets/Cornito/so101_tea2/tree/3da2ed9f4d512bf7b5e06821a7a9c70d9c416c99/videos/chunk-000/observation.images.phone) - 101 per-episode MP4 files, 1,623,417,441 bytes
- [`observation.images.wrist`](https://huggingface.co/datasets/Cornito/so101_tea2/tree/3da2ed9f4d512bf7b5e06821a7a9c70d9c416c99/videos/chunk-000/observation.images.wrist) - 101 per-episode MP4 files, 616,287,833 bytes

Together the 202 MP4 files occupy 2,239,705,274 bytes. Numeric
rows remain aligned with the original videos through `episode_index`,
`frame_index`, and the source per-episode metadata.

## Validation

The generated TsFile passed row-count equality, schema/TAG checks, per-device
Time monotonicity, Java full-table readback, physical codec inspection, and file-size
comparison against the 101 source Parquet shards. Local JSON and Markdown
validation reports are retained with the conversion workspace and are not part
of the eventual Hugging Face upload set.

## Minimal Read Example

```python
from tsfile import TsFileReader

reader = TsFileReader("data/cornito_so101_tea2.tsfile")
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "action_0",
    "observation_state_0",
]

with reader.query_table("cornito_so101_tea2", columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())
reader.close()
```

## Citation

The source dataset card provides no paper or completed BibTeX citation. Cite
the original Hugging Face dataset and Corneille Marechal (`Cornito`) when using
this converted artifact.