---
license: apache-2.0
authors:
  - Quinten Noens
task_categories:
  - robotics
tags:
  - tsfile
  - timeseries
  - tabular
  - robotics
  - lerobot
modality:
  - timeseries
  - tabular
pretty_name: Block2Target TsFile
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/qnoens_block2target.tsfile
size_categories:
  - 10K<n<100K
---

# Block2Target TsFile

Apache TsFile conversion of [qnoens/block2target](https://huggingface.co/datasets/qnoens/block2target), a LeRobot v2.1 robotics dataset.

## Source and provenance

- Original dataset: [qnoens/block2target](https://huggingface.co/datasets/qnoens/block2target)
- Inspected source revision: `70f0d984883d25eba7837c258c18d721f61315fb`
- Original author/uploader: [Quinten Noens (`qnoens`)](https://huggingface.co/qnoens)
- License: Apache-2.0
- Tasks: move a block to the blue rectangle, green circle, pink pentagon, or orange triangle
- Split: `train: 0:200`; sampling rate: 10 fps
- Coverage: 200 episodes, 23,071 frame rows, 4 tasks, 200 Parquet shards
- Source frame layout: `data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet`

The source card provides no paper, homepage, or completed BibTeX citation.

## Converted files

- TsFile: `data/qnoens_block2target.tsfile` (2,437,996 bytes)
- Table: `qnoens_block2target`; rows: 23,071; devices/episodes: 200
- Metadata: source `meta/` is mirrored, with `meta/info.json` updated for the TsFile artifact.

## Schema and conversion

`Time = round(timestamp * 1000)` as INT64 milliseconds, restarting at zero per episode. Source `timestamp` is dropped as redundant with `Time / 1000`. `frame_index` is retained and `index` is renamed to `sample_index`.

TAG/device columns: `episode_index`, `task_index`.

FIELD columns: `next_reward` FLOAT, `next_success` BOOLEAN, `seed`/`frame_index`/`sample_index` INT64, `gripper_state` FLOAT; plus flattened `state[7]`, `robot_pose[6]`, `joints[6]`, and `action[7]` scalar FLOAT fields. No rows or numeric vector dimensions are dropped.

Actual storage profile: Time `TS_2DIFF + LZ4`; FLOAT/DOUBLE `GORILLA + LZ4`; INT32/INT64 `TS_2DIFF + LZ4`; BOOLEAN `RLE + LZ4`; TAGs use TsFile table-model device/tag storage.

## Videos (not included)

The original repository contains 800 MP4 files (200 per stream; 367,041,730 bytes) under [videos/](https://huggingface.co/datasets/qnoens/block2target/tree/70f0d984883d25eba7837c258c18d721f61315fb/videos): `videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4`. Streams are `scene_image`, `scene_image_original`, `wrist_image`, and `wrist_image_original`. Videos are not copied or uploaded; `episode_index` and `frame_index` preserve frame alignment.

## Validation

Local validation: **PASS**. Source/staged/TsFile rows: 23,071 / 23,071 / 23,071; duplicate device-Time keys: 0; monotonicity violations: 0; TsFile/source Parquet size ratio: 53.53%. The Apache TsFile Java reader queried every row and inspected actual codec metadata.

## Minimal read example

```python
from tsfile import TsFileReader
reader = TsFileReader("data/qnoens_block2target.tsfile")
with reader.query_table("qnoens_block2target", ["episode_index", "task_index", "Time", "frame_index", "next_success", "action_0"], batch_size=65536) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()
```

Please cite the original Hugging Face dataset and Quinten Noens (`qnoens`).