Datasets:

THULab
/

koch_test

license: unknown
task_categories:
  - robotics
language:
  - en
tags:
  - lerobot
  - robotics
  - imitation-learning
  - action-state
  - tutorial
  - tsfile
  - format:tsfile
  - timeseries
pretty_name: koch_test TsFile
size_categories:
  - 1K<n<10K

koch_test

This repository contains a TsFile conversion of the original Hugging Face dataset MoofireX/koch_test.

Source Dataset

The source repository contains one Parquet split and LeRobot metadata files:

data/train-00000-of-00001.parquet
meta_data/episode_data_index.safetensors
meta_data/info.json
meta_data/stats.safetensors

The converted repository keeps the same top-level structure:

data/train.tsfile
meta_data/episode_data_index.safetensors
meta_data/info.json
meta_data/stats.safetensors

Source Schema

The source Parquet split has 2,755 rows and 7 columns:

Column	Source type	Conversion
`observation.state`	fixed-size list float[6]	flattened to `state_0` ... `state_5`
`action`	fixed-size list float[6]	flattened to `action_0` ... `action_5`
`episode_index`	int64	FIELD
`frame_index`	int64	FIELD
`timestamp`	float	renamed to `episode_timestamp_s` FIELD
`next.done`	bool	renamed to `next_done` FIELD
`index`	int64	renamed to `sample_index` FIELD

Conversion Notes

Source format converted to TsFile: lowercase data/train-00000-of-00001.parquet.
Converted format: TsFile table model.
Output file: data/train.tsfile.
Table name: koch_test.
TAG column: episode_id, synthesized from episode_index.
Time column: synthesized from frame_index * 100 ms.
The original timestamp value is preserved as episode_timestamp_s.
observation.state and action are not dropped semantically; each 6-element vector is expanded to scalar fields because TsFile import requires scalar columns.
Row count is preserved: 2,755 source rows -> 2,755 TsFile rows.

Metadata

The source meta_data/info.json reports:

{
  "codebase_version": "v1.6",
  "fps": 30,
  "video": true,
  "encoding": {
    "vcodec": "libsvtav1",
    "pix_fmt": "yuv420p",
    "g": 2,
    "crf": 30
  }
}

The meta_data/ files are mirrored unchanged from the source repository for traceability. The TsFile conversion only uses the Parquet split.

Citation

Please cite the original dataset:

@dataset{koch_test,
  author = {MoofireX},
  title = {koch_test},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/MoofireX/koch_test}
}