README.md

---
license: apache-2.0
authors:
  - williamdgomez
task_categories:
  - robotics
tags:
  - tsfile
  - timeseries
  - tabular
  - robotics
  - lerobot
  - manipulation
modality:
  - timeseries
  - tabular
pretty_name: SNOW.SB TsFile
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/williamdgomez_snow_sb.tsfile
size_categories:
  - 10K<n<100K
---

# SNOW.SB TsFile

This repository is an Apache TsFile conversion of the LeRobot dataset
[`williamdgomez/SNOW.SB`](https://huggingface.co/datasets/williamdgomez/SNOW.SB).
The original Hugging Face repository and its commit history identify
`williamdgomez` as the dataset author/uploader. The source dataset was created
with LeRobot v3.0 for the `lekiwi_client` robot.

## Source Dataset

- Dataset: [`williamdgomez/SNOW.SB`](https://huggingface.co/datasets/williamdgomez/SNOW.SB)
- Author/uploader: [`williamdgomez`](https://huggingface.co/williamdgomez)
- License: Apache-2.0
- Task: Pick up the block and place it in the bin
- Split: `train` (`0:91`)
- Sampling rate: 24 fps
- Scale: 91 episodes, 73,578 frames, 1 task
- Source data shards: 91 Parquet files under `data/chunk-000/file-*.parquet`
- Source video files: 273 MP4 files, 91 per camera stream

The source video streams are stored at [`videos/`](https://huggingface.co/datasets/williamdgomez/SNOW.SB/tree/main/videos):

- `videos/observation.images.front/chunk-000/file-{file_index:03d}.mp4`
- `videos/observation.images.wrist/chunk-000/file-{file_index:03d}.mp4`
- `videos/observation.images.top/chunk-000/file-{file_index:03d}.mp4`

Videos are not included in this TsFile repository. They remain available in
the original Hugging Face dataset and can be aligned with rows using
`episode_index`, `frame_index`, and the source `meta/episodes` offsets.

## Converted Files

- TsFile: `data/williamdgomez_snow_sb.tsfile`
- Table: `williamdgomez_snow_sb`
- Rows: 73,578
- Devices: 91 (`episode_index` x `task_index`)
- Time precision: milliseconds
- Metadata: `meta/` mirrors the source metadata; `meta/info.json` records the
  TsFile path and conversion mapping.

## Schema

`Time` is `round(timestamp * 1000)` and is stored in milliseconds. The source
`timestamp` is dropped after this conversion because it is exactly the seconds
representation of `Time`; `Time` restarts at zero for each episode.

| Column group | TsFile columns | Type / role |
|---|---|---|
| Time | `Time` | INT64/TIMESTAMP, TIME |
| Tags | `episode_index`, `task_index` | TAG/device segments |
| Frame metadata | `frame_index`, `sample_index` | INT64 FIELD; `sample_index` is source `index` |
| Action | `action_0` ... `action_8` | FLOAT FIELD, flattened from `action[9]` |
| Observation | `observation_state_0` ... `observation_state_8` | FLOAT FIELD, flattened from `observation.state[9]` |

Vector names preserve their source prefixes (`.` becomes `_`) and append a
zero-based element index. The three video features
(`observation.images.front`, `.wrist`, `.top`) are omitted from the TsFile
because TsFile stores the numeric time-series table while the original MP4
files remain at the source URL.

## Encoding and Compression

The local conversion used the type-aware TsFile writer profile requested for
this dataset:

- FLOAT/DOUBLE: `GORILLA` with `LZ4`
- INT32/INT64 and `Time`: `TS_2DIFF` with `LZ4`
- BOOLEAN (if present): `RLE` with `LZ4`
- TAG columns: TsFile table/device TAG mechanism

The generated TsFile is 1,456,157 bytes. The merged staged Parquet is
1,235,578 bytes, and the 91 source Parquet shards total 3,152,836 bytes.

## Validation

Apache TsFile Python SDK readback succeeded. The TsFile metadata row count and
query readback both equal the staged Parquet count: 73,578 rows. All 91
episode devices have monotonic `Time` values with no duplicate
(`episode_index`, `task_index`, `Time`) keys.

## Usage

```python
from tsfile import TsFileReader

reader = TsFileReader("data/williamdgomez_snow_sb.tsfile")
table = reader.get_all_table_schemas()["williamdgomez_snow_sb"]
columns = [c.get_column_name() for c in table.get_columns() if c.get_column_name() != "Time"]

with reader.query_table("williamdgomez_snow_sb", columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
```

## Citation

The original dataset card does not provide a paper or BibTeX citation.