jzpeng/typoverflow_vifailback

VIFailBack (TsFile)

This dataset is an Apache TsFile conversion of typoverflow/vifailback, a LeRobot v3.0 ALOHA AgileX Piper robot dataset.

Modalities: Time-series, Tabular. The converted repository contains numeric robot state, target, timing, episode/task tags, debug signals, and mirrored source metadata. Camera videos remain in the original Hugging Face dataset.

Source Dataset and Author

• Original dataset: typoverflow/vifailback
• Pinned source revision: 9145e6c58471ec61d5a11d7a1bbb39d540fc5710
• Original author, repository owner, uploader, and commit author: Chenxiao Gao (typoverflow)
• License: MIT
• Robot type: aloha_agilex_piper
• LeRobot codebase version: v3.0
• Sampling rate: 25 fps
• Split: train, episodes 0 through 4713
• Source scale: 4,714 episodes, 1,190,262 frames, 98 tasks
• Source frame shards: 8 Parquet files totaling 516,628,399 bytes
• Paper, homepage, and completed citation: not provided by the source card

The source card states that the dataset was created using LeRobot. Task text for all 98 task indices is preserved in meta/tasks.parquet; the source card does not provide a more detailed dataset narrative, so none is inferred here.

Converted Files

• TsFile: data/typoverflow_vifailback_train.tsfile
• Table: typoverflow_vifailback_train
• Rows: 1,190,262
• Episode/task devices: 4,714
• Columns: 183 total — 1 TIME, 2 TAG, 180 FIELD
• TsFile size: 467,828,268 bytes (90.55% of the source frame-Parquet size)
• Time precision: milliseconds
• Metadata: meta/ is mirrored from the pinned source, with only meta/info.json rewritten to describe the TsFile artifact and conversion.
• Conversion config: typoverflow_vifailback.yaml

The latest pinned source metadata includes meta/critical_frames.jsonl and meta/qc_warnings.jsonl; both are preserved unchanged.

TsFile Schema

Time is an INT64 millisecond timestamp computed as round(timestamp * 1000). It restarts within each episode and spans 0..36,960 ms across the dataset.

TAG columns (TsFile table-model device segments):

• episode_index — source episode number, 0 through 4713
• task_index — source task number, 0 through 97

Scalar FIELD columns:

• frame_index — frame number within an episode
• sample_index — source global index, renamed to avoid ambiguity

The remaining 178 FLOAT FIELD columns are flattened from the source's scalar and vector signals while preserving their full source prefixes. Dots are replaced by underscores and vector positions receive a numeric suffix. They cover these source groups:

• raw_state.* and raw_target.*: joint position/velocity, end-effector XYZ and quaternion, gripper state, and base velocity
• state.* and target.*: joint position/velocity, end-effector XYZ and 9D rotation, gripper state, and base velocity
• debug.*: left/right gripper end-effector XYZ and 6D rotation

For example, raw_state.left_joint_pos[6] becomes raw_state_left_joint_pos_0 through raw_state_left_joint_pos_5. The complete machine-readable list and roles are recorded in meta/info.json.

Encoding and Conversion Notes

• FLOAT/DOUBLE FIELD: GORILLA + ZSTD
• INT32/INT64 FIELD: TS_2DIFF + ZSTD
• Time: TS_2DIFF + LZ4
• BOOLEAN FIELD policy: RLE + LZ4; this source schema contains no BOOLEAN field
• TAG: stored through the TsFile device/TAG mechanism, not duplicated as ordinary FIELD columns
• All eight source frame shards are merged into one train-split TsFile. Filter by episode_index and task_index to select a trajectory or task.
• The source timestamp column is the only numeric source value not stored as a separate FIELD. It is redundant after the exact millisecond mapping (timestamp = Time / 1000 seconds at 25 fps).
• The source index is retained as sample_index; frame_index is unchanged.
• No published rows, episodes, tasks, or numeric signal dimensions are dropped.

Videos

Videos are not duplicated in the converted repository. The pinned source has 145 frame-aligned MP4 files totaling 27,938,508,868 bytes (26.02 GiB) in three streams:

• videos/observation.images.cam_high/ — 56 MP4 files, 11,040,155,298 bytes (10.28 GiB)
• videos/observation.images.cam_left_wrist/ — 45 MP4 files, 8,526,126,153 bytes (7.94 GiB)
• videos/observation.images.cam_right_wrist/ — 44 MP4 files, 8,372,227,417 bytes (7.80 GiB)

The source template is videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4. Numeric rows remain aligned with the original videos through episode_index, frame_index, and the source episode metadata under meta/episodes/.

Validation

The generated TsFile passed Apache TsFile Java 2.2.1 schema inspection and a full 1,190,262-row query readback. Validation confirmed:

• source, staged, and TsFile row-count equality
• exactly two TAG columns and 180 FIELD columns
• the requested physical encoding/compression profile
• TsFile size below the combined source frame-Parquet size
• SHA-256 4580f9bfa834395dba4298be61761a3a5d97dd77d93f009e7c3eef0cbad493a8

The local JSON and Markdown validation reports are intentionally not included in this upload-ready dataset.

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/typoverflow_vifailback_train.tsfile")
table_name = "typoverflow_vifailback_train"
columns = ["episode_index", "task_index", "frame_index", "sample_index"]

with reader.query_table(table_name, 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 citation. Cite the original Hugging Face dataset and Chenxiao Gao (typoverflow) when using this converted artifact.