level3/README.md

RoboSteer Level3 preprocessing assets

English | 中文说明

This folder releases three Level3 assets as portable tables and a reconstruction tool. It is independent of the sibling level1/ assets. The original server paths and video symlinks are not distributed.

Assets

Folder	Contents	Reconstructed layout
text_audio_instructions/	17,575 UTF-8 instructions across 7,764 tasks, including original audio/text modality and order	<task_id>/<item_id>__text.txt or __audio.txt
video_symlinks/	8,475 video references with paths relative to a user-provided public-data root, byte sizes and SHA-256	<item_id>.mp4
interleaved_input_list_for_GEM/	26,050 ordered items: 17,575 text items and 8,475 video references	<task_id>/order_NNN.txt or .mp4

No raw audio transcripts, audio files, image assets or video bytes are included. The audio-origin instructions are processed motion instructions, not raw Whisper transcripts. The original GEM preparation skipped 7,764 image items. Original order numbers are preserved, including gaps left by skipped images.

release_manifest.json records the comparison between the grouped text and the actual GEM text. If differences exist, text_overrides.parquet preserves them for exact reconstruction; the grouped instruction table remains unchanged.

Download

pip install huggingface_hub pyarrow
hf download PhoebeCC/RoboSteer-Preprocessing \
  --repo-type dataset --include 'level3/*' --local-dir ./RoboSteer-Preprocessing

For reproducibility, add --revision <commit-id> using the commit of the Level3 release.

External video dependency

Prepare the matching RoboSteer public data locally:

<public-root>/
└── Data_Processed/Level3/Video_Slices/
    └── <original segment filename>.mp4

The release identifies every required video using source_relpath, source_size and source_sha256. A matching folder layout alone is insufficient: each video must also match its recorded bytes. This release does not download or regenerate the original media.

Reconstruct the three original assets

python RoboSteer-Preprocessing/level3/scripts/materialize_level3.py \
  --public-root /data/public \
  --output-root ./restored_level3 \
  --video-mode symlink

The result is:

restored_level3/
├── text_audio_instructions/
├── video_symlinks/
└── interleaved_input_list_for_GEM/

This reproduces the selected assets directly under the output root. To match the organized project layout, pass --output-root /path/to/cxt/level3/inputs instead. Use a new output directory for initial validation.

The tool verifies release checksums and all source videos before writing. It writes exact UTF-8 text bytes, including existing newlines. It does not overwrite conflicting files. Existing matching files are accepted. Missing or different source videos produce a JSON error report on stdout and exit code 2. Conflicting or missing reconstructed outputs in verification mode produce exit code 3. Original input videos remain untouched.

Use --video-mode copy if symlinks are unavailable. This copies video bytes into both video layouts and therefore uses additional disk space. Default symlinks point to the supplied local public-data root; moving that root later requires rebuilding them.

Verify the restored assets:

python RoboSteer-Preprocessing/level3/scripts/materialize_level3.py \
  --public-root /data/public \
  --output-root ./restored_level3 \
  --verify-only

Verification checks all expected entries; it does not remove or reject unrelated extra files. It does not run GEM or evaluate video decoding or instruction quality.

Read the tables directly

import pyarrow.parquet as pq

instructions = pq.read_table(
    "RoboSteer-Preprocessing/level3/text_audio_instructions/data.parquet"
).to_pylist()

• Instructions: item_id, task_id, order, source_modality, file_name, text, sha256.
• Videos: item_id, task_id, order, file_name, source_relpath, source_size, source_sha256.
• GEM items: task_id, item_id, order, file_name, kind, sha256, source_relpath (null for text). Text is joined by item_id; optional overrides take precedence only for GEM reconstruction.

file_name is relative to its reconstructed asset directory. Media paths are relative to --public-root, never to the HF cache. The video table contains references, not embedded Video objects; viewing it on Hugging Face does not make external videos playable.

Validation and provenance

The text was exported from the existing Level3 preprocessing outputs. Video references were taken from the original symlink targets, and the task/order layout from the actual GEM input directories. Text and video integrity are recorded as SHA-256; no historical processing status values were reused as quality labels.

See release_manifest.json, validation.json and SHA256SUMS. The reconstruction utility requires Python 3.9+ and pyarrow. The release does not include Qwen/Whisper/GEM models, logs, server credentials, or raw absolute server paths. Dataset terms and upstream media conditions should be consulted at the repository and source-data level.