Joshkn/Distilled_handvqa

Distilled HanDyVQA

This dataset contains distilled training trajectories for hand-object interaction questions from HanDyVQA. HanDyVQA uses egocentric video from Ego4D. We selected four question categories and distilled each selected sample with an adapted O-VAD pipeline. This is a training derivative, not an official HanDyVQA release or an independent evaluation set.

Selection and video extraction

The source is train.jsonl at HanDyVQA revision 30d7f45cbcd14cfbd28862e50e865a5b8f878d98. Its 1,116 training questions include six categories. We retained every question in process, state, location, and parts, and excluded action and objects. No additional random subsampling was applied to these four categories.

Category	Selected and accepted samples	SFT rows
process	189	756
state	187	748
location	190	760
parts	184	736
Total	750	3,000

Each question identifies its source using video_id, custom_id, narration_uid, and time. The upstream extraction script requests a five-second interval, from time - 2.5 to time + 2.5. It resizes the clip to height 480, preserves the aspect ratio, and uses an even output width. media_manifest.jsonl records the source identifiers, requested intervals, observed clip properties, and original clip hashes. Actual frame counts can differ at clip boundaries.

Distillation

Stage 1 identifies relevant entities and creates segmentation masks with SAM3. Stage 2 uses native SAM2.1 tracking and records temporal observations and structured states. Stage 3 produces a report and an answer option. Automatic runs use a GPT-based teacher and can use label-informed corrective retries. Acceptance checks include answer agreement and structural consistency. Acceptance does not guarantee that every mask or visual inference is correct.

Five samples use Codex visual review instead of GPT API descriptions: parts_0879, parts_0986, parts_1819, process_1276, and process_1565. These recoveries still use real SAM3 segmentation and native SAM2.1 tracking. They use manually selected segmentation prompts and, where recorded, user clarification. They are not blind evaluations or identical to fully automatic runs. See codex_distilled_samples.json for their provenance.

The dataset owner corrected process_1565 from answer index 2 (upward) to 3 (downward). Indices are zero-based. label_corrections.json preserves this change. The report still distinguishes visible hand motion from the obscured switch endpoint and unverified electrical state.

Files and training schema

• sft_training_summary.json: one readable JSON array containing all 3,000 SFT rows.
• sft_training_summary.jsonl: the same rows in JSONL format, suitable for dataset loaders.
• sft_training_original.jsonl: the corresponding export with fuller Stage 2 observations.
• categories/{category}/: category-specific SFT files and build indices.
• archives/{category}.tar.gz: Stage 1/2/3 records, selected indexed masks, native tracking RLE, frame mappings, and response records.
• selected_train.jsonl, build_input.jsonl, and media_manifest.jsonl: selection, build, and source-media indices.
• source_provenance.json, codex_distilled_samples.json, label_corrections.json, and validation.json: provenance and checks.
• code/: the SFT builders and a helper for local rebuilding.

Every SFT row has exactly five fields: system, instruction, input, history, and output. Each question produces four rounds: recog, caption, track, and the final answer. The first three assistant targets use seeded tool-call templates. They are not verbatim teacher reasoning. The summary export constructs its overall description from Stage 2 structured fields without another model call. It removes detailed clip lists from SFT history. The archived stage records retain those details.

input contains a dataset-root-relative video path, not embedded video data. All public build indices also use dataset-root-relative paths. Use code/rebuild_sft.py to resolve them before rebuilding. Response records document API outputs or explicitly marked Codex annotations. Do not treat all response records as training targets.

Source media and local use

This public release contains annotations, not Ego4D RGB videos, extracted RGB frames, or contour-overlay images. The selected mask PNGs contain only segmentation IDs. Tracking masks use RLE encoding. Original pixels, together with these masks and frame mappings, are needed to reconstruct overlays. Private runtime contexts, credentials, model weights, failed attempts, and local experiment backups are excluded.

Obtain source videos through the official Ego4D access process. Then use the pinned HanDyVQA extraction script with selected_train.jsonl and --output_root videos. Extract the annotation archives from the dataset root:

for category in process state location parts; do
  tar -xzf "archives/${category}.tar.gz"
done
python code/rebuild_sft.py --dataset-root . --output-dir rebuilt

The builders check original video hashes. Re-encoding with a different FFmpeg version can change those hashes. A hash mismatch requires checking the reconstructed media. Do not disable the integrity check to hide a mismatch. The released SFT JSON can be loaded independently, but visual training still requires the corresponding licensed videos.

Attribution and terms

HanDyVQA's dataset card specifies CC BY-SA 4.0 for its annotations. This derivative annotation release retains that license and identifies its changes above. The included software retains its MIT notice in code/LICENSE. That software license does not relicense the source data. Ego4D media remains subject to the Ego4D agreements. This repository grants no additional rights to redistribute that media.

Please credit HanDyVQA and the Ego4D Consortium when using this derivative. Ego4D attribution: Ego4D Consortium (2020), Egocentric Live 4D Perception (Ego4D) Database, a resource for research on first-person activities. See also Grauman et al., Ego4D and the O-VAD adaptation.