tianxia2/egolife-judge-sft-842-question-split

EgoLife six-user QA judge SFT data (842-question split)

Training and held-out test data for binary PASS/FAIL judges (formality, groundedness, answerability) over six-user multiple-choice questions generated from EgoLife egocentric video. This is the data used by multi-user/training/judge_sft (branch xth of GODAVID0819/EgoQA-two-user).

Media notice: the frames are sampled from EgoLife videos and remain subject to the EgoLife dataset license and terms of use. Keep this repository private unless redistribution is permitted.

Layout

manifests/
  train/train.jsonl            # 2694 judge rows (training split)
  train/data_audit.json
  train/prompt_snapshots.json
  train/excluded_task_rows.jsonl
  test/train.jsonl             # 672 judge rows (held-out test split; file name kept from the pipeline)
  test/...
manifests_video_split/         # same rows, re-split so train and test share no video
  train/train.jsonl            # 2694 rows, 96 packets
  test/train.jsonl             # 672 rows, 24 packets
  split_audit.json             # test packet list and per-task/label counts
  split_report.html            # visual check: counts, day balance, timeline, CLIP similarity, frame pairs
packets/
  <packet_id>.tar              # 120 tars, one per 600 s six-user packet

Each packets/<packet_id>.tar extracts to

<packet_id>/packet.json        # users, frame order, packet-relative frame paths
<packet_id>/checksums.sha256
<packet_id>/frames/user_<i>_<AGENT>/frame_XXXX_<t>s.jpg   # 300 frames per user, 0.5 FPS

Splits

Both splits contain exactly the same rows and the same per-task, per-label counts; only the assignment of rows to train/test differs.

split	formality	groundedness	answerability	total
train	672	674 (456 pass / 218 fail)	1348	2694
test	168	168 (117 pass / 51 fail)	336	672

• manifests/ (question split): no QA candidate appears in both splits, but test questions reuse video packets seen in training (94 of the 120).
• manifests_video_split/ (video split): train uses 96 packets and test 24; no packet is shared. Packets whose 10-minute windows overlap in time (29 pairs, up to 7 minutes) are always kept in the same split, so no footage appears on both sides. Every day DAY1–DAY7 has test packets (18–22% of that day's groundedness rows). Test packets look like training packets without sharing footage: the median best-frame CLIP cosine of test frames against train frames is 0.917, versus 0.906 for train packets against train packets outside their own overlap group. Adjacent (back-to-back) windows may fall on different sides, e.g. test 11:20–11:30 and train 11:30–11:40.

Using the frames

Manifest rows reference packets by absolute cluster path in frame_packet (/scratch/hm2991/egolife_rlhf_evidence_v1/packets/<packet_id>). After extracting the tars, rewrite that prefix to your local directory:

mkdir -p packets_extracted && for t in packets/*.tar; do tar -xf "$t" -C packets_extracted; done
python - <<'PY'
import json, pathlib
root = pathlib.Path("packets_extracted").resolve()
for split in ("manifests/train", "manifests/test",
              "manifests_video_split/train", "manifests_video_split/test"):
    p = pathlib.Path(f"{split}/train.jsonl")
    rows = [json.loads(l) for l in p.open()]
    for r in rows:
        if r.get("frame_packet"):
            r["frame_packet"] = str(root / pathlib.Path(r["frame_packet"]).name)
    p.with_name("train.local.jsonl").write_text("".join(json.dumps(r, ensure_ascii=False) + "\n" for r in rows))
PY