EgoVITA Training Data

Training data for EgoVITA: Learning to Plan and Verify for Egocentric Video Reasoning (paper · code · model).

The plan-then-verify pipeline is trained in two stages: SFT (learn the two-turn <ego_plan>/<exo_verify> → <answer> format) then GRPO (reinforce it with dense rewards).

Data ownership & licensing

This repository contains only text annotations (questions, plan/verify/answer targets, and relative video filenames) produced for this project. No video is redistributed here. The underlying videos belong to their respective source datasets (HD-EPIC, EgoProceL / EGTEA Gaze+, CMU-MMAC, EPIC-Tents, MECCANO, and the EgoIT-99K sources: ADL, Charades-Ego, EgoTaskQA, HoloAssist, IndustReal). To use this data you must obtain each video from its original source and comply with that dataset's own license and terms.

Files

File	Rows	Stage
combined_sft_2turn.jsonl	12,879	Two-turn SFT
combined_rl_2turn.jsonl	40,201	GRPO (plan-then-verify)

Video paths in these files are relative (e.g. hdepic_clips/....mp4). You assemble a local videos/ directory from the public source datasets (below) and point the trainer at it. The videos themselves are not redistributed here — obtain them from their original sources under each dataset's own license.

Row schema

{
  "messages": [
    {"role": "system", "content": "You are an egocentric video reasoner..."},
    {"role": "user", "content": "<video>\n<question>\nOptions:\nA. ...\n...\nReason in two stages..."}
  ],
  "videos": ["hdepic_clips/P02-20240209-184316__426_465.mp4"],
  "solution": "<answer> C </answer>",
  "turn2_system": "You are a careful exocentric verifier and answerer...",   // RL only
  "turn2_prompt": "Given your first-person plan and third-person verification above, ...", // RL only
  "sample_id": "<md5>"
}

• SFT rows are the full two-turn conversation (system → user → assistant plan/verify → answer).
• RL rows carry turn2_system / turn2_prompt for the second turn, and solution (gold answer) for the answer reward. sample_id uniquely identifies each row.

Video sources

Paths use three relative roots. Assemble under one videos/ directory:

videos/
  hdepic_clips/     # HD-EPIC segments        (filename: <video>__<start>_<end>.mp4)
  ego_procel/       # EgoProceL: EGTEA_Gaze+, CMU-MMAC, EPIC-Tents, MECCANO, pc_assembly_disassembly
  egoit_clips/      # EgoIT-99K-derived: ADL, Charades-Ego, EgoTaskQA, HoloAssist, IndustReal

Obtain each under its own license/terms:

• HD-EPIC — https://hd-epic.github.io/
• EgoProceL — https://sid2697.github.io/egoprocel/
• EgoIT-99K sources — ADL, Charades-Ego, EgoTaskQA, HoloAssist, IndustReal (from their project pages)

hdepic_clips and egoit_clips are cut segments: the filename encodes <start>_<end> in seconds. If you only have the source long videos, cut the segments to match those spans (or adapt the data loader to seek [start, end] in the full video).

Download

pip install -U "huggingface_hub[cli]"
hf download yogkul2000/EgoVITA --repo-type dataset --local-dir ./egovita_data

Then place the assembled videos under ./egovita_data/videos/ with the three roots above.

Integrity check

cd egovita_data
python - <<'PY'
import json, os
for f in ["combined_sft_2turn.jsonl", "combined_rl_2turn.jsonl"]:
    tot=miss=0
    for l in open(f):
        for v in (json.loads(l).get("videos") or []):
            tot+=1
            if not os.path.exists(os.path.join("videos", v)): miss+=1
    print(f"{f}: {tot-miss}/{tot} videos found under videos/")
PY

Using with the training scripts

In ms-swift/egovita/sft_*.sh and rl_*.sh (see the code repo): set DATA= to the jsonl, run the trainer from the dir containing videos/ (or map the videos/ root onto the relative paths in the loader).

Teacher cache (for the confidence reward — regenerate, not shipped)

GRPO uses a confidence reward that, for the first EGOVITA_CONF_WARMUP (default 200) steps, compares the policy's own <exo_verify> against a teacher's verification. Those teacher verifications are read from a teacher_cache.jsonl keyed by sample_id. It is not included here (it is derivable and teacher-specific). To regenerate it:

1. Pick a teacher VLM (the paper uses a stronger instruct model to produce reference <exo_verify>).
2. Run precompute_teacher.py (in the code repo) over combined_rl_2turn.jsonl, producing rows {"sample_id": ..., "teacher_exo_verify": "..."}.
3. Point EGOVITA_TEACHER_CACHE= at the resulting file.

Without it, GRPO still runs — the confidence reward is simply 0 during the teacher-guided warmup and the run relies on format + answer + ACMG rewards until it switches to self-ranking after warmup.

Citation

@inproceedings{kulkarni2026egovita,
  title={EgoVITA: Learning to Plan and Verify for Egocentric Video Reasoning},
  author={Kulkarni, Yogesh and Fazli, Pooyan},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}