# Video Cold-Start CoT Dataset (OMNEX-VL format)

Generated 2026-05-31 for the video-hallucination project's Stage-1 cold-start SFT.

## Files
- `coldstart_cot_all.json`  — 12,262 valid samples (master, post-filter)
- `coldstart_cot_10k.json`  — 10,000-sample training slice (seed=0 shuffle of master)
- `coldstart_cot.raw.jsonl.gz` — every generation attempt + reject reasons (audit)
- `build_coldstart_cot.py` / `sft_coldstart.py` / `coldstart_sft.sh` — exact code used

## Schema (per sample, matches OMNEX-VL sft.py)
{ "problem", "data_type":"video", "path":<abs mp4>,
  "process_and_answer": "<prethink>...</prethink>\n<caption>...</caption>\n<think>...</think>\n<answer>...</answer>",
  "meta": {"idx","source","GT"} }

## How it was built
- Source: TemporalBench short_qa (9,740) + long_qa (2,522), MCQ with letter GT.
- Teacher: Qwen2.5-VL-72B-Instruct (vLLM, TP=8), video sampled fps=2 / max 64 frames.
- GT given to teacher at gen-time (LLaVA-CoT style), stripped from stored prompt.
- Rejection sampling: 4 tags present AND <answer>==GT; quality filters
  (answer-leak phrases, min stage length, n-gram repetition). Pass rate 98.2%.