# Video Tool Call SFT Dataset

## Contents
- `final_sft_dataset.json` : 16,522 SFT samples (Plan B 12525 + Seeker 3997, 8670 unique videos)
- `llava_video/videos/...` : Source mp4 files (138 GB total)

## How to merge & extract
This dataset is split into multiple ~40GB parts to fit HuggingFace's 50GB single-file limit.

```bash
# Download all parts (video_tool_call.tar.part-aa, ab, ac, ad, ...)
cat video_tool_call.tar.part-* | tar -xvf -

# Or stream-extract while downloading (Linux):
cat video_tool_call.tar.part-* | tar -xf -
```

## Schema
Each sample in `final_sft_dataset.json`:
```json
{
  "doc_id": "...",
  "video": "llava_video/videos/...",
  "question": "...",
  "options": ["A. ...", "B. ..."],
  "gt_answer": "A",
  "trajectory": {"turns": [...], "final_answer": "A", "n_turns": 2, "terminated": "answered"},
  "metadata": {"source": "...", "question_type": "...", ...}
}
```

## Source distribution
| count | source |
|---:|---|
| 4239 | llava178k_academic_2_3m_mc |
| 3342 | seeker:multi_clue_w_tool_same |
| 2980 | llava178k_academic_1_2m_mc |
| 2767 | llava178k_1_2_m_youtube_mc |
| 2067 | llava178k_nextqa_1_2m_mc |
| 655  | seeker:multi_clue_w_tool_diff |
| 472  | llava178k_2_3_m_youtube_mc |