Video_tool_call / README.md
RUBBISHLIKE's picture
Add README
7035d50 verified
|
Raw
History Blame Contribute Delete
1.22 kB

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.

# 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:

{
  "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