| # 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 | |
|
|
