Tphuc15's picture
Document private DFlash raw subset
c095b93 verified
|
Raw History Blame Contribute Delete
2.39 kB
---
pretty_name: DFlash post-training raw video subset
language:
- en
task_categories:
- video-text-to-text
tags:
- video
- dflash
---
# DFlash post-training raw video subset
This is a bounded raw-data subset prepared for academic experiments on DFlash post-training. It contains the selected original video archives and original caption/open-ended QA annotation files from [lmms-lab/LLaVA-Video-178K](https://huggingface.co/datasets/lmms-lab/LLaVA-Video-178K), pinned to revision `6d8c562dc26d70042a0d9704d1cae58c94b89098`.
## Contents
- 8 original video archives and 12 original annotation JSON files; total source payload is about 44.04 GB.
- 7551 video archive entries, 7536 unique video IDs, and 7536 annotated unique IDs after excluding overlap with the local VDC test metadata.
- `raw_video_inventory.jsonl` maps each source archive member to its original annotation row references. `metadata/collection_summary.json` records the catalog counts.
- `metadata/SHA256SUMS` records the checksums of the downloaded source files. `source/README.md` preserves the upstream dataset card.
| Upstream duration group | Archive entries |
|---|---:|
| `0_30_s_youtube_v0_1` | 3900 |
| `1_2_m_academic_v0_1` | 366 |
| `1_2_m_youtube_v0_1` | 960 |
| `2_3_m_academic_v0_1` | 200 |
| `2_3_m_youtube_v0_1` | 504 |
| `30_60_s_youtube_v0_1` | 1621 |
## Use and attribution
The upstream dataset card states that use is limited to academic research and education. This private copy does not broaden those terms. Review `source/README.md` and the [upstream dataset card](https://huggingface.co/datasets/lmms-lab/LLaVA-Video-178K) before use. Please cite the upstream dataset and its paper, [Video Instruction Tuning With Synthetic Data](https://arxiv.org/abs/2410.02713).
This repository contains raw source archives and annotations only. It does not contain generated target-model responses, model logits or hidden states, feature caches, decoded-video caches, calibrated visual-token budgets, or train/validation/test splits. Duration groups are source labels and are not calibrated visual-token budgets.
## Download from the training server
Grant the server's Hugging Face token read access to this private dataset, then use:
```python
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="Tphuc15/dflash-posttraining-raw",
repo_type="dataset",
token=True,
)
```