---
language:
- en
license: cc-by-4.0
pretty_name: OSSL-v2
tags:
- video-to-music
- music-generation
task_categories:
- other
---

# Open Screen Soundtrack Libary Version 2 (OSSL-v2)

Paired **film video ↔ soundtrack music** clips for video-to-music generation.

**Paper:** [Dialogue-Aware Video-to-Music Generation Using Public Domain Film Collections](https://huggingface.co/papers/2608.11576)

## Layout

```
ossl-v2-hf/
├── metadata.csv                     # one row per movie (film + source metadata)
├── splits/{train,test}.txt          # clip_ids per split
├── public_train_test_remapped.pkl   # {"train":[clip_id...], "test":[clip_id...]}
├── train/{video,audio}/<movie_id>/<clip_id>.{mp4,wav}
└── test/{video,audio}/<movie_id>/<clip_id>.{mp4,wav}
```

Media are grouped into per-movie subfolders (`<movie_id>/`), where `movie_id` is
the prefix of `clip_id`. A clip's files are therefore at
`<split>/video/<movie_id>/<clip_id>.mp4` and `<split>/audio/<movie_id>/<clip_id>.wav`.

## Splits

| Split | Clips | Movies |
|---|---|---|
| train | 31,011 | — |
| test | 3,332 | — |
| **total** | **34,343** | **1,812** |

## `metadata.csv`

One row per movie (`movie_id` 0…1811, ascending). Clip-level info is recoverable
from `splits/{train,test}.txt` and the `clip_id` prefix (`<movie_id>_<clip_idx>_<start>_<end>`).

| column | description |
|---|---|
| `movie_id` | dense movie index, `0…1811` (matches the `clip_id` prefix) |
| `split` | `train` or `test` (a movie's clips never cross splits) |
| `n_clips` | number of clips from this movie |
| `movie_title`, `year`, `genres`, `imdb_id`, `imdb_rating`, `num_votes`, `runtime_min`, `release_date`, `director`, `title_type`, `imdb_url`, `imdb_description` | film metadata (IMDb) |
| `youtube_id`, `youtube_title`, `youtube_url`, `youtube_duration`, `youtube_published`, `youtube_view_count`, `youtube_channel`, `youtube_channel_id`, `youtube_description`, `youtube_thumbnail` | source YouTube video |
| `imdb_list_position`, `imdb_list_created`, `imdb_list_modified` | IMDb-list bookkeeping (provenance) |


## Loading

```python
from datasets import load_dataset
ds = load_dataset("McAuley-Lab/OSSL-v2", split="train")
```

## Citation

If you use OSSL-v2, please cite:

```bibtex
@inproceedings{osslv2,
  title     = {Dialogue-Aware Video-to-Music Generation Using Public Domain Film Collections},
  author    = {Haven Kim, Zachary Novack, Julian McAuley, Hao-Wen Dong},
  booktitle = {arXiv preprint arXiv:2608.11576},
  year      = {2026}
}
```