| --- |
| license: cc-by-nc-4.0 |
| task_categories: |
| - audio-classification |
| - video-classification |
| language: |
| - en |
| tags: |
| - conversational-ai |
| - full-duplex |
| - turn-taking |
| - benchmark |
| - videofdb |
| - seamless-interaction |
| pretty_name: OpenVideoFDB (manifests) |
| size_categories: |
| - n<1K |
| --- |
| |
| # OpenVideoFDB manifests |
|
|
| Clip manifests for **OpenVideoFDB**, an open implementation of NVIDIA's VideoFDB benchmark |
| for full-duplex audio-visual conversational agents |
| ([code](https://github.com/OpenHRIai/OpenVideoFDB) · [paper](https://arxiv.org/abs/2605.30256)). |
|
|
| This dataset contains **no media**. Each row names a window inside a public recording of |
| Meta's [Seamless Interaction](https://huggingface.co/datasets/facebook/seamless-interaction) |
| corpus; the clips are re-cut locally and deterministically with |
| `uv run ovfdb rebuild --manifest <file>`. |
|
|
| ## Files |
|
|
| | file | rows | what | |
| |---|---|---| |
| | `ovfdb_v0.jsonl` | 118 | first mined set: 5 dialogue dynamics (60 rows) + 6 nonverbal dynamics (58 rows). Day-1 windows were selected using the partner's reaction, so the human reference passes the timing rules by construction; kept for reproducibility of `results/v0_summary*.md`. | |
| | `ovfdb_v1.jsonl` | 60 | the 5 dialogue dynamics re-mined with user-cue-only rules and fixed post-cue windows. | |
| | `ovfdb_v1_reviewed.jsonl` | 53 | v1 after a human review (7 rejects dropped). Use this one. | |
| | `reviews/ovfdb_v1.jsonl` | 56 | the raw human decisions (accept / reject / unsure, notes, window fixes). | |
|
|
| ## Row schema (one row = one clip × bucket) |
|
|
| `clip_id` (Seamless `file_id` + `_cNN`), `conversation_id`, `bucket` (perception | |
| generation), `dynamic_type` (one of the 11 VideoFDB dynamics), `dynamic_start_s`, |
| `dynamic_end_s`, `duration_s` (clip-relative seconds), `user_audio_path`, |
| `user_video_path`, `partner_audio_path`, `partner_video_path` (local paths after rebuild), |
| `system_prompt` (agent-side prompt, GPT-4o), `agent_speaks_first`, `av_caption` |
| (Gemini Flash caption of the side-by-side clip), `user_words` (Seamless word timings, |
| clip-relative), `source_meta` (`user_file_id`, `partner_file_id`, raw `t0`/`t1`, |
| `mining_rule`, rule evidence, VLM verdict, review decision). |
|
|
| ## Dynamics |
|
|
| Perception: Pause Handling, Gaze Avoidance with Pause, Adaptor Handling, Nonverbal |
| Interruption, Face Emotion Display, Nonverbal Backchanneling, Laughter. |
| Generation: Face Emotion Display, Nonverbal Backchanneling, Laughter, Verbal Interruption, |
| Verbal Backchanneling, Turn-taking, Emotion Matching. |
|
|
| ## Known limitations |
|
|
| - Source recordings are studio face-to-face conversations (two cameras), not video calls. |
| - Captions are machine-generated and occasionally attribute speech to the wrong speaker; the |
| evaluation payload tells the judge to trust the per-channel transcripts for who said what. |
| - Verbal Backchanneling windows in v1 are breath points inside long user turns; real |
| listeners say "mhm" there only ~10 % of the time (the paper's annotated windows give 69 %). |
| - Audio bleed between the two microphones can make a partner's laugh look like the user's. |
|
|
| ## Licence and attribution |
|
|
| CC BY-NC 4.0, the licence of the source corpus. Please attribute Meta AI's Seamless |
| Interaction dataset (Agrawal et al., 2025) and cite the VideoFDB paper |
| (arXiv 2605.30256). The NVIDIA VideoFDB dataset itself is gated and non-redistributable and |
| is not part of this release. |
|
|
