Datasets:
Social Signal Bench
A benchmark for co-occurring social signals and engagement in conversation, in video and audio.
Social Signal Bench asks a model which social signals, such as hesitation, skepticism or confidence, a speaker shows in a short clip, and whether the speaker is engaged. It has two datasets, Social Signal Bench Video (ten signals, audiovisual clips) and Social Signal Bench Audio (six signals, audio-only clips). This repository holds their public training sets; models are scored on private test sets with expert labels.
At a glance
| Social Signal Bench Video | Social Signal Bench Audio | |
|---|---|---|
| Size (train + test) | 10,173 clips, 25.1 h | 8,392 clips, 22.0 h |
| Labels | 10 social signals + engagement status | 6 social signals + engagement status |
| Public training set | 7,801 clips: 10.0 h expert, 9.5 h pseudo-labelled | 7,298 clips: 10.0 h expert, 9.6 h pseudo-labelled |
| Private test set | 2,372 clips, 85 recordings, four domains | 1,094 clips, 47 recordings, four domains |
| Expert annotation | at least 3 trained raters per clip | at least 3 trained raters per clip |
| Best model (EW-F1) | 44.0, post-trained Qwen2.5-Omni-7B | 58.1, post-trained Qwen2-Audio-7B-Instruct |
Contents
- Leaderboard
- News
- Task
- Benchmark composition
- Quick start
- Annotation
- Metrics
- Evaluation on the private test sets
- Reproduce
- Data format
- Use and limits
- Citation
- License
Leaderboard
Scores on the private test sets under a shared prompt and scorer. All numbers are percentages. The two datasets use
different signal sets, so compare scores only within a table. Machine-readable copies are in
leaderboard/. β marks models post-trained on Social Signal Bench.
Video (2,372 clips, four domains)
| Rank | Model | Access | Input | EW-F1 | Micro-F1 | Macro-F1 | Engagement macro-F1 |
|---|---|---|---|---|---|---|---|
| π₯ | β Qwen2.5-Omni-7B, post-trained on Social Signal Bench | open | video + audio | 44.0 | 39.2 | 31.0 | 58.4 |
| π₯ | Muse Spark 1.3 | API | video + transcript | 40.9 | 35.6 | 29.8 | 43.0 |
| π₯ | Gemini 3.8 Flash | API | video + audio | 40.6 | 35.2 | 29.4 | 35.1 |
| 4 | Qwen3.5-Omni-Plus | open | video + audio | 40.6 | 38.6 | 33.1 | 40.0 |
| 5 | Gemini 4 Argon | API | video + audio | 40.4 | 37.1 | 32.8 | 39.4 |
| 6 | GPT-6.1 Sol | API | 16 frames + transcript | 39.6 | 34.7 | 28.7 | 33.5 |
| 7 | Gemma-4-31B-it | open | 16 frames + transcript | 35.3 | 33.0 | 26.8 | 38.2 |
| 8 | Qwen2.5-Omni-7B (untuned) | open | video + audio | 20.6 | 13.7 | 10.2 | 20.7 |
Audio (1,094 clips, four domains)
| Rank | Model | Access | EW-F1 | Micro-F1 | Macro-F1 | Engagement macro-F1 |
|---|---|---|---|---|---|---|
| π₯ | β Qwen2-Audio-7B-Instruct, post-trained on Social Signal Bench | open | 58.1 | 54.7 | 47.4 | 45.8 |
| π₯ | Gemini 3.1 Pro | API | 54.0 | 51.2 | 44.7 | 46.5 |
| π₯ | Gemini 3.8 Flash | API | 51.0 | 47.1 | 43.2 | 35.8 |
| 4 | Qwen3.5-Omni-Plus | open | 47.5 | 44.1 | 42.0 | 47.2 |
| 5 | GPT-Audio | API | 44.4 | 42.3 | 37.8 | 37.7 |
| 6 | Kimi-Audio-7B-Instruct | open | 29.8 | 26.7 | 20.8 | 26.2 |
| 7 | Qwen2-Audio-7B-Instruct (untuned) | open | 28.7 | 20.7 | 17.8 | 12.4 |
| 8 | Audio-Flamingo-3 | open | 25.4 | 25.4 | 14.2 | 32.9 |
| 9 | Ultravox-v0.5-Llama-3.1-8B | open | 23.6 | 16.4 | 20.7 | 31.1 |
Leaderboard runs use an unpublished prompt with signal definitions, observable cues and decision rules; the public
prompt in prompts/ is a shorter version, so scores from it are not directly comparable. The
post-trained models were trained on internal training data from all four domains, of which the public release is
a subset. Muse Spark 1.3 returned 20 empty responses, scored as no signal; Qwen2-Audio-7B-Instruct (untuned) and
Audio-Flamingo-3 use constrained JSON decoding.
News
- 2026-11: Pseudo-labels complete for both configurations; dataset card, scripts and leaderboard added.
- 2026-10: Train configurations
videoandaudioreleased. This replaces the dev/test layout of v0.1 (tagv0.1).
Task
Each item is a short clip of one clearly visible (video) or audible (audio) main speaker. Given a clip, a system returns:
engagement_status: one ofengaged,neutral,disengaged;signals: the set of social signals present in the clip. An empty set is a valid answer: many clips of natural conversation show no signal.
The labels describe what trained observers perceive, not a speaker's inner state.
| Signal | Perceived behavior | Cluster | Video | Audio |
|---|---|---|---|---|
| Agreement | Visible alignment or affirmation | Agreement | β | β |
| Confidence | Assured, steady stance toward one's message | Confidence | β | β |
| Interest | Motivated attention to something new | Engagement & Interest | β | β |
| Disagreement | Open opposition to another's position | Skepticism & Disagreement | β | |
| Skepticism | Doubting stance toward another's statement | Skepticism & Disagreement | β | |
| Frustration | Annoyance from obstacles or blocked goals | Stress & Frustration | β | β |
| Stress | Tension and anxiety, effort to stay in control | Stress & Frustration | β | |
| Confusion | Momentary breakdown in understanding | Uncertainty & Hesitation | β | |
| Hesitation | Disruption in speaking or responding | Uncertainty & Hesitation | β | β |
| Uncertainty | Self-doubt, partial commitment to a message | Uncertainty & Hesitation | β | β |
Engagement and disengagement form the separate engagement head. Speech alone expresses Disagreement, Skepticism, Stress and Confusion less consistently, so the audio dataset leaves them out.
Benchmark composition
The public training sets contain English online video from Creative Commons licensed YouTube videos; every clip is 5 to 20 seconds long. The private test sets span four conversational domains (two collections of online video, sales calls, and dyadic conversations) and share no recording or channel with the training data.
| Dataset | Split | Labels | Clips | Hours | Sources | Empty |
|---|---|---|---|---|---|---|
| Video | Train (public) | Expert | 4,321 | 10.0 | 319 | 37.2% |
| Video | Train (public) | Pseudo | 3,480 | 9.5 | 453 | 19.0% |
| Video | Test (private) | Expert | 2,372 | 5.6 | 85 | 47.2% |
| Video | Total | 10,173 | 25.1 | 857 | 33.3% | |
| Audio | Train (public) | Expert | 3,974 | 10.0 | 312 | 17.2% |
| Audio | Train (public) | Pseudo | 3,324 | 9.6 | 447 | 12.8% |
| Audio | Test (private) | Expert | 1,094 | 2.4 | 47 | 29.3% |
| Audio | Total | 8,392 | 22.0 | 806 | 17.0% |
Sources are YouTube videos for the training data and parent recordings for the test data. Empty is the share of clips with no annotated signal at any probability.
Each row says where its labels come from in label_source:
human: expert labels projected from the event annotations of the source recording.model: pseudo-labels from a teacher model trained on a larger proprietary social-signal dataset that spans all four domains and was annotated with the same protocol. There is one teacher per configuration.
Pseudo-labels are more concentrated than expert labels: Confidence dominates, few clips carry two or more signals, and rare signals and disengagement come mainly from the expert-labelled clips (see Signal counts).
Quick start
from datasets import load_dataset
video = load_dataset("Interhuman/signal-bench", "video", split="train")
audio = load_dataset("Interhuman/signal-bench", "audio", split="train")
row = audio[0]
print(row["clip_id"], row["label_source"], row["engagement_status"])
print([s["name"] for s in row["signals"]]) # signals present at any probability
expert = audio.filter(lambda r: r["label_source"] == "human")
Annotation
Every expert-labelled clip, including every test clip, is annotated independently by at least three trained raters
with at least a bachelor's degree in behavioral or social science. They work from a written guide and annotate a
signal only when it is clearly observable, on spans of at least two seconds, with an intensity (1 to 5) and a
confidence (low, medium, high). Disagreement is measured against the taxonomy: a vote for Hesitation where peers
marked Uncertainty is partly supported, one where peers marked Agreement is not. Only votes with no support in the
same or an adjacent cluster go to expert review. Reviewed spans are merged into a reference that keeps every
signal with an evidence weight, and clips are cut into 5 to 20 second segments that keep the annotated signal time.
The probability of each signal in this release (high, medium, low) summarizes that evidence.
Metrics
| Metric | Where | Definition |
|---|---|---|
| Macro-F1 | public + leaderboard | Mean of per-signal F1 over the dataset's signals (10 video, 6 audio) |
| Micro-F1 | public + leaderboard | F1 over all signal decisions pooled |
| Engagement macro-F1 | public + leaderboard | Three-class macro-F1 over engaged, neutral, disengaged |
| Engagement accuracy | public | Share of clips whose engagement_status is predicted correctly; dominated by the frequent engaged class |
| EW-F1 | leaderboard only | Evidence-Weighted F1: weights each reference signal by its annotation evidence and gives graded credit to related signals (1 exact, 0.66 same cluster, 0.33 adjacent cluster); needs evidence weights that are not part of this release |
A reference signal counts at any probability. A class with no reference and no predicted positive has F1 0.
Use macro-F1 as the primary public metric. Confidence is so frequent that a constant answer scores well on
micro-F1: on the expert-labelled validation split of scripts/make_split.py (seed 0), a
model that always predicts Confidence reaches 42.2 micro-F1 but 5.9 macro-F1 on video, and 58.9 micro-F1 but 13.6 macro-F1
on audio.
Evaluation on the private test sets
The test sets stay private so that the leaderboard stays a fixed reference. Submissions are scored by the maintainers on the same clips with the same scorer, and results are versioned.
Reproduce
Requirements: Python 3.10 or newer, pip install -r scripts/requirements.txt, and a free Hugging Face account
for the space backend. The scripts read the parquet files directly and download them from the Hub when
--data-dir is not given; inside a clone of this repository, add --data-dir . to every command to use the local
copy instead.
# 1. Channel-disjoint validation splits for both datasets (expert-labelled validation clips;
# train = all clips of the other channels). Seed 0 gives 614 video and 632 audio validation clips.
python scripts/make_split.py --config video --out-dir splits
python scripts/make_split.py --config audio --out-dir splits
# 2a. Reference models: Inter-2 through its Hugging Face Spaces (no GPU needed; uses HF_TOKEN or `hf auth login`)
python scripts/infer.py --config audio --backend space \
--clip-ids splits/audio/val_ids.txt --out preds_inter2_audio.jsonl --limit 5 # smoke test
python scripts/infer.py --config video --backend space \
--clip-ids splits/video/val_ids.txt --out preds_inter2_video.jsonl --limit 50 --resume
# 2b. Your own model: implement predict(media, config, prompt) -> answer text (see examples/my_model.py)
python scripts/infer.py --config audio --backend python --entry examples/my_model.py:predict \
--clip-ids splits/audio/val_ids.txt --out preds_mine.jsonl
# 3. Score, on the same clips the predictions cover
python scripts/score.py --config audio --predictions preds_mine.jsonl --clip-ids splits/audio/val_ids.txt
python scripts/score.py --config video --predictions preds_inter2_video.jsonl --clip-ids splits/video/val_ids.txt
A run with --limit covers only part of the validation split; the clips it skipped count as empty answers, so
score a partial run only to check the pipeline, or continue it with --resume until it covers the split.
--resume keeps the finished rows of --out, retries clips that failed (for example when the GPU quota ran out)
and continues with the rest.
The space backend sends each clip to the Inter-2 playground Spaces:
Inter-2 Omni for video (the clip with its
audio track, task "Social signals" with the model's own system prompt) and
Inter-2 Audio for audio (task "Combined";
only engagement and the six signals are scored). Decoding is greedy and the answer is constrained to the task's
JSON schema. The Spaces run on ZeroGPU, so every clip counts against your daily Hugging Face GPU quota: use them
for spot checks and reference runs on a subset, and --resume to continue a run across days. The Inter-2 Omni
model is the post-trained video model of the leaderboard; the Inter-2 Audio model continues the post-trained audio
model with emotion-recognition training, so it is not the exact audio leaderboard checkpoint.
score.py scores against expert labels by default (--label-source human). A clip without a prediction, with a
failed inference or with an unparsable answer counts as an empty answer with neutral engagement, as on the
leaderboard; score.py prints how many clips fall in each case.
Data format
Dataset rows
| Field | Type | Description |
|---|---|---|
clip_id |
string | Unique clip identifier |
youtube_id, youtube_url |
string | Source video |
channel_id, channel_title, video_title |
string | Source channel and video |
published_at |
string | Publication date of the source video |
start_ms, end_ms, duration_ms |
int | Clip position inside the source video, and clip length |
transcript |
string | Transcript of the clip |
engagement_status |
string | engaged, neutral or disengaged; null if the clip is not labelled |
signals |
list | {name, probability} with probability high, medium or low; an empty list means labelled with no signal, null means not labelled |
selection |
string | How the clip was selected for annotation |
media_license |
string | Licence of the media clip (CC BY 3.0) |
label_source |
string | human (expert) or model (pseudo-label) |
video / audio |
struct | mp4 clip with its audio track (video config) or 16 kHz mono wav (audio config) |
Predictions (scripts/score.py input)
One JSON object per line:
{"clip_id": "-00eAb6FnJE_w005__s002", "engagement_status": "neutral", "signals": [{"name": "Hesitation", "probability": "high"}]}
signals may also be a list of names. Names, probabilities and engagement_status are matched case-insensitively;
unknown names are ignored. Rows written by infer.py also carry parse_ok, error, raw and latency_s.
Missing values
A value is only filled from a source record and is never estimated. Where no record exists, the column is empty:
| Column | video empty |
audio empty |
|---|---|---|
start_ms |
4,321 | 3,974 |
end_ms |
4,321 | 3,974 |
transcript |
4,321 | 3,974 |
published_at |
7,108 | 6,200 |
selection |
3,313 | 3,077 |
engagement_status, signals |
0 | 2 |
channel_title |
0 | 0 |
video_title |
0 | 0 |
The two audio rows without engagement_status and signals are pseudo-labelled clips that carry no label;
the scripts skip them.
The expert-labelled clips were cut from annotated source segments, and their offsets inside the full YouTube video, their transcripts and their publication dates are not filled yet. A later commit will fill them without changing the schema.
Signal counts
Clips in which the signal is present at any probability.
| Signal | Video expert | Video pseudo | Audio expert | Audio pseudo |
|---|---|---|---|---|
| Agreement | 170 | 149 | 170 | 86 |
| Confidence | 1,440 | 1,832 | 2,409 | 2,555 |
| Interest | 349 | 272 | 506 | 168 |
| Disagreement | 170 | 21 | ||
| Skepticism | 266 | 71 | ||
| Frustration | 410 | 141 | 565 | 233 |
| Stress | 201 | 18 | ||
| Confusion | 139 | 20 | ||
| Hesitation | 461 | 336 | 771 | 490 |
| Uncertainty | 681 | 367 | 835 | 236 |
Use and limits
- Intended use: research on social-signal recognition and engagement, model development on the training sets, and comparison on the hosted test sets.
- What the labels mean: perceptions of trained observers, not clinical or ground-truth mental states.
- Coverage: the public data is English online video with one clearly visible speaker; the test sets add sales calls and dyadic conversations, so models trained only on this release are also tested across domains.
- Not allowed: surveillance without consent, decisions with high stakes for individuals (hiring, education, workplace monitoring), identifying people, and cloning voices or faces.
- Pseudo-labels come from a model and carry its errors; validate on the expert-labelled clips.
Citation
@article{interhuman2026socialsignalbench,
title = {Social Signal Bench: Evidence-Weighted Evaluation of Co-occurring Social Signals in Video and Audio},
author = {Siddharth Ravi, Anders R. Bargum, Hugo Anselme, Patrizio Cugia di Sant'Orsola, Alberto Cardaci, Flora Kincso Parragi, Vincent Borowiec, Line H. Clemmensen},
journal = {arXiv preprint arXiv:2611.22704},
year = {2026},
url = {https://arxiv.org/abs/2611.22704}
}
License
The labels, metadata and code are licensed under CC BY-NC 4.0, for non-commercial use only. Each media clip keeps the CC BY 3.0 licence of its YouTube video. The source channel, title and YouTube URL are on every row.
- Downloads last month
- 127