fluid-concepts/friend-bench

Can a model — or a human — tell how two people are related from a 20-second clip of how they interact?

🌐 Built on Seamless Interaction

FriendBench is a suite of benchmarks for social perception from thin-slice dyadic interaction — inferring facts about two people's relationship from a brief clip of how they interact, built on the Seamless Interaction dataset. Each released set is a config of this repository.

• 🎧 Multi-modal — text, audio, and video for every clip
• 🎯 Objective label — records whether a pair had actually met before, not a viewer's impression
• 🧪 Matched human + model baselines — ~90 crowd raters per modality and 26 zero-shot models
• ⏱️ Controlled design — same ice-breaker prompt for every dyad, balanced and stratified

Released sets

Config	Task	Modalities	N	Status
validation	Prior familiarity (familiar vs. strangers)	text · audio · video	96	released

Additional configs (e.g. larger familiarity sets and interpersonal-circumplex relationship-type sets) may be added to this repository in future releases.

Load a set by its config name:

from datasets import load_dataset
ds = load_dataset("fluid-concepts/friend-bench", "validation")   # the default config

Paper scope. The MINT 2026 paper describes validation only — its task, design, human ratings, and model baselines all refer to that config. Any configs added later are related sets released separately and not covered by that paper. Cite validation specifically when reproducing or comparing to the paper.

Config names double as version tags: a future revision is added as a new config (e.g. validation_v2) beside the existing one, which stays loadable for reproducibility.

---

Config: validation

The dataset evaluated in the MINT 2026 paper. A balanced benchmark for inferring prior familiarity between two people from a brief, thin-slice clip of how they interact. Given one 20-second clip of a dyad answering an "Either-Or" ice-breaker prompt, the task is to judge whether the two people are familiar (they know each other) or strangers (they are meeting for the first time).

The clip is the only evidence. Every dyad answers the same type of prompt, so the topic of conversation cannot reveal the answer. The signal is in the manner of the interaction, not its content. The set is released in three modalities: text, audio, and video.

🎬 Example clips

Familiar (sample_000)

Strangers (sample_016)

(Videos render on the Hugging Face dataset page once published.)

What's in the set

• 96 dyads, one clip each (mean 20.2 s, range 15.6–25.0 s).
• Fully crossed and balanced across three binary factors (12 cells × 8 dyads):
  • relationship — 48 familiar / 48 stranger (the label).
  • gender_config — 48 mixed / 48 same-gender (see the note on estimated gender below).
  • temporal_position — 48 early / 48 late (where the clip falls in the interaction).
• Balanced across three recording sites (vendor_id V00 / V02 / V03).
• Participant-disjoint: no individual appears in more than one dyad.
• Every dyad passed automatic quality filters and an LLM prompt-adherence screen.

Files

validation/
├── data/
│   ├── metadata.csv                 one row per sample; the answer key + design factors
│   ├── audio/  sample_000.wav …     16 kHz mono mix (the audio-arm stimulus)
│   ├── video/  sample_000.mp4 …     side-by-side dyad video, H.264 (the video-arm stimulus)
│   └── text/   sample_000.txt …     turn-segmented transcript (the text-arm stimulus)
├── ratings/
│   ├── human_ratings.csv            per-trial crowd ratings (~90 raters per modality; de-identified)
│   └── model_predictions.csv        per-trial zero-shot predictions from 26 models
├── baselines/results.csv            leaderboard: accuracy per system and modality
├── eval.py                          score your own predictions against the answer key
└── make_websafe.py                  regenerate the browser-friendly media (see below)

Repository-level README.md, TERMS.md, and CITATION.cff are shared across all configs. The video stimulus keeps the clip's audio, so the video condition is audiovisual. The audio and text conditions each isolate a single channel.

Model-facing vs. web-safe media

The bundled media are the model-facing stimuli the models were scored on: audio.wav is 16 kHz mono and video.mp4 is the raw side-by-side render. The human rater study used higher-fidelity, browser-friendly derivatives (48 kHz MP3; 720p 8-bit video). Rather than ship a second copy, we include make_websafe.py, which regenerates equivalent derivatives from the released data/video/*.mp4 with the same ffmpeg pipeline and per-speaker normalization:

python validation/make_websafe.py     # writes websafe/audio/*.mp3 and websafe/video/*.mp4

These match the study media in format and processing but are not bit-identical to our original study files (exact bytes depend on your ffmpeg build; adjust --abr if you need a specific bitrate).

🚀 Loading

Pull the manifest and media straight from the Hub:

import pandas as pd
from huggingface_hub import hf_hub_download, snapshot_download

REPO = "fluid-concepts/friend-bench"

# The manifest: 96 rows, sample_id -> label + design factors + media paths
meta_path = hf_hub_download(REPO, repo_type="dataset",
                            filename="validation/data/metadata.csv")
meta = pd.read_csv(meta_path)
meta[["sample_id", "relationship", "temporal_position", "gender_config"]].head()

# Grab one whole modality (e.g. all 96 audio clips) ...
snapshot_download(REPO, repo_type="dataset", allow_patterns=["validation/data/audio/*"])
# ... or everything for one sample across modalities
snapshot_download(REPO, repo_type="dataset", allow_patterns=["validation/data/*/sample_000.*"])

The layout is plain files plus one manifest, so once downloaded no special library is needed — join the manifest to the media by its path columns:

row = meta.iloc[0]
audio_path = "validation/" + row["audio_path"]       # e.g. "validation/audio/sample_000.wav"
video_path = "validation/" + row["video_path"]
transcript = open("validation/" + row["text_path"]).read()
label = row["relationship"]                              # "familiar" or "stranger"

Manifest columns (data/metadata.csv)

Column	Description
sample_id	Stimulus key (sample_NNN).
relationship	Label / answer key: familiar or stranger.
gender_a, gender_b	Estimated perceived gender of each participant (M/F), predicted from voice by a model — not self-reported.
gender_config	mixed or same-gender dyad, derived from the estimated genders above.
temporal_position	early or late — where the clip falls in the interaction.
clip_duration	Clip length in seconds.
n_turns	Number of speaking turns in the clip.
cell	Design cell <vendor>_<fam|str>_<mixed|same> (stratification key; reconstructable from vendor_id + relationship + gender_config).
eo_prompt_a, eo_prompt_b	The Either-Or ice-breaker prompt shown to each participant (elicitation context; does not reveal the label).
vendor_id	Recording site (V00 / V02 / V03).
session_id	Seamless Interaction source session.
file_id_a, file_id_b	Seamless Interaction per-participant file ids.
clip_start, clip_end	Clip offsets (seconds) into the source recording.
audio_path, video_path, text_path	Relative paths to the three modality files.

Gender labels throughout the dataset (manifest and ratings/*.csv) are voice-based model estimates, not self-reported; treat them as a noisy covariate.

Evaluate your model

Write a CSV with two columns, sample_id and prediction (values familiar / stranger, or blank for a declined trial), then run:

python validation/eval.py my_predictions.csv

It reports overall accuracy over answered trials, coverage, and a breakdown by relationship (per-class recall), gender, temporal position, and design cell. Accuracy is scored over answered trials, so a model is not penalized for declining rather than misjudging.

Baselines

Human raters and 26 zero-shot models were evaluated on the same 96 stimuli. Full per-system numbers are in validation/baselines/results.csv. Headline accuracy (%), by modality:

Modality	Best model	Acc.	Human crowd	Acc.	Human (indiv.)	Acc.
Text	gpt_text_mini	56.2	majority vote	51.0	per-rater mean	49.9
Audio	gemini_audio_pro	66.7	majority vote	63.5	per-rater mean	56.3
Video	gemini_video	66.7	majority vote	71.9	per-rater mean	60.9

Chance is 50%. The task is hard: text carries little signal for either humans or models, audio adds reliable signal, and video is humans' strongest modality. The best models match the human crowd on accuracy but reach it with a stronger bias toward answering "stranger," where humans stay balanced. See the paper for the full analysis.

All figures are recomputed from the shipped ratings and match the paper. The crowd is a per-stimulus majority vote; a tie is scored as 0.5 (the expected accuracy under random tie-breaking, the standard unbiased convention).

Human ratings

ratings/human_ratings.csv is one row per rater × stimulus trial (8,941 trials, 280 raters). Each row carries the trial (response, correct, confidence, rt_ms), the stimulus join keys, and the rater's covariates: coarse demographics (age band, self-reported gender, english), the TSIS social-intelligence scale (tsis_total, tsis_mean, tsis1–tsis7), and post-task cue-usage self-reports (visual_*, vocal_*, verbal_*, inter_*, other_environ).

The data are de-identified: raters are a study-local pseudonym (rater_id, R001…), and no platform ids, device/OS/browser strings, or other quasi-identifiers are included. The rater's own gender here is self-reported, distinct from the voice-estimated gender_a/gender_b of the recorded participants.

Notes and caveats

Public reference set — not a blind test set. This config ships the ground-truth relationship label and the full per-trial human and model responses, so you can reproduce the paper and score your own system directly. It is therefore a public reference/validation set, not a contamination-controlled held-out test set. Treat results on it as reproduction, and expect the labels to enter training corpora over time — do not report numbers on it as a leak-proof leaderboard. A separate blind test set (labels and responses withheld) is planned for leaderboard use.

• Source pointers (vendor_id, session_id, file_id_a/b, clip_start, clip_end) are kept in the manifest so you can trace provenance or pull modalities not bundled here (e.g. body pose) from the upstream dataset.
• Held-out labels. Finer-grained relationship type (e.g. friends vs. romantic partner) is not released here; it is reserved for a future relationship-type evaluation.
• Zero-shot evaluation set. There is no train/test split; the whole set is for zero-shot evaluation only.

📄 License

FriendBench is derived from Seamless Interaction and is released under the same license, CC-BY-NC 4.0 (Attribution–NonCommercial 4.0):

• Share — copy and redistribute in any medium or format
• Adapt — remix, transform, and build upon the material

Under these terms:

• Attribution — credit Seamless Interaction and this benchmark, with a link to the license
• NonCommercial — no commercial use without explicit permission

Full terms, including the re-identification and research-use conditions, are in TERMS.md.

📑 Citation

Cite FriendBench and the Seamless Interaction dataset. Machine-readable metadata is in CITATION.cff; a formal citation will be added on publication.

🙏 Acknowledgments

Built on the Seamless Interaction dataset from Meta FAIR.