Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
FRIENDSHIP-BENCH contains recordings of real human conversations. This dataset is intended for model benchmarking under CC BY-ND 4.0. By requesting access, you agree to comply with the license terms and not attempt to identify any speakers.
Log in or Sign Up to review the conditions and access this dataset content.
FRIENDSHIP-BENCH (data)
Data for FRIENDSHIP-BENCH, a benchmark of how natural a real-time voice model's reaction is at the social moments that matter between friends and partners. Code, protocol and scoring: https://github.com/symbiateam/liva-friendship-bench
- 1,081 items in seven moment types: affection, teasing, direct question, flirting, vulnerable disclosure, self-deprecation, joke / punchline.
- Public split (806 items): audio, speaker transcript and the real listener's reply.
- Private split (275 items): audio and speaker transcript. The real listeners' replies are not released.
Files
| Path | What |
|---|---|
audio/<id>/input.wav |
the speaker's own track. from 2 s before the turn to its end then 6 s of silence |
items.jsonl |
one line per item (fields below) |
| Field | Meaning |
|---|---|
id |
item id (folder name under audio/) |
split |
public or private |
conversation |
conversation id; all items of a conversation are in the same split |
moment |
affection, tease, question, flirt, vulnerable, self_deprecation, punchline |
speaker_text |
transcript of the speaker's turn |
turn_end_s |
end of the speaker's turn in input.wav (seconds). the reply window is 0.3 s before to 6 s after |
real_listener_text |
what the real listener said (public split only) |
Download
huggingface-cli download liva-ai/friendship-bench-data --repo-type dataset --local-dir data
Licence
CC BY-ND 4.0
- Downloads last month
- 80