vnahata/avcaps-retrieval

AVCaps audio–visual retrieval (MTEB)

Retrieval tasks over AVCaps, an audio-visual dataset derived from VidOR in which each clip is captioned three separate ways — from the audio alone, from the visuals alone, and from both together.

That separation is the point: the audio-only, video-only and combined directions can be scored independently on identical clips, rather than inferred from a single caption set that mixes the modalities.

Prepared for MTEB as six tasks: AVCapsA2TRetrieval / AVCapsT2ARetrieval, AVCapsV2TRetrieval / AVCapsT2VRetrieval, AVCapsVA2TRetrieval / AVCapsT2VARetrieval.

Contents

config	rows	description
media	200	clips with separate video and audio columns
audio_captions	725	captions written from audio only
visual_captions	788	captions written from visuals only
av_captions	808	captions written from both

Caption configs carry id, text, media_id; media_id joins to media.

How this was built

Official test split only — train and validation are untouched, so the evaluation set matches the split the authors held out.

The published clips are muxed mp4s. Audio is demuxed to a separate mono 16 kHz stream, and each task exposes only the media columns its captions actually describe — so an audio-caption task cannot be quietly solved off the video track.

Construction script: scripts/data/avcaps_retrieval/create_data.py in the MTEB repo.

Licence

AVCaps is released under CC-BY-NC-4.0, so this derived dataset carries the same non-commercial restriction.

Citation

@article{sudarsanam2025avcaps,
  author = {Sudarsanam, Parthasaarathy and Martin-Morato, Irene and Hakala, Aapo and Virtanen, Tuomas},
  journal = {IEEE Open Journal of Signal Processing},
  pages = {691--704},
  title = {{AVCaps}: An Audio-Visual Dataset with Modality-Specific Captions},
  volume = {6},
  year = {2025},
}