Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 333365683 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

WavLate training data

The data used to train and evaluate WavLate, a late-interaction (multi-vector) model for retrieval between audio clips and text captions.

Config Split Rows What a row is
training train 441,198 one training clip: its captions, its 10 verified hard negatives, and where its audio is stored
evaluation audiocaps_test 964 one AudioCaps test clip with its 5 captions and the audio
evaluation clotho_evaluation 1,045 one Clotho evaluation clip with its 5 captions and the audio
from datasets import load_dataset

train = load_dataset("chungimungi/WavLate-training-data", "training", split="train")
test = load_dataset("chungimungi/WavLate-training-data", "evaluation", split="clotho_evaluation")

Training config

Sources (456,546 captions in total, test-set overlaps removed):

Source Clips
WavCaps FreeSound 256,154
WavCaps AudioSet strongly labelled 99,278
AudioCaps train 49,838
WavCaps BBC Sound Effects 31,201
Clotho development 3,837
WavCaps SoundBible 1,232

342 clips of the 441,540 in the source manifest had no usable audio and are left out.

Column Type Meaning
id string clip identifier used throughout training
source_dataset, subset, source_id string where the clip comes from (wavcaps / audiocaps / clotho, the WavCaps subset, and the id in that source)
duration_s float clip length in seconds (training used the first 30 s)
captions list of strings the clip's reference captions
source_license string the clip's licence when the source states one (Freesound clips in Clotho); empty otherwise
hard_negative_ids list of strings 10 other clips that a pooled model scores as close to this clip, and that a cross-encoder judged to describe different audio
hard_negative_captions list of strings the first caption of each hard negative, for reading
hard_negative_teacher_scores list of floats cosine similarity between this clip's audio and the negative's caption under frozen pooled MGA-CLAP
hard_negative_reranker_scores list of floats the highest cross-encoder score between the negative's captions and this clip's captions
positive_teacher_score float pooled MGA-CLAP cosine similarity between this clip's audio and its best own caption
audio_repo, audio_shard, audio_member string the dataset repo, tar shard and file name holding the clip's audio (32 kHz FLAC)
audio_offset, audio_size, audio_sha256 int, int, string byte offset and size of the FLAC file inside the tar, and its SHA-256
audio_sample_rate int 32000

How the hard negatives were mined

  1. Every clip and caption was embedded with pooled MGA-CLAP (one vector each).
  2. For each clip, all training captions were ranked by their similarity to the clip's audio. A candidate must score at least 0.01 below the clip's own best caption, and the clip itself, clips cut from the same YouTube video and identical captions are skipped.
  3. Each candidate was checked with the cross-encoder cross-encoder/ettin-reranker-32m-v1: it scores every pair of (candidate caption, own caption). A candidate is rejected when its best score reaches 9.94, the 0.9 quantile of the scores between two captions of the same clip (measured on multi-caption training clips). In plain terms, a candidate that reads like another description of the same clip is not used as a negative.
  4. The walk continues until 10 candidates pass. 6.5M candidates were examined, 32.2% were rejected, and 15 clips ended with fewer than 10.

Evaluation config

The exact test pools used for every number in the WavLate model card: AudioCaps test clips that are still downloadable (964 of 975) and the full Clotho evaluation split (1,045). The audio column holds the original audio as FLAC (AudioCaps at 32 kHz, Clotho at 44.1 kHz). Retrieval is scored with all 5 captions of a clip as correct answers.

Licences

This repository collects data from several sources, each under its own terms:

  • WavCaps captions and audio: academic use only, see the WavCaps licence and the terms of FreeSound, BBC Sound Effects and SoundBible.
  • AudioCaps: captions under the AudioCaps licence; audio from YouTube via AudioSet.
  • Clotho: audio from Freesound under per-clip Creative Commons licences; captions under the Clotho licence.

Use this data for non-commercial research and check each source's terms before redistribution.

Downloads last month
-

Models trained or fine-tuned on chungimungi/WavLate-training-data

Article mentioning chungimungi/WavLate-training-data