Evidencefirst Audio: Captions and QA Pairs for AudioSet Clips

EvA Open Data provides audio clips paired with descriptive captions and instruction-based question-answer data. The audio is sourced from the AudioSet Strong Labels dataset and stored in parquet shards. The dataset was authored by SatsukiVie and last updated on Hugging Face in May 2026.

Use Cases

• Train audio captioning models based on the provided caption and caption_with_asr fields.
• Develop audio question-answering systems using the instruction QA data.
• Fine-tune multimodal language models on audio-text pairs sourced from AudioSet.
• Benchmark audio understanding models on tasks combining descriptive and instructional data.

Strengths

• Includes multiple data modalities: audio, descriptive captions, and instruction-based QA pairs.
• Audio clips are sourced from the established AudioSet Strong Labels dataset.
• Data is structured with specific fields like id, audio_path, caption, and caption_with_asr.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license information are unknown.

Provenance

Source

Audio clips sourced from AudioSet Strong Labels.

Collection Method

Likely processed and paired with generated captions and QA data.

Freshness

Last updated 2026-05-28 09:48:32; freshness should be verified.