vhands/audio-reasoning-qa-post-public

audio-reasoning-qa-post-public

Question-answering and multi-task audio reasoning annotations across 15 public audio QA datasets. Spans general audio QA (Clotho-AQA, HeySQuAD), music reasoning (MU-LLaMA, MusicBench, LLARK-MTAT, Music-AVQA), speech-grounded QA (LibriSQA, GigaSpeech), and NVIDIA-aggregator skill subsets (TemporalQA, CountingQA, AudioSet-Speech-QA, GigaSpeech-Long-QA). Closes a substantial slice of the public audio-reasoning SFT gap (compare to NVIDIA AudioSkills-XL coverage).

Audio is not bundled in this repo. See download.sh and per-dataset data/<name>.info.json for the fetch recipe; run postlink_audio.py after fetching to rewrite the JSONL audio_path fields with absolute local paths.

Total: 750,141 rows across 15 datasets.

Datasets

Schema

Each JSONL record:

{
  "dataset": "<DatasetName>",
  "split": "train",
  "audio_path": "<basename or relative path>",
  "instruction": "<question or task prompt>",
  "response": "<answer / label / caption>",
  "meta": { "...": "..." }
}

After postlink_audio.py, a sibling data/<name>.with_audio.jsonl is produced with audio_path rewritten to absolute local paths.

How to use

git clone https://huggingface.co/datasets/vhands/audio-reasoning-qa-post-public
cd audio-reasoning-qa-post-public
bash download.sh                    # print per-dataset audio recipes
bash download.sh <DatasetName>      # print recipe for one dataset
# (manually fetch audio into audio/<DatasetName>/ per the recipe)
python3 postlink_audio.py           # rewrite JSONLs with absolute audio paths

Licenses

The bundle is offered under cc-by-nc-sa-4.0 — the most restrictive license among contained datasets. Per-dataset licenses are in the table above and in data/<name>.info.json. Attribution to upstream creators is required when redistributing.

Files

audio-reasoning-qa-post-public/
├── README.md
├── dataset_manifest.json
├── domain_reasoning_qa.jsonl           # joint, 750,141 rows
├── data/
│   ├── <DatasetName>.jsonl         (×15)
│   └── <DatasetName>.info.json     (×15)
├── audio/                          # populated by you
├── download.sh
└── postlink_audio.py

Companion bundles

• —vhands/audio-emotion-post-public — 5 emotion datasets, 45k rows
• —vhands/audio-reasoning-qa-post-public — 15 reasoning-QA datasets, 750k rows
• —vhands/audio-music-mir-post-public — 7 music/MIR datasets, 137k rows
• —vhands/audio-event-classification-post-public — 7 sound-event/scene datasets, 95k rows

Together these cover ~1 M rows of audio-LLM post-training annotations across 5 domains.