README.md

---
license: cc-by-nc-4.0
task_categories:
  - audio-classification
language:
  - en
pretty_name: MIA_LALM
---

# MIA_LALM

Audio datasets for **Membership Inference Attacks against Large Audio Language
Models** ([arXiv](https://arxiv.org/abs/2603.28378),
[code](https://github.com/snooow1029/ALM_MIA)).

The audio is distributed as one zstd-compressed tar per dataset family. A few
large archives download far faster than ~90k individual WAV files (no per-file
overhead, no HTTP 429), and each extracts to the exact layout the attack runners
expect.

## Download

With the [code repository](https://github.com/snooow1029/ALM_MIA) cloned,
download each family into ``MIA_on_dataset/data/`` and extract it there — the
loader picks it up automatically, no environment variable needed:

```bash
cd MIA_on_dataset
for ds in voxpopuli spgispeech gigaspeech librispeech tedlium clotho cochlscene nsynth; do
  hf download Snooow1029/MIA_LALM "${ds}_mia_dataset.tar.zst" \
    --repo-type dataset --local-dir data
  tar --zstd -xf "data/${ds}_mia_dataset.tar.zst" -C data
  rm "data/${ds}_mia_dataset.tar.zst"
done
```

Each dataset then lives under ``MIA_on_dataset/data/<family>_mia_dataset/``,
matching the manifest paths the runners expect (e.g.
``voxpopuli_mia_dataset/tts_61_2/tts_dataset.json``). To keep the data
elsewhere, extract anywhere and ``export AUDIO_MIA_DATA_ROOT=/that/path``.

## Layout

Each archive expands to ``<family>_mia_dataset/...`` containing audio plus the
JSON/CSV manifests. The same manifests are also available loose in this repo for
browsing.