---
license: apache-2.0
language:
- en
task_categories:
- audio-to-audio
tags:
- mteb
- audio-retrieval
- composed-retrieval
configs:
- config_name: queries
  data_files:
  - split: test
    path: queries/test-*
- config_name: corpus
  data_files:
  - split: test
    path: corpus/test-*
- config_name: qrels
  data_files:
  - split: test
    path: qrels/test-*
---

# MMEdit-AT2A

MMEdit-AT2A is an MTEB-formatted composed audio retrieval adaptation of
[`CocoBro/MMEdit-TestSet`](https://huggingface.co/datasets/CocoBro/MMEdit-TestSet).
Each query combines an unedited source recording with a natural-language edit
instruction, and the corpus contains the corresponding edited recordings.

## Schema

- `queries`: `id`, `audio` (unedited source), and `text` (edit instruction)
- `corpus`: `id` and `audio` (edited target)
- `qrels`: `query-id`, `corpus-id`, and binary `score`

The test split contains 3,317 queries and 3,317 corpus items. Exact
byte-identical target duplicates are all marked relevant, producing 3,337 qrels.

## Construction

The source is pinned to commit
`ae4f9a772180a2a3c77c2e865b398e7d6f60bcee`. Triplets are joined by `audio_id`.
The construction preserves the original WAV payloads without padding,
resampling, or loudness normalization. IDs are deterministic:
`q-{audio_id}` for queries and `t-{audio_id}` for corpus items.

## Known source-format discrepancies

Although the source card describes all clips as 24 kHz, mono, and ten seconds:

- 1,394 source files and 1,394 target files differ from ten seconds by more
  than 0.05 seconds (minimum observed duration: 2.3985 seconds).
- 143 targets (`replace_one` and `replace_time`) are 16 kHz, 32-bit PCM rather
  than 24 kHz, 16-bit PCM.
- All 6,634 WAV files are mono.

These properties are retained so this derivative stays faithful to the source.

## Benchmark limitation: source-only leakage

The native global candidate pool has a substantial shortcut: a model can often
identify the edited target from source audio alone without using the instruction.
A deterministic, non-neural 64-bin log-mel fingerprint baseline obtained the
following full-pool results:

| Diagnostic | Recall@1 | mAP@10 | nDCG@10 |
|---|---:|---:|---:|
| Random expectation | 0.0003 | 0.0009 | 0.0014 |
| Source-only spectral fingerprint | 0.5849 | 0.6094 | 0.6216 |
| Source duration + spectral fingerprint | 0.8113 | 0.8376 | 0.8495 |

Consequently, scores on this benchmark must not be interpreted as measuring
instruction use alone. Results should be accompanied by an audio-only ablation
where possible. The native 3,317-pair design is intentionally preserved rather
than introducing synthetic edits or reviewer-unapproved hard negatives.

## Source reuse and duplicate handling

The audit found five byte-identical source groups, four of which have multiple
distinct edited targets. It also found ten byte-identical target groups. The
corpus retains every native item, while qrels mark every byte-identical copy of
a query's target as relevant.

## License and citation

The source dataset is published under Apache-2.0. Please also cite the original
MMEdit work:

```bibtex
@article{tao2025mmedit,
  author = {Tao, Ye and Xu, Xuenan and Wu, Wen and Wang, Shuai and Wu, Mengyue and Zhang, Chao},
  journal = {arXiv preprint arXiv:2512.20339},
  title = {MMEDIT: A Unified Framework for Multi-Type Audio Editing via Audio Language Model},
  url = {https://arxiv.org/abs/2512.20339},
  year = {2025},
}
```