--- license: other license_name: mixed-upstream language: - zh - en task_categories: - audio-classification - audio-text-to-text tags: - audio - llm-as-a-judge - instruction-following - rubric - text-to-audio - text-to-speech - text-to-music pretty_name: AnyAudio-Judge Corpus (opus) size_categories: - 10K 中文简介见文末。 ## Why a new corpus Manual inspection of the original 105K corpus (rubrics and negatives produced by Qwen3-30B-A3B) found: | Problem | Example | |---|---| | Rubrics too fine-grained | "is the pitch low" and "is the pitch stable" as two items; ~37 items per `mix` sample | | Hallucinated rubrics | items / evidence mention details absent from the caption ("sustained", "no decay") | | Over-rewritten negatives | a negative claimed to change only the pitch but rewrote gender, speed and emotion — every item answered *no* | | Template negatives | sound negatives mostly appended "with a steady rhythmic pulse" / "with a long reverberant tail" | | Wrong labels | items about unchanged parts of the caption labelled *no* | Design goals of this corpus: - **Rubrics**: faithful to the caption (no hallucination), moderate granularity — modifiers of the same dimension are merged into one item. - **Fine-grained negatives (`neg_change`)**: only **1–2 edits** to the original caption, **at most 2 items answered *no***, all other items *yes*. Edits are diverse and plausible, simulating a generator that is "almost right" (one level off in degree, missing dynamics, wrong local/temporal detail, near-synonym substitution, extra detail …). Crude reversals (male↔female, fast↔slow) are capped at 10 %. - **Swap negatives (`neg_swap`)**: the instruction is another clip's caption; items are checked against the true caption, so large-scale *no* (with natural, coincidental *yes*). ## Contents ``` data/{speech,sound,music,mix}.jsonl structured samples (29,791) sft/train.jsonl ms-swift training file (34,291 rows) — exactly what our 7B judge was trained on audio/sound/*.wav 4,200 AudioCaps clips audio/music/*.wav 4,100 MusicBench clips docs/ construction specs (Chinese) given to the sub-agents tools/validate.py the validator every shard had to pass (0 ERROR) viewer/ local browsing / QC web page ``` **Audio availability**: `sound` and `music` audio is included. `speech` and `mix` audio comes from in-house data and is **not released**; those rows keep the `uuid`, captions, rubrics and answers, with `audio = null`. ### Statistics (structured data) | Subset | Audio source | pos | neg_change | neg_swap | Total | Avg. items | *no* items | |---|---|---:|---:|---:|---:|---:|---:| | speech | in-house expressive speech (not released) | 3,150 | 4,500 | 1,350 | 9,000 | 4.6 | 24 % | | sound | AudioCaps | 2,940 | 4,200 | 1,260 | 8,400 | 2.1 | 43 % | | music | MusicBench | 2,870 | 4,100 | 1,230 | 8,200 | 4.4 | 27 % | | mix | in-house cinematic mixed audio (not released) | 1,467 | 2,096 | 628 | 4,191 | 10.2 | 18 % | | **Total** | | **10,427** | **14,896** | **4,468** | **29,791** | 4.6 | 25 % | - pos : neg_change : neg_swap ≈ 35 : 50 : 15. Each audio clip yields one `pos` and one or two `neg_change` (two negatives edit different dimensions). - `neg_change`: 72 % have 1 edit, 28 % have 2 edits; 72 % have exactly one *no* item, 28 % two, none more. - Edit types (letters are defined per subset in `docs/SPEC_*.md`): A 23.3 %, B 21.3 %, C 15.8 %, D 16.9 %, E 17.2 %, F 5.4 %. - Built in two rounds with the same method: `v0.1` (10,000 samples) and `v0.3` (19,791 new samples; audio disjoint from v0.1). The `version` field records which round a sample came from. ### Structured record format (`data/*.jsonl`) ```json { "id": "v0.3-sound-000001", "version": "v0.3", "subset": "sound", "type": "neg_change", // pos | neg_change | neg_swap "label": "no", // yes iff every item is yes "uuid": "60940", // AudioCaps audiocap_id / MusicBench file id / in-house id "audio": "audio/sound/60940.wav", // null for speech / mix "source_dataset": "AudioCaps", "caption_ori": "A steam engine running on railroad tracks as steam hisses followed by compressed air releasing", "caption": "A diesel locomotive running on railroad tracks as steam hisses followed by compressed air releasing", "modifications": [{"dimension": "声源", "type": "A", "original": "A steam engine running on railroad tracks", "modified": "A diesel locomotive running on railroad tracks"}], "questions": [ {"dimension": "声音事件", "question": "是否有内燃机车在铁轨上运行的声音?", "basis": "A diesel locomotive running on railroad tracks"}, {"dimension": "时序", "question": "是否先有蒸汽嘶嘶声,随后出现压缩空气释放的声音?", "basis": "steam hisses followed by compressed air releasing"}], "answers": [ {"id": 0, "answer": "no", "evidence": "听到的是蒸汽机车在铁轨上运行,不是内燃机车。"}, {"id": 1, "answer": "yes", "evidence": "先是蒸汽嘶嘶作响,接着出现压缩空气释放的声音。"}] } ``` `caption_ori` is the true description of the audio; `caption` is the instruction being judged. `neg_swap` rows also carry `swap_uuid` (whose caption was borrowed). `basis` is always a verbatim substring of `caption`. Rubric questions / evidence are in Chinese; `basis` follows the caption language. ### SFT format (`sft/train.jsonl`) ms-swift `messages` + `audios` format. It contains the 29,791 multi-item samples above (re-wrapped with diverse system prompts, question phrasings, numbering styles and output formats — 35 % keep the original JSON format) plus 4,500 single-item yes/no samples. Rubrics, answers and evidence are never changed by the re-wrapping. Audio paths are relative to the repository root; rows with `"audio_released": false` (speech / mix) point to files that are not shipped. ## Construction 1. A main agent wrote the per-subset specs (`docs/SPEC_*.md`) and sampled source clips into shards. 2. Parallel Claude sub-agents processed one shard at a time (`docs/AGENT_PROMPT_v0.3.md`), **writing every caption, rubric item, answer and evidence by hand — scripts were only allowed for JSON plumbing and substring assertions**. 3. Every shard had to pass `tools/validate.py` with 0 ERROR (structure, `basis` ⊂ caption, no meta-language in captions, `pos` all *yes*, 1–2 edits and 1–2 *no* items per `neg_change`, complete pos / neg / swap counts per clip; item-count range and same-dimension double edits are warnings). Edit-count / edit-type proportions were enforced through the sub-agent instructions. Rows containing raw English speaker tags (`Speaker_A` …) were dropped. Item-count ranges: speech 2–8, sound 1–4, music 2–7, mix 5–14. For music, key / chord / time signature / usage speculation are not turned into items, and BPM is only mapped to fast / slow. **Known limitation**: the writers could not listen to the audio. Ground truth comes only from the original human / model caption (plus attribute metadata for speech). Errors already present in the source captions are not corrected. ## Viewer Online (read-only): [cucl2/AnyAudio-Judge-Corpus-opus-viewer](https://huggingface.co/spaces/cucl2/AnyAudio-Judge-Corpus-opus-viewer). Browse locally (all 29,791 samples, filter by subset / type / label / heuristic flags, play sound & music audio): ```bash git lfs install && git clone https://huggingface.co/datasets/cucl2/AnyAudio-Judge-Corpus-opus cd AnyAudio-Judge-Corpus-opus python3 viewer/server.py --port 8080 # add --readonly to disable annotation # open http://localhost:8080 ``` Without `--readonly`, you can annotate samples in the page; annotations are saved to `viewer/viewer.annotations.json`. ## License Annotations (captions edits, rubrics, answers, evidence, specs, code) are released under Apache-2.0. Audio and original captions remain under their upstream licenses — AudioCaps (AudioSet / YouTube) and MusicBench — please consult them before any commercial use. ## Citation ```bibtex @article{li2026anyaudio, title={AnyAudio-Judge: A Dynamic Rubric-Based Benchmark and Evaluator for Audio Instruction Following}, author={Li, Haitao and Tan, Tian and Yang, Yuguang and Yang, Shan and Chen, Xie}, journal={arXiv preprint arXiv:2606.03116}, year={2026} } ``` --- ## 中文简介 本数据集是 AnyAudio-Judge 训练语料的重制版:由 Claude 子代理按 `docs/` 中的构造规范**逐条手写**拆题、正样本答案、精细改动负样本和互换负样本,替代原先由 Qwen3-30B 批量生成的数据,解决拆题过细、幻觉、负样本整段重写、模板化改动、未改动处被标 no 等问题。 - 结构化数据 29,791 条(speech 9,000 / sound 8,400 / music 8,200 / mix 4,191),pos : 精细改动 : 互换 ≈ 35 : 50 : 15;精细改动负样本只改 1~2 处,答 no 的题最多 2 道。 - `sft/train.jsonl` 为训练文件(34,291 行,含问法 / 格式多样化与 4,500 条单题样本)。 - sound(AudioCaps)与 music(MusicBench)附音频;speech 与 mix 音频来自内部数据,**不公开**,只保留 uuid 与文本标注。 - `viewer/` 为本地浏览 / 质检网页:`python3 viewer/server.py --port 8080`。 - 局限:撰写者听不到音频,真实情况只来自原 caption,原 caption 本身的错误无法发现。