cucl2/AnyAudio-Judge-7B

1---2license: apache-2.03language:4  - en5  - zh6library_name: transformers7base_model: Qwen/Qwen2.5-Omni-7B8tags:9  - audio10  - audio-language-model11  - instruction-following12  - rubric-based-evaluation13  - judge-model14pipeline_tag: audio-text-to-text15---16 17# AnyAudio-Judge-7B18 19`AnyAudio-Judge-7B` is a **dynamic rubric-based audio judge** built on top of [Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B). It predicts, for each yes/no rubric item describing one verifiable aspect of an audio caption, whether the audio satisfies that aspect — together with a short evidence string.20 21This is the smaller variant of the AnyAudio-Judge family. The larger `AnyAudio-Judge-30B` (initialized from Qwen3-Omni-30B-A3B-Captioner) is the variant reported in the paper. The 7B model is trained on the same SFT corpus and is intended for users who need a more efficient evaluator.22 23> Companion benchmark: [`cucl2/AnyAudio-Judge-Bench`](https://huggingface.co/datasets/cucl2/AnyAudio-Judge-Bench)24> Companion corpus: [`cucl2/AnyAudio-Judge-Corpus`](https://huggingface.co/datasets/cucl2/AnyAudio-Judge-Corpus)25> Companion 30B model: [`cucl2/AnyAudio-Judge-30B`](https://huggingface.co/cucl2/AnyAudio-Judge-30B)26 27## Training28 29- **Base**: Qwen2.5-Omni-7B30- **Corpus**: 105K (audio, instruction, rubric, CoT) tuples (see `cucl2/AnyAudio-Judge-Corpus`)31- **Stage**: full-parameter SFT for 1 epoch32  - 16 × H20 96GB33  - per-device batch size 4, grad accumulation 134  - learning rate 1e-535 36## Usage37 38```python39from anyaudio_judge import AnyAudioJudge, decompose_instruction40 41caption = "A gentle, delicate female voice, with soft and smooth pitch, calm and restrained throughout."42rubric  = decompose_instruction(caption)  # external LLM call43 44judge = AnyAudioJudge.from_pretrained("cucl2/AnyAudio-Judge-7B")45result = judge.judge("./demo.wav", rubric)46print("alignment_score:", result.score)47for item in result.items:48    print(item.question, "->", item.answer)49```50 51(See the [GitHub repo](https://github.com/) for the full pipeline including external rubric decomposition.)52 53## License54 55Apache-2.0, inheriting the license of the base Qwen2.5-Omni-7B model.56 57## Citation58 59```bibtex60@misc{anyaudiojudge2026,61  title  = {AnyAudio-Judge: A Dynamic Rubric-Based Benchmark and Evaluator for Audio Instruction Following},62  author = {Anonymous Authors},63  year   = {2026},64  note   = {Preprint, under submission}65}66```67