| --- |
| base_model: Qwen/Qwen2.5-Omni-7B |
| language: |
| - en |
| license: other |
| license_name: apache-2.0 |
| license_link: https://huggingface.co/Qwen/Qwen2.5-Omni-7B/blob/main/LICENSE |
| pipeline_tag: audio-text-to-text |
| tags: |
| - audio |
| - audio-reasoning |
| - reinforcement-learning |
| --- |
| |
| # AudioRubrics |
|
|
| The model from [Reinforcement Learning with Evolving Rubrics as Rewards for Audio Reasoning](https://huggingface.co/papers/2608.02831): Qwen2.5-Omni-7B post-trained with GRPO using self-evolving, audio-grounded rubric rewards and an overthinking penalty. |
|
|
| - Project page: https://audiorubrics.github.io |
| - Code: https://github.com/tianyi-lab/AudioRubrics |
| - Rubric dataset: https://huggingface.co/datasets/umd-zhou-lab/AVQA-Audio-Rubrics |
|
|
| This is the full merged checkpoint (thinker merged back into the complete Omni model) and can be served directly with vLLM: |
|
|
| ```bash |
| vllm serve umd-zhou-lab/AudioRubrics --served-model-name omni --trust-remote-code \ |
| --max-model-len 8192 --limit-mm-per-prompt '{"audio":1}' |
| ``` |
|
|
| See the [GitHub repository](https://github.com/Yu-Fangxu/AudioRubrics) for training and evaluation instructions. |
