| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3-8B |
| tags: |
| - rubric-generation |
| - sft |
| - grubric |
| --- |
| |
| # SFTQwen3-8B-OpenRubrics-v1 |
|
|
| Qwen3-8B full fine-tuned on [OpenRubrics v1](https://huggingface.co/datasets/maxreciprocate/OpenRubrics) (~35.4k examples) for evaluation rubric generation. |
|
|
| ## Training |
|
|
| - **Base model:** Qwen/Qwen3-8B |
| - **Dataset:** OpenRubrics v1 (35,406 examples) |
| - **Epochs:** 1 |
| - **Learning rate:** 8e-6 (cosine schedule) |
| - **Effective batch size:** 128 (per-device=2, gradient accumulation=8, 8 GPUs) |
| - **Max sequence length:** 3072 |
|
|
| ## Task |
|
|
| Given a user prompt, the model generates a structured evaluation rubric in `[Hard Rule]` / `[Principle]` format. These rubrics are used to judge LLM response quality. |
|
|
| ## Evaluation |
|
|
| - ~83.5% format validity on Chatbot Arena prompts |
| - Used as the baseline rubric generator in the GRUBRIC pipeline |
|
|
