We present SciRM and SciRM-Ref, cost-efficient open-source reward models tailored for scientific writing evaluation. Our models are trained via a two-stage reinforcement learning framework (using GRPO) that first optimizes scientific evaluation preferences, then refines reasoning capabilities. The multi-aspect evaluation design and joint training across diverse tasks enable fine-grained assessment and robustness to dynamic criteria and scoring rubrics. Key properties: — ✅ Multi-aspect evaluation — assesses multiple evaluation dimensions per task — ✅ Dynamic scoring rubrics — conditioned on an explicit evaluation constitution at both train and inference time — ✅ Cross-task generalization — a single model handles diverse and previously unseen scientific writing tasks without task-specific retraining — ✅ Open-source and cost-efficient — no reliance on proprietary LLMs Although SciRM and SciRM-Ref have increased reasoning capab Therefore, adherence to the system prompt and providing a clear evaluation criteria is strongly recommended for best performance. SciRM is more stable than SciRM-Ref in terms of output formatting. This is an…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: UKPLab
Теги: qwen2, reward-model, scientific-writing, evaluation, reinforcement-learning, grpo, conversational, en
Лайков: 3 | Загрузок: 135
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
