FirstPass: Grounding AI Scientific Judgment in Multi-Round Editorial Outcomes

FirstPass is a fine-tuned model based on Qwen2.5-7B-Instruct, trained on a dataset of 3,668 multi-round peer-review dialogues from various scientific domains. It utilizes Low-Rank Adaptation and introduces response-only loss masking, achieving an accuracy of 80.5% in predicting editorial outcomes, significantly outperforming existing models like Gemini-3.1-flash-lite-preview. This tool enhances the peer review process by providing anticipatory critiques and revision predictions, thereby offering authors a reliable assessment akin to that of a trusted colleague across multiple disciplines.