hyun1905/qwen3-4b-instruct-2507-soundness-judge-sft
The hyun1905/qwen3-4b-instruct-2507-soundness-judge-sft is a 4 billion parameter Qwen3-based instruction-tuned model designed as a research reward model. It predicts the factual soundness of review points for scientific papers, classifying them into three categories based on the truthfulness of their material premises. This model is specifically trained to evaluate generated review points in the context of paper-review feedback.
Loading preview...
Model Overview
The hyun1905/qwen3-4b-instruct-2507-soundness-judge-sft is a 4 billion parameter Qwen3-based instruction-tuned model developed as a soundness reward model for research purposes. Its primary function is to assess the factual status of review points provided for scientific papers.
Key Capabilities
This model evaluates a given paper and a single review point, predicting its factual soundness based on the material premises. It assigns one of three scores:
0: Indicates that at least one material premise in the review point is clearly false.1: Suggests that no material premise is clearly false, but some remain unresolved.2: Signifies that every material premise in the review point is true.
Intended Use and Limitations
This model is specifically trained for paper-review feedback within an RL judge service, where it processes one generated review point at a time. It uses logits for the score tokens after a JSON prefix {"score":. It is important to note that this is a research reward model, and its scores should not be interpreted as verified factual or scientific judgments.