Qwen3-4B Soundness Judge SFT
This checkpoint is the trained soundness reward model used by the review-point
RL experiments. Given a paper and one review point, it predicts the factual
status of the review point's material premises:
0: at least one material premise is clearly false.
1: no material premise is clearly false, but at least one remains unresolved.
2: every material premise is true.
The RL judge service evaluates one generated review point at a time and uses
the logits for the three score tokens after the JSON prefix {"score":.
This is a research reward model trained for paper-review feedback. Its scores
should not be treated as verified factual or scientific judgments.