SWE-Critix: Post-training Generalist Verifiers for Assessing Coding Agent Trajectories

SWE-Critix presents an LLM post-training pipeline for evaluating coding-agent trajectories. The trained verifier takes a trajectory as input and determines whether the trajectory successfully completes the task. Before making the binary judgment, the verifier also produces a chain-of-thought (CoT) rationale that justifies its decision.

SWE-Critix Post-training Pipeline

We release the model weights, datasets, and training scripts

• SFT and RL checkpoints
• SFT, RL, and Test datasets
• SFT and RL training scripts (tested on Ascend 910 NPUs)

Visit https://github.com/SWE-Critix/SWE-Critix to learn more.