code-critic-model/Qwen3-4B-Critic-SFT-Qwen-only
Qwen3-4B-Critic-SFT-Qwen-only is a 4 billion parameter critic model developed by Shubham Gandhi, Yiqing Xie, Atharva Naik, Ruichen Zhu, and Carolyn Rose, based on the Qwen3-4B-Instruct-2507 architecture. This model is specifically trained to provide structured critiques for coding agent trajectories, identifying error categories, evidence, and recovery actions. It excels at evaluating the performance of large code agents, particularly those based on Qwen3-Next-80B-A3B, and is optimized for code agent steering rather than direct code generation.
Loading preview...
Model Overview
code-critic-model/Qwen3-4B-Critic-SFT-Qwen-only is a 4 billion parameter critic model, part of the research presented in the paper "Steer, Don't Solve: Training Small Critic Models for Large Code Agents." Developed by Shubham Gandhi et al., this model is specifically fine-tuned to critique the trajectories of coding agents, particularly those generated by Qwen3-Next-80B-A3B-Instruct.
Key Capabilities
- Structured Critiques: The model reads a coding agent's trajectory and outputs a structured critique, including detected error categories, supporting evidence, a suggested recovery action, task status, and overall guidance.
- Agent Steering: Unlike traditional code generation models, this critic model focuses on evaluating and guiding large code agents rather than directly writing code patches.
- Specialized Training Data: It was trained exclusively on critiques of Qwen3-Next-80B-A3B trajectories, using 1,915 examples from R2E-Gym instances (matplotlib, moto, sympy) and Claude Opus 4.6 as the teacher model.
- Performance Improvement: When integrated with coding agents, this critic model demonstrates an improvement in resolve rates on benchmarks like SWE-bench Verified. For instance, it boosted Qwen3-Next-80B-A3B's resolve rate from 20.0% to 24.6%.
Training Details
The model was trained using full-parameter SFT with LLaMA-Factory, based on Qwen/Qwen3-4B-Instruct-2507. It utilizes a sequence length of 32,768 tokens and focuses loss calculation on the final critique turn. This specific version represents an ablation study focusing solely on Qwen-only training data, as detailed in Table 3 of the associated research paper.
Good For
- Enhancing Code Agent Performance: Ideal for developers looking to improve the reliability and effectiveness of large language model-based coding agents by providing targeted, actionable feedback.
- Research in Agentic AI: Useful for researchers exploring critic-based steering mechanisms for complex AI systems, particularly in code generation and debugging domains.
- Evaluating Code Trajectories: Provides a structured method for analyzing and understanding the failure modes and successes of automated coding processes.