code-critic-model/Qwen3-8B-Critic-SFT-CWM-only
The code-critic-model/Qwen3-8B-Critic-SFT-CWM-only is an 8 billion parameter critic model developed by code-critic-model, based on the Qwen3-8B architecture with a 32,768 token context length. It is specifically fine-tuned to provide structured critiques of coding agent trajectories, trained exclusively on critiques of CWM-32B agent outputs. This model is designed to assist larger code agents by identifying error categories, providing evidence, and suggesting recovery actions, rather than generating code itself.
Loading preview...
Overview
This model, Qwen3-8B-Critic-SFT-CWM-only, is an 8 billion parameter critic model derived from the Qwen3-8B base, developed by code-critic-model. It is a specialized component from the research detailed in the paper "Steer, Don't Solve: Training Small Critic Models for Large Code Agents" (arXiv:2606.21811). Its primary function is to analyze a coding agent's trajectory and provide structured feedback, including detected error categories, supporting evidence, a recovery action, task status, and overall guidance.
Key Capabilities
- Structured Code Critique: Provides detailed feedback on coding agent trajectories, identifying issues without writing code patches.
- Trajectory Analysis: Reads and evaluates an agent's progress at specified intervals (e.g., every
ksteps). - Targeted Training: Uniquely trained on a dataset of 4,532 examples derived solely from critiques of CWM-32B agent trajectories on R2E-Gym tasks.
- Context Length: Supports a sequence length of 32,768 tokens, allowing for analysis of extensive code agent interactions.
Performance Insights
As an ablation study component, this critic model, despite being trained only on CWM trajectories, demonstrates a positive impact on the resolve rate of other coding agents like Qwen3-Next-80B-A3B and Qwen3-32B on SWE-bench Verified benchmarks. For instance, Qwen3-Next-80B-A3B's resolve rate increased from 20.0% to 23.6% with this critic. This highlights its ability to transfer critical insights to agents beyond its specific training data source.
Use Cases
This model is ideal for integration into larger code agent systems where an external, specialized critic is needed to provide actionable feedback and steer the agent's problem-solving process, enhancing overall code generation and debugging efficiency.