code-critic-model/Qwen3-4B-Critic-SFT

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 11, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The code-critic-model/Qwen3-4B-Critic-SFT is a 4 billion parameter Qwen3-based critic model, fine-tuned by code-critic-model with a 32,768 token context length. It is designed to steer large code agents by providing structured critiques on code trajectories, identifying error categories, evidence, and recovery actions. This model excels at improving the resolve rate of various coding agents on benchmarks like SWE-bench Verified by offering guidance rather than directly solving problems. It serves as the starting point for the DPO critic Qwen3-4B-Critic-SFT-DPO.

Loading preview...

Model Overview

code-critic-model/Qwen3-4B-Critic-SFT is a 4 billion parameter critic model based on the Qwen3-4B-Instruct-2507 architecture, developed by code-critic-model. It is specifically fine-tuned to act as a "critic" for large coding agents, providing structured feedback rather than generating code solutions directly. The model operates by analyzing agent trajectories and outputting critiques that include detected error categories, supporting evidence, suggested recovery actions, and overall guidance.

Key Capabilities & Features

  • Critic Functionality: Designed to sit alongside a frozen coding agent, reading its trajectory and providing short, structured critiques every k steps.
  • Structured Output: Critiques include error categories, evidence, recovery actions, task status, and one line of overall guidance.
  • Enhanced Agent Performance: Demonstrates significant improvements in the resolve rate of various coding agents (e.g., Qwen3-32B, Qwen3-Next-80B-A3B, GPT-OSS-20B) on the SWE-bench Verified benchmark when integrated.
  • Training Data: Fine-tuned on a dataset of 6,447 examples from R2E-Gym and SWE-bench Verified, with trajectories generated by CWM-32B and Qwen3-Next-80B-A3B-Instruct, and critiques provided by Claude Opus 4.6.
  • Context Length: Supports a sequence length of 32,768 tokens, allowing for analysis of extensive code agent trajectories.

When to Use This Model

This model is ideal for developers looking to:

  • Improve Code Agent Reliability: Integrate a critic to guide and improve the performance of existing large language model-based coding agents.
  • Automate Code Review Feedback: Generate automated, structured feedback on code generation processes without requiring direct code modification by the critic.
  • Research in Agent Steering: Utilize a small, specialized critic model for experiments in agent steering and reinforcement learning from human feedback (RLHF) or DPO, as it serves as the SFT base for DPO critics.

Finetunes
3 models