critic-sft-cwm-qwen

The main SFT corpus from Steer, Don't Solve: Training Small Critic Models for Large Code Agents. It trains both Qwen3-8B-Critic-SFT and Qwen3-4B-Critic-SFT.

Each record is one critique point: a coding agent's trajectory up to some step, followed by the critique that Claude Opus 4.6 wrote for it. The teacher was prompted with the paper's high-level prompt, which asks for error detection and one or two sentences of guidance and forbids code and commands in the answer.

Records	6,447
Trajectories	983 (500 from CWM-32B, 483 from Qwen3-Next-80B-A3B-Instruct)
Records per agent	4,532 from CWM, 1,915 from Qwen3-Next
Tasks	500 R2E-Gym instances from matplotlib, moto, and sympy, disjoint from SWE-bench Verified
Teacher	Claude Opus 4.6 (us.anthropic.claude-opus-4-6-v1), queried every 5 agent steps
Agent scaffold	mini-swe-agent, one bash command per step
Length limit	32,768 tokens per record

This corpus is the union of critic-sft-cwm-only and critic-sft-qwen-only. The paper's Table 3 compares critics trained on each of the three.

Format

One JSON object per line with a single field, messages, in ShareGPT layout.

• system: the critic's instructions. Twelve error categories in three groups (specification, reasoning, coordination), the response format, and one worked example.
• user / assistant turns: the agent's trajectory. The first user turn holds the PR description and task instructions; later user turns hold command output; assistant turns hold the agent's reasoning and its bash command. Critiques the agent received earlier in the trajectory appear inside the user turn that follows them.
• final assistant turn: the teacher's critique for this point in the trajectory. This is the only turn the critic is trained to produce.

A critique looks like this:

SPECIFICATION ERRORS:
1. Task Specification Violations: DETECTED: No
2. Role Specification Violations: DETECTED: No
3. Step Repetition: DETECTED: Yes
EVIDENCE: The agent has repeatedly failed to insert the fix correctly, restoring the file and retrying with similar approaches.
RECOVERY_ACTION: Write the edit script to a file first, then execute it, to avoid shell escaping issues.
4. Termination Condition Unawareness: DETECTED: No

REASONING ERRORS:
...
TASK_STATUS: Needs correction
OVERALL_GUIDANCE: ...

Records longer than 32,768 tokens under the Qwen3 chat template had their trajectory prefix shortened to fit. metadata.json records how many were affected, along with the source run directories and the list of task instance ids per agent.

Files

File	Purpose
prm_sft_train.jsonl	the records
dataset_info.json	LLaMA-Factory dataset registration (sharegpt format)
metadata.json	provenance: source runs, agents, teacher, counts, instance ids

Using it

With datasets:

from datasets import load_dataset
ds = load_dataset("code-critic-model/critic-sft-cwm-qwen", split="train")
print(ds[0]["messages"][-1]["content"])   # the teacher critique

With LLaMA-Factory, point dataset_dir at a download of this repository and use dataset: prm_sft_train. The exact training config is finetuning/qwen3_8b_critic_full_sft_l40s_train_multiturn_resumable.yaml in the critic-training repository; it uses the qwen3_nothink template, a 32,768 token cutoff, and mask_history: true so that loss is computed on the final critique only.

Citation

@misc{gandhi2026steerdontsolvetraining,
  title={Steer, Don't Solve: Training Small Critic Models for Large Code Agents},
  author={Shubham Gandhi and Yiqing Xie and Atharva Naik and Ruichen Zhu and Carolyn Rose},
  year={2026},
  eprint={2606.21811},
  archivePrefix={arXiv},
  primaryClass={cs.SE},
  url={https://arxiv.org/abs/2606.21811}
}