File size: 3,308 Bytes
4f6d42e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 | ---
license: apache-2.0
task_categories:
- text-generation
language:
- en
size_categories:
- 1K<n<10K
tags:
- code-agent
- critic
- process-reward-model
- swe-bench
- sft
- llama-factory
configs:
- config_name: default
data_files:
- split: train
path: prm_sft_train.jsonl
---
# critic-sft-cwm-only
The CWM-only SFT corpus from [Steer, Don't Solve: Training Small Critic Models for Large Code Agents](https://arxiv.org/abs/2606.21811). It trains [Qwen3-8B-Critic-SFT-CWM-only](https://huggingface.co/code-critic-model/Qwen3-8B-Critic-SFT-CWM-only), the `CWM-only` arm of the corpus ablation in Table 3.
Each record is one critique point: a CWM-32B trajectory up to some step, followed by the critique that Claude Opus 4.6 wrote for it under the paper's high-level prompt. The main corpus, [critic-sft-cwm-qwen](https://huggingface.co/datasets/code-critic-model/critic-sft-cwm-qwen), is this set plus 1,915 records from Qwen3-Next trajectories.
| | |
|---|---|
| Records | 4,532 |
| Trajectories | 500, all from CWM-32B |
| 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 |
| Teacher prompt | high-level (`prm_issue_res_instructions` in the repository) |
| Agent scaffold | mini-swe-agent, one bash command per step |
| Length limit | 32,768 tokens per record |
The same 500 tasks and trajectories, critiqued with the detailed prompt instead, are in [critic-sft-cwm-only-detailed-prompt](https://huggingface.co/datasets/code-critic-model/critic-sft-cwm-only-detailed-prompt).
## Format
One JSON object per line with a single field, `messages`, in ShareGPT layout: a `system` turn with the critic instructions, alternating `user` / `assistant` turns carrying the agent's trajectory (task and command output on the user side, agent reasoning and bash command on the assistant side), and a final `assistant` turn holding the teacher critique. Critiques the agent received earlier in the trajectory appear inside the user turn that follows them. The full description of the format, with an example critique, is on the [critic-sft-cwm-qwen](https://huggingface.co/datasets/code-critic-model/critic-sft-cwm-qwen) card.
## Files
| File | Purpose |
|---|---|
| `prm_sft_train.jsonl` | the records |
| `dataset_info.json` | LLaMA-Factory dataset registration (`sharegpt` format) |
| `metadata.json` | provenance: source run, agent, teacher, counts, instance ids |
## Using it
```python
from datasets import load_dataset
ds = load_dataset("code-critic-model/critic-sft-cwm-only", split="train")
```
For LLaMA-Factory, point `dataset_dir` at a download of this repository and use `dataset: prm_sft_train` with the config `finetuning/qwen3_8b_critic_full_sft_l40s_train_multiturn_resumable.yaml` from the [critic-training](https://github.com/shubhamrgandhi/critic-training) repository.
## Citation
```bibtex
@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}
}
```
|