code-critic-model/Qwen3-8B-Critic-SFT-Detailed-Prompt
Text Generation • 1B • Updated • 320
messages listlengths 13 103 |
|---|
[
{
"role": "system",
"content": "You are a supervisor monitoring an LLM-based coding agent. Your role is to detect trajectory-level errors and provide corrective guidance to prevent task failure.\nAnalyze the provided agent trace for the following error categories. For each detected error, provide specific e... |
[
{
"role": "system",
"content": "You are a supervisor monitoring an LLM-based coding agent. Your role is to detect trajectory-level errors and provide corrective guidance to prevent task failure.\nAnalyze the provided agent trace for the following error categories. For each detected error, provide specific e... |
[{"role":"system","content":"You are a supervisor monitoring an LLM-based coding agent. Your role is(...TRUNCATED) |
[{"role":"system","content":"You are a supervisor monitoring an LLM-based coding agent. Your role is(...TRUNCATED) |
[{"role":"system","content":"You are a supervisor monitoring an LLM-based coding agent. Your role is(...TRUNCATED) |
[{"role":"system","content":"You are a supervisor monitoring an LLM-based coding agent. Your role is(...TRUNCATED) |
[{"role":"system","content":"You are a supervisor monitoring an LLM-based coding agent. Your role is(...TRUNCATED) |
[{"role":"system","content":"You are a supervisor monitoring an LLM-based coding agent. Your role is(...TRUNCATED) |
[{"role":"system","content":"You are a supervisor monitoring an LLM-based coding agent. Your role is(...TRUNCATED) |
[{"role":"system","content":"You are a supervisor monitoring an LLM-based coding agent. Your role is(...TRUNCATED) |
End of preview. Expand in Data Studio
critic-sft-cwm-only-detailed-prompt
The detailed-prompt SFT corpus from Steer, Don't Solve: Training Small Critic Models for Large Code Agents. It trains Qwen3-8B-Critic-SFT-Detailed-Prompt, the comparison arm of the prompt ablation in Table 4.
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. The difference from critic-sft-cwm-only is the teacher prompt. Here the teacher used the detailed prompt, which allows concrete code-level suggestions, so critiques are longer and often contain code. The high-level prompt used for every other corpus forbids code and asks for one or two sentences of guidance.
| Records | 3,135 |
| 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 | detailed (prm_issue_res in the repository) |
| Agent scaffold | mini-swe-agent, one bash command per step |
| Length limit | 32,768 tokens per record |
Fewer records survive the 32,768 token limit than in the high-level corpus because detailed critiques are longer. The paper's Table 5 measures how much of the agent's final patch gets copied from detailed critiques, which motivates the high-level 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 is on the 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
from datasets import load_dataset
ds = load_dataset("code-critic-model/critic-sft-cwm-only-detailed-prompt", 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 repository.
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}
}
- Downloads last month
- 162
Models trained or fine-tuned on code-critic-model/critic-sft-cwm-only-detailed-prompt