critic-model
critic-sft-cwm-only-detailed-prompt
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… See the full description on the dataset page: https://huggingface.co/datasets/code-critic-model/critic-sft-cwm-only-detailed-prompt.critic-sft-qwen-only
critic-sft-qwen-only
The Qwen-only SFT corpus from Steer, Don't Solve: Training Small Critic Models for Large Code Agents. It trains Qwen3-8B-Critic-SFT-Qwen-only and Qwen3-4B-Critic-SFT-Qwen-only, the Qwen-only arms of the corpus ablation in Table 3.
Each record is one critique point: a Qwen3-Next-80B-A3B-Instruct 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, is… See the full description on the dataset page: https://huggingface.co/datasets/code-critic-model/critic-sft-qwen-only.critic-sft-cwm-qwen
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… See the full description on the dataset page: https://huggingface.co/datasets/code-critic-model/critic-sft-cwm-qwen.critic-sft-cwm-only
critic-sft-cwm-only
The CWM-only SFT corpus from Steer, Don't Solve: Training Small Critic Models for Large Code Agents. It trains 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, is this set plus 1,915 records from Qwen3-Next trajectories.… See the full description on the dataset page: https://huggingface.co/datasets/code-critic-model/critic-sft-cwm-only.PRM_1541i
Not the paper's DPO data. This is an earlier set of 1,541 preference pairs used for development DPO sweeps. The released critic Qwen3-4B-Critic-SFT-DPO from Steer, Don't Solve was trained on a different set of 1,409 pairs built with the same procedure. See the organization page for the artifacts behind the paper.
PRM_1541i
1541 preference pairs for training a critic over coding-agent trajectories. Each example is a
multi-turn agent transcript paired with two candidate… See the full description on the dataset page: https://huggingface.co/datasets/code-critic-model/PRM_1541i.UBIQUANT_critic_model_data