lichangh20/qwen3-coder-30b-swegym-train-eval-100-kl-cache

该数据集包含Qwen3-Coder-30B模型在SWE-Gym Train-Eval-100数据集上的教师滚动输出，用于离线KL/JSD散度评估。数据集包含100个任务的文件，每个文件包含模型生成的令牌、提示长度、响应长度、损失掩码、滚动对数概率、奖励分数等信息。适用于使用Qwen3 tokenizer的4B类学生模型，并且与Qwen3-4B-Instruct变体兼容。This dataset contains teacher rollouts from the Qwen3-Coder-30B model on the SWE-Gym Train-Eval-100 dataset, used for offline KL/JSD divergence evaluation. The dataset includes files for 100 tasks, each containing model-generated tokens, prompt length, response length, loss mask, rollout log probabilities, reward score, and other information. It is suitable for 4B-class student models using the Qwen3 tokenizer and is compatible with Qwen3-4B-Instruct variants.