andyrdt/gpt-oss-20b-rollouts

该数据集包含了从GPT-OSS-20B生成的回滚数据，具有解析的Harmony通道（助手思考/最终）。数据集的模式包括用户内容、系统推理努力、助手思考和助手内容。安全性和jailbreak相关的数据集包括HarmBench、HarmfulGeneration-HarmBench、StrongREJECT等。数学和推理相关的数据集包括gsm8k、MATH-500、mmlu等。聊天和指令相关的数据集包括WildChat-1M、ultrachat_200k、oasst1等。代码编程相关的数据集包括openai_humaneval、mbpp、apps等。

This dataset contains rollouts generated from GPT-OSS-20B with parsed Harmony channels (assistant thinking/final). The dataset schema includes user content, system reasoning effort, assistant thinking, and assistant content. Safety and jailbreak-related datasets include HarmBench, HarmfulGeneration-HarmBench, StrongREJECT, etc. Math and reasoning-related datasets include gsm8k, MATH-500, mmlu, etc. Chat and instruction-related datasets include WildChat-1M, ultrachat_200k, oasst1, etc. Code programming-related datasets include openai_humaneval, mbpp, apps, etc.