costadev00/dolly-15k-rlhf-instructgpt-format

该数据集名为Dolly 15k RLHF Datasets in InstructGPT Format，是基于databricks/databricks-dolly-15k改造的专门用于人类反馈强化学习(RLHF)的数据集。它包含四种配置：1) sft：监督微调示例，包含prompt、completion和text；2) rm_schema：奖励模型架构/提示池，包含空的chosen和rejected字段、reference_response以及ready_for_rm=false标记；3) rm_synthetic：使用Dolly的reference_response作为chosen，采样GPT-2 SFT输出作为rejected的奖励模型代理对；4) ppo：用于PPO/RLHF展开的仅包含prompt的示例。数据集采用InstructGPT风格的纯文本格式，包含训练集(12010条)、验证集(1502条)和测试集(1499条)，适用于文本生成、问答和摘要等NLP任务。

This dataset, named Dolly 15k RLHF Datasets in InstructGPT Format, is derived from databricks/databricks-dolly-15k and specifically designed for Reinforcement Learning from Human Feedback (RLHF). It includes four configurations: 1) sft: supervised fine-tuning examples with prompt, completion, and text; 2) rm_schema: reward-modeling schema/prompt pool with empty chosen and rejected fields, reference_response, and ready_for_rm=false; 3) rm_synthetic: reward-modeling proxy pairs where Dolly reference_response is used as chosen and sampled GPT-2 SFT output as rejected; 4) ppo: prompt-only examples for PPO/RLHF rollouts. The dataset uses a plain textual InstructGPT-style format and contains train (12,010), validation (1,502), and test (1,499) splits, suitable for NLP tasks like text generation, question answering, and summarization.