jayshah5696/humanize-rl-tasks-v03

humanize-rl RL任务v03数据集包含492个强化学习任务，旨在训练和评估在特定约束下生成类人散文的模型。每一行将一个编写的指令与一个约束规范配对，该规范可由humanize-rl奖励环境（确定性检查+第一层启发式+岭回归风格评分器）使用。492 RL tasks designed to train and evaluate models that produce human-sounding prose under specific constraints. Each row pairs an authored instruction with a constraint spec consumable by the humanize-rl reward environment (deterministic checks + Layer-1 heuristics + ridge regression style scorer).