Creative Rubrics Preferences
收藏资源简介:
本数据集名为“Creative Rubrics Preferences”,由Komorebi AI Technologies创建,包含900条样本。数据集利用合成生成的偏好数据,这些数据基于细粒度的评分标准,用于定义诸如写作风格等期望属性。数据集的创建过程包括定义评分标准、生成评分条件下的文本、系统提示的生成以及构建偏好对。该数据集用于训练语言模型,使其能够根据明确的、人类可解释的指令动态调整其行为,无需重新训练。数据集的应用领域为AI对齐,旨在解决语言模型与人类偏好、行为和安全协议的对齐问题。
This dataset, named "Creative Rubrics Preferences", was developed by Komorebi AI Technologies and comprises 900 samples. It utilizes synthetically generated preference data grounded in fine-grained scoring rubrics that delineate desired attributes such as writing styles. The dataset creation pipeline includes defining scoring rubrics, generating texts that adhere to the specified scoring criteria, formulating system prompts, and constructing preference pairs. This dataset is intended for training language models to dynamically adjust their behaviors in accordance with explicit, human-interpretable instructions without the need for retraining. Its application domain is AI alignment, with the core objective of resolving alignment issues between language models and human preferences, behaviors, and safety protocols.
数据集概述:creative-rubrics-preferences
基本信息
- 语言: 英语 (en)
- 许可证: Apache 2.0
- 数据规模: 小于1K样本 (n<1K)
- 任务类型: 文本生成 (text-generation)
- 标签: dpo, preferences, creative, gpt-4.5, o3-mini, R1
数据集内容
- 特征:
- model: 字符串类型,表示生成文本的模型
- prompt: 字符串类型,表示生成文本的提示
- chosen: 字符串类型,表示被选中的生成文本
- rejected: 字符串类型,表示被拒绝的生成文本
- qualifier: 字符串类型,表示生成文本的限定条件
- 数据分割:
- train: 包含900个样本,大小为4647924字节
数据集用途
- 用于微调自定义写作风格的任务
- 基于论文《Configurable Preference Tuning with Rubric-Guided Synthetic Data》的研究
相关资源
- 论文链接: https://huggingface.co/papers/2506.11702
- 代码仓库: https://github.com/vicgalle/configurable-preference-tuning
- 相关数据集: https://huggingface.co/datasets/vicgalle/creative-rubrics-gpt-4.5-o3-R1
示例
- 电影评论示例: 关于《疯狂动物城》的摄影风格评论,风格华丽且富有创意
- 天气预报示例: 描述一个拥有五个太阳的星球的天气预报,风格荒诞且富有想象力




