weixinchen/GRATH
收藏Hugging Face2024-07-17 更新2024-07-22 收录
下载链接:
https://hf-mirror.com/datasets/weixinchen/GRATH
下载链接
链接失效反馈官方服务:
资源简介:
GRATH数据集用于论文《GRATH: Gradual Self-Truthifying for Large Language Models》中的成对真实性训练数据。数据集包含由Llama2-7B、Llama2-13B和Zephyr-7B生成的三个子集,每个子集包含pretrained_gen和selftruth_gen两个分割。pretrained_gen分割包含由预训练基础模型生成的数据,用于DPO自我真实性训练;selftruth_gen分割包含由自我真实性模型生成的数据,用于逐步自我真实性训练。每个数据条目包含一个问题、一个正确答案和一个错误答案。
The GRATH dataset is used for pairwise truthfulness training data in the paper GRATH: Gradual Self-Truthifying for Large Language Models. The dataset contains three subsets generated by Llama2-7B, Llama2-13B, and Zephyr-7B, each containing pretrained_gen and selftruth_gen splits. The pretrained_gen split contains data generated by the pretrained base model, used for DPO self-truthifying training; the selftruth_gen split contains data generated by the self-truthified model, used for gradual self-truthifying training. Each data entry includes a question, a correct answer, and an incorrect answer.
提供机构:
weixinchen
原始信息汇总
数据集概述
数据集名称
- GRATH
数据集配置
- llama2_7b
- 数据文件
pretrained_gen分割:llama2_7b/iter0_arcc_num_1119_useGT_False_useFS_True.jsonselftruth_gen分割:llama2_7b/iter1_arcc_num_1119_useGT_False_useFS_True_combined.json
- 数据文件
- llama2_13b
- 数据文件
pretrained_gen分割:llama2_13b/iter0_arcc_num_1119_useGT_False_useFS_True_seed_0.jsonselftruth_gen分割:llama2_13b/iter1_arcc_num_1119_useGT_False_useFS_True_seed_0_combined.json
- 数据文件
- zephyr_7b
- 数据文件
pretrained_gen分割:zephyr_7b/iter0_arcc_num_1119_useGT_False_useFS_True_seed_0.jsonselftruth_gen分割:zephyr_7b/iter1_arcc_num_1119_useGT_False_useFS_True_combined.json
- 数据文件
任务类别
- 文本分类
- 问答
- 文本生成
语言
- 英语
数据集大小
- n<1K
数据格式
- 每行包含一个真实性数据对,包括一个问题、一个正确答案和一个错误答案。
示例
json { "question": "George wants to warm his hands quickly by rubbing them. Which skin surface will produce the most heat?", "correct": "The palms of his hands will produce the most heat.", "incorrect": "The soles of his feet will produce the most heat." }
数据生成模型
- llama2_7b:由 Llama2-7B 生成
- llama2_13b:由 Llama2-13B 生成
- zephyr_7b:由 Zephyr-7B 生成
数据分割说明
pretrained_gen分割:由预训练的基础模型生成的数据,用于自真实化(DPO)。selftruth_gen分割:由自真实化模型生成的数据,用于逐步自真实化(DPO)。



