LongReward-10k
收藏资源简介:
LongReward-10k数据集包含10,000个长上下文问答实例,涵盖英语和中文,每个实例最多可达64,000字。数据集分为三个部分:sft、dpo_glm4_9b和dpo_llama3.1_8b。sft部分包含通过GLM-4-0520模型生成的SFT数据,用于监督微调两个模型:LongReward-glm4-9b-SFT和LongReward-llama3.1-8b-SFT。dpo_glm4_9b和dpo_llama3.1_8b部分是长上下文偏好数据集,用于训练DPO模型:LongReward-glm4-9b-DPO和LongReward-llama3.1-8b-DPO。这些模型基于相应的SFT模型和LongReward方法进行训练。
The LongReward-10k dataset consists of 10,000 long-context question-answering instances covering both English and Chinese, with each instance containing up to 64,000 words. The dataset is divided into three splits: sft, dpo_glm4_9b, and dpo_llama3.1_8b. The sft split contains SFT data generated by the GLM-4-0520 model, which is used for supervised fine-tuning of two models: LongReward-glm4-9b-SFT and LongReward-llama3.1-8b-SFT. The dpo_glm4_9b and dpo_llama3.1_8b splits are long-context preference datasets used for training DPO models: LongReward-glm4-9b-DPO and LongReward-llama3.1-8b-DPO. All these models are trained based on their corresponding SFT models and the LongReward method.
LongReward-10k
概述
- 数据集名称: LongReward-10k
- 许可协议: Apache 2.0
- 任务类别: 文本生成
- 语言: 英语、中文
- 标签: 长上下文、大语言模型、RLAIF
- 数据规模: 1K<n<10K
数据集详情
- 数据文件:
sft:sft.jsonldpo_glm4_9b:dpo-glm4-9b.jsonldpo_llama3.1_8b:dpo-llama3.1-8b.jsonl
数据集描述
- LongReward-10k 包含10,000个长上下文问答实例(英语和中文,最长可达64,000字)。
sft分割包含由 GLM-4-0520 生成的SFT数据,遵循 LongAlign 中的自指导方法。dpo_glm4_9b和dpo_llama3.1_8b分割是长上下文偏好数据集,其中获胜和失败的响应从相应的SFT模型中采样,并由 LongReward 方法排名。
相关模型
- LongReward-glm4-9b-SFT: 基于 GLM-4-9B 训练。
- LongReward-llama3.1-8b-SFT: 基于 Meta-Llama-3.1-8B 训练。
- LongReward-glm4-9b-DPO: 基于
dpo_glm4_9b分割训练。 - LongReward-llama3.1-8b-DPO: 基于
dpo_llama3.1_8b分割训练。
引用
@article{zhang2024longreward, title = {LongReward: Improving Long-context Large Language Models with AI Feedback}, author = {Jiajie Zhang and Zhongni Hou and Xin Lv and Shulin Cao and Zhenyu Hou and Yilin Niu and Lei Hou and Yuxiao Dong and Ling Feng and Juanzi Li}, journal = {arXiv preprint arXiv:2410.21252}, year = {2024} }




