metadata
task_categories:
- text-generation
language:
- zh
- en
license: cc-by-sa-4.0
- 2024.5.28 update: the data format is converted from "chatml" to "messages", which is more convenient to use
tokenizer.apply_chat_template
. And the version without "Original text paraphrasing" is added.
Long Context Instruction-tuning dataset with "Original text paraphrasing"
- consist of multiple tasks
- Chinese and English
- sample length ranging from 4k to 32k
- the answer contains "Original text paraphrasing" part
- Paper
长文本指令微调数据
- 此数据集由多种长文本任务数据集组合而成。
- 包含中文和英文两部分
- Paper
源数据
此处给出各个数据集的链接集合。也可以直接点击我的个人主页查看所有数据集。
中文
英文
中英
预训练长文本语料库(中英)LongData-Corpus