README.md

metadata

task_categories:
  - text-generation
language:
  - zh
  - en
license: cc-by-sa-4.0

🔥 Updates

[2024.6.4] Add a slim version. The sample number is reduced from about 20k to 10k.

[2024.5.28]

1. The data format is converted from "chatml" to "messages", which is more convenient to use tokenizer.apply_chat_template. The old version has been moved to "legacy" branch.
2. The version without "Original text paraphrasing" is added.

📊 Long Context Instruction-tuning dataset with "Original text paraphrasing"

• Paper
• Github
• consist of multiple tasks
• Chinese and English
• sample length ranging from 4k to 32k
• the answer contains "Original text paraphrasing" part

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

Dataset Composition (original version)

Dataset Composition (slim version)

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

1. 图书总结

2. 论文摘要 涉及到知网数据，受限访问。

3. 论文问答 涉及到知网数据，受限访问。

4. 多文档问答（检索）

英文

1. 多文档问答（检索）

中英

1. 长论文多任务

2. 从ShareGPT中筛选的长对话（中英）

3. 预训练长文本语料库（中英）LongData-Corpus