指令微调数据集

大型语言模型指令微调的所有可用数据集

黄金标准数据集

• P3: https://github.com/bigscience-workshop/promptsource, https://huggingface.co/datasets/bigscience/P3
  • 涵盖多种NLP任务的英语提示数据集集合
  • 270个数据集中的2000种提示类型
• xP3: https://huggingface.co/datasets/bigscience/xP3mt
  • 46种语言的13个训练任务混合，提示使用20种语言（从英语机器翻译而来）
• Natural Instructions v2: https://github.com/allenai/natural-instructions
  • 1,616个多样化NLP任务及其专家编写的指令基准，涵盖76种不同任务类型和55种不同语言。
• The Flan Collection: https://github.com/google-research/FLAN/tree/main/flan/v2
  • 包含此处部分数据集的超集
  • 1836个任务，1500万个示例
• Open Assistant: https://huggingface.co/datasets/OpenAssistant/oasst1
  • 人工标注的助手式对话语料，包含161,443条消息，分布在66,497个对话树中，涉及35种不同语言，标注了461,292个质量评级
• LIMA: 1000条高质量指令
  • https://huggingface.co/datasets/GAIR/lima
• databricks-dolly-15k: https://github.com/databrickslabs/dolly/tree/master/data
• PRESTO: https://github.com/google-research-datasets/presto
  • 55万条人类与虚拟助手之间的多语言上下文对话
• BB3x: https://parl.ai/projects/bb3x/
• InstructCTG: https://github.com/MichaelZhouwang/InstructCTG
  • 受控生成框架 https://arxiv.org/abs/2304.14293
• CrossFit: https://github.com/INK-USC/CrossFit
• tasksource: https://arxiv.org/abs/2301.05948
• ExMix: https://arxiv.org/abs/2111.10952
• InstructEval: https://github.com/declare-lab/instruct-eval
• M3IT: https://huggingface.co/datasets/MMInstruction/M3IT
  • https://arxiv.org/abs/2306.04387
  • 240万多模态实例和400条指令，涵盖40个任务和80种语言
• MIMIC-IT: 多模态上下文指令微调: https://arxiv.org/abs/2306.05425
• MultiInstruct: https://github.com/VT-NLP/MultiInstruct
• COLLIE: https://github.com/princeton-nlp/Collie
• Mind2Web: 面向网络的通用智能体 https://osu-nlp-group.github.io/Mind2Web/
• Android in the Wild: 大规模Android设备控制数据集: https://github.com/google-research/google-research/tree/master/android_in_the_wild
• FLASK: 基于对齐技能集的细粒度语言模型评估 https://github.com/kaistAI/FLASK
• Safe-RLHF: https://arxiv.org/abs/2310.12773
  • https://arxiv.org/pdf/2310.12773.pdf#https%3A//github.com/PKU-Alignment/safe-rlhf
• HelpSteer: https://huggingface.co/datasets/nvidia/HelpSteer

次优标准/使用语言模型生成

• Self-Instruct: https://github.com/yizhongw/self-instruct
• Unnatural Instructions: https://github.com/orhonovich/unnatural-instructions
• Alpaca: https://huggingface.co/datasets/tatsu-lab/alpaca
  • Alpaca-Clean: https://github.com/gururise/AlpacaDataCleaned
• Code Alpaca: https://github.com/sahil280114/codealpaca
• AlpacaGPT3.5Customized: https://huggingface.co/datasets/whitefox44/AlpacaGPT3.5Customized
• GPT4All: https://github.com/nomic-ai/gpt4all
  • GPT4All-pruned: https://huggingface.co/datasets/Nebulous/gpt4all_pruned
• ShareGPT: https://huggingface.co/datasets/RyokoAI/ShareGPT52K
• GPTeacher: https://github.com/teknium1/GPTeacher
• CAMEL🐪: https://www.camel-ai.org/
• 人类与ChatGPT对比语料库: https://github.com/Hello-SimpleAI/chatgpt-comparison-detection
• InstructionWild: https://github.com/XueFuzhao/InstructionWild
• 使用GPT-4进行指令调优: https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM
• Guanaco: https://huggingface.co/datasets/JosephusCheung/GuanacoDataset
• LongForm数据集: https://github.com/akoksal/LongForm/tree/main/dataset
  • 为多样化语料样本生成LLM指令（27,739对指令和长文本）
• UltraChat: https://huggingface.co/datasets/stingning/ultrachat
• LLaVA视觉指令150K: https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K
  • GPT生成的多模态指令执行数据
• GPT4Tools: https://github.com/StevenGrove/GPT4Tools
  • 用于调用多个多模态模型API的指令数据
• LaMini-Instruction: https://huggingface.co/datasets/MBZUAI/LaMini-instruction
  • 258万对指令和回复
• Evol-Instruct 70k: https://github.com/nlpxucan/WizardLM
• Dynosaur: https://dynosaur-it.github.io/
• Alpaca-Farm: https://github.com/tatsu-lab/alpaca_farm
  • https://huggingface.co/datasets/tatsu-lab/alpaca_farm
• ign_clean_instruct_dataset_500k: https://huggingface.co/datasets/ignmilton/ign_clean_instruct_dataset_500k
• airoboros: https://github.com/jondurbin/airoboros
• UltraFeedback: https://huggingface.co/datasets/openbmb/UltraFeedback
• WildChat: 570K真实用户-ChatGPT交互语料 https://wildchat.allen.ai/
• 反馈收集: https://arxiv.org/abs/2310.08491
  • https://huggingface.co/datasets/kaist-ai/Feedback-Collection

偏好数据集（可用于训练奖励模型）

• HH-RLHF: https://huggingface.co/datasets/Anthropic/hh-rlhf
  • 包含人类对模型输出的有害性和有用性评分。该数据集包含约16万个人工评分示例，每个示例由聊天机器人的一对回复组成，其中一个是人类偏好的。
• OpenAI WebGPT: https://huggingface.co/datasets/openai/webgpt_comparisons
  • 包含约2万个比较，每个示例包括一个问题、一对模型答案和元数据。答案由人类根据偏好评分。
• OpenAI总结: https://huggingface.co/datasets/openai/summarize_from_feedback
  • 包含约9.3万个示例，每个示例包含人类对模型生成摘要的反馈。人类评估员从两个选项中选择更优秀的摘要。
• 斯坦福人类偏好数据集（SHP）: https://huggingface.co/datasets/stanfordnlp/SHP
  • 38.5万个人类对18个不同主题的问题/指令回复的集体偏好
• Stack Exchange偏好: https://huggingface.co/datasets/HuggingFaceH4/stack-exchange-preferences
• SLF5K: https://huggingface.co/datasets/JeremyAlain/SLF5K
• qa-from-hf: https://github.com/lil-lab/qa-from-hf
• Nectar: https://huggingface.co/datasets/berkeley-nest/Nectar
• JudgeLM-100K: https://huggingface.co/datasets/BAAI/JudgeLM-100K
• UltraFeedback: https://huggingface.co/datasets/openbmb/UltraFeedback

杂项

• OIG: https://huggingface.co/datasets/laion/OIG
  • 这里一些数据集的超集
• oa_leet10k: https://huggingface.co/datasets/ehartford/oa_leet10k
  • 用多种编程语言解决的LeetCode问题
• ProSocial对话: https://huggingface.co/datasets/allenai/prosocial-dialog
• ConvoKit: https://convokit.cornell.edu/documentation/datasets.html
• CoT-Collection: https://github.com/kaist-lklab/CoT-Collection
• DialogStudio: https://github.com/salesforce/DialogStudio
• Chatbot Arena对话 https://huggingface.co/datasets/lmsys/chatbot_arena_conversations
• lmsys 1M: https://huggingface.co/datasets/lmsys/lmsys-chat-1m
• 对话编年史: https://conversation-chronicles.github.io/