SupraLabs/supra-wild-titles-130k
收藏资源简介:
Wild Title是一个由SupraLabs策划的数据集系列,专门用于聊天标题生成的训练和评估。它包含从小众和高度专业化的对话样本中特意划分出来的数据,这些样本来自主要的标题数据集。这种分离使研究人员和开发者能够在更广泛、更全面的标题生成任务范围内稳健地训练和评估模型。与主要捕捉通用、高频对话的标准聊天标题数据集不同,Wild Title隔离了那些在传统基准测试中通常代表性不足的不常见、特定领域和边缘情况的交互。该数据集基于allenai/WildChat的用户交互日志,具有高度可变的序列长度,用户输入范围从简短的1-2字符短语到长达1000-2500词的提示。这种设计引入了结构多样性和分布方差,有助于小型语言模型(SLMs)和大型语言模型(LLMs)实现更优的泛化能力,同时通过减少对稀有或长尾令牌的参数分配来优化令牌效率。
Wild Title is a dataset series curated by SupraLabs designed specifically for training and evaluation in chat title generation. It comprises niche and highly specialized conversation samples intentionally partitioned from our primary title datasets. This separation enables researchers and developers to robustly train and evaluate models across a more diverse, comprehensive spectrum of title-generation tasks. Unlike standard chat title datasets that primarily capture general-purpose, high-frequency conversations, Wild Title isolates uncommon, domain-specific, and edge-case interactions that are typically underrepresented in conventional benchmarks. The dataset features highly variable sequence lengths, with user inputs ranging from brief 1–2 character phrases to extensive 1,000–2,500 word prompts. This design introduces explicit structural diversity and distribution variance, enabling Small Language Models (SLMs) and Large Language Models (LLMs) to achieve superior generalization capabilities while optimizing token efficiency by mitigating parameter allocation on rare or long-tail tokens.
数据集概述:SupraLabs/supra-wild-titles-130k
- 数据集来源:由 SupraLabs 提供,位于 Hugging Face 平台。
- 任务类型:特征提取(Feature Extraction)、摘要生成(Summarization)。
- 语言:英语(English)。
- 数据集规模:100K 至 1M 条记录,具体为 134k 行。
- 标签:chat-titles, chat_titles, chat, title, chat-title, chat_title。
- 许可证:odc-by(Open Data Commons Attribution License)。
- 数据集结构:
- 子集:
default,包含train分割,共 134k 行。 - 数据字段:
user(字符串,用户查询内容)、title(字符串,生成的标题)。
- 子集:
- 数据示例:包含各种用户查询及其对应标题,例如用户指令“Write a very long, elaborate, descriptive and detailed shooting script...”对应的标题为“John Waters Omelette Scene”,以及其他类似配对。




