stories
收藏资源简介:
stories数据集是一个包含1,712个英文故事的系统生成集合,专为微调较小的自回归语言模型而设计,特别适用于连贯故事写作类任务,建议与其他数据结合使用以增强效果。其主要动机是作为后续处理(例如,将句子翻译成德语、乌尔都语等其他语言)的基础语料,从而辅助训练专注于翻译任务的小型语言模型。数据由六个不同的对话/指令微调自回归语言模型生成,包括:meta-llama/Llama-3.1-8B-Instruct、meta-llama/Llama-3.3-70B-Instruct、google/gemma-4-31B-it、openai/gpt-oss-120b、zai-org/GLM-4.7和Qwen/Qwen3.5-27B。生成过程通过脚本系统性地覆盖了九种故事叙述组合,涉及三种时态组(过去时、现在时、现在+将来时)和三种叙述视角(第一人称“I”、第三人称单数“He/She”、第三人称复数“They”)。数据规模的控制通过统计句子终结符(., !, ?)来近似估算每个组合的句子总量。生成后对数据进行了清理:移除了模型原生推理功能可能产生的推理内容以保持简洁,并将一些过度拟合的UTF标点符号(如弯引号“”和特定连字符‐)替换为标准ASCII字符(直引号"和减号-)。生成工作由Cerebras Systems Inc.和Groq提供的免费、高效推理服务支持,耗时约一周完成,且模型未进行量化处理。数据集采用Apache License, Version 2.0许可发布,并可通过提供的Python脚本示例便捷访问。
The stories dataset is a systematically generated collection of 1,712 English stories, designed to provide resources for fine-tuning smaller autoregressive language models, particularly suitable for coherent story writing tasks, and it is recommended to be used in combination with other data for enhanced effectiveness. Its primary motivation is to serve as a foundational corpus for subsequent processing (e.g., translating sentences into other languages such as German, Urdu), thereby assisting in training small language models focused on translation tasks. The data is generated by six different conversational/instruction-tuned autoregressive language models, including: meta-llama/Llama-3.1-8B-Instruct, meta-llama/Llama-3.3-70B-Instruct, google/gemma-4-31B-it, openai/gpt-oss-120b, zai-org/GLM-4.7, and Qwen/Qwen3.5-27B. The generation process systematically covers nine story narrative combinations through scripts, involving three tense groups (past, present, present+future) and three narrative perspectives (first-person I, third-person singular He/She, third-person plural They). Data scale control is not directly based on story count but approximated by counting sentence terminators (., !, ?) to estimate the total number of sentences per combination. After generation, the data was cleaned: removing reasoning content that may arise from the models native inference capabilities to maintain conciseness, and replacing some overfitted UTF punctuation marks (such as curly quotes “ ” and specific hyphens ‐) with standard ASCII characters (straight quotes " and hyphens -). The generation work was supported by free and efficient inference services provided by Cerebras Systems Inc. and Groq, taking about a week to complete, and the models were not quantized. The dataset is released under the Apache License, Version 2.0, and can be conveniently accessed via provided Python script examples.
数据集概述
数据集名称:stories
许可协议:Apache-2.0
任务类别:文本生成(text-generation)
语言:英语(en)
标签:合成数据(synthetic)、故事(story)、文本(text)
数据规模:1,000 至 10,000 条样本(实际包含 1,712 条故事)
数据集描述
该数据集包含 1,712 条系统生成的故事,由多种自回归语言模型生成,所有故事均使用英语表达。数据集采用 Apache License 2.0 许可发布,推荐用于微调较小的自回归语言模型,特别是在连贯故事写作等任务上,若结合更多数据效果更佳。生成该数据集的目的还在于进一步处理,将所有句子翻译成其他语言(如德语)和低资源语言(如乌尔都语)。
下载与使用
可通过编程语言(如 Python 及其 requests 库)轻松下载数据并选取特定样本。以下是一个最小示例:
python from requests import get
data: list[str] = get("https://huggingface.co/datasets/hamzah0asadullah/stories/raw/main/data.json").json()
len(data) -> 1712
print(data[int(1712 / 2)])
生成方式
使用六种不同的对话模型生成文本,生成脚本由 Python 编写。脚本遍历了三种时间时态(过去、现在、现在+未来)和三种故事叙事人称(我、他/她、他们(指一群人))的所有九种组合。对于每种组合,脚本通过统计字符 .、! 和 ? 的数量来近似估算所有故事包含的句子总数。
使用的生成模型
- meta-llama/Llama-3.1-8B-Instruct
- meta-llama/Llama-3.3-70B-Instruct
- google/gemma-4-31B-it
- openai/gpt-oss-120b
- zai-org/GLM-4.7
- Qwen/Qwen3.5-27B
生成细节
- 当模型原生支持推理时,该功能被启用,但推理内容被剥离以保持数据简洁。
- 某些模型上过拟合的 UTF 字符被替换为正常的 ASCII 对应字符,例如将
“和”替换为",将连字符(‐)替换为简单减号(-)。 - 模型未进行量化处理,提供推理服务的机构为 Cerebras Systems Inc. 和 Groq,这使得数据集能够在约一周内免费快速生成。





