SkySyrup/tinystories_german

---
language:
- de
license: apache-2.0
---

### What have you done
this dataset is a german interpretation of the roneneldan/TinyStories dataset

that dataset is amazing- I wanted to make a german version to experiment with the bilinguality of tiny language models (more coming on that soon!!!) (i wrote a paper :D)

this is the result of a bunch of work and months of screwing around

it was made with basically 0 budget;

----
argos-translate contains 200k opennmt translated tinystories

german_GEMINI_async-combined contains about 180k synthetic generated german stories with some extremely generous token donations from google (potentially unwilling)

v2_GERMAN.txt contains about 80k stories generated with leo-hessenai-7b-chat

v4_GERMAN.txt contains about 50-60k stories generated with leo-mistral-hessenai-7b-chat

----

there were some smaller datasets that weren't uploaded because they sucked due to the models used to generate them

the models trained with these datasets and a custom tokenizer show a lot of promise, they aren't quite the level of the english stories but they're pretty good for my standards

it took me about 5 months of on-and-off generating and working to create these datasets;

## Dataset usage

do whatever you want with it, i don't like copyright

but if you build something cool maybe say it contains/ was based off this dataset and link to it pretty please

## Limitations and biases
seriously? this dataset was made with models not intended for nsfw/rp and also not prompted in such a way

get off my back ai safety people

原始信息汇总

数据集概述

该数据集是基于roneneldan/TinyStories数据集的德语版本，旨在探索小型语言模型的双语能力。数据集的制作过程历时数月，且基本无预算。

数据集组成

• argos-translate 包含约200k条通过OpenNMT翻译的TinyStories。
• german_GEMINI_async-combined 包含约180k条合成生成的德语故事，部分由Google慷慨捐赠（可能非自愿）。
• v2_GERMAN.txt 包含约80k条通过leo-hessenai-7b-chat生成的故事。
• v4_GERMAN.txt 包含约50-60k条通过leo-mistral-hessenai-7b-chat生成的故事。

数据集制作

数据集的制作过程中，有一些较小的数据集未上传，因为它们的质量较差，主要是因为生成它们所用的模型不佳。使用这些数据集和自定义分词器训练的模型显示出很大的潜力，虽然它们尚未达到英语故事的水平，但已符合我的标准。

数据集使用

该数据集的使用不受限制，作者不支持版权。如果使用该数据集构建了有趣的项目，建议提及该数据集并链接到其来源。

局限性与偏见

该数据集由不适用于NSFW/RP的模型生成，且未以相关方式提示。