oliverkinch/autodata-da-sft
收藏资源简介:
Autodata DA 是一个用于指令微调的丹麦语合成数据集,通过两步骤的LLM流程从Dynaword语料库生成。每个示例都是丹麦语中真实的聊天机器人对话,涵盖多样化的日常用例:总结文档、回答关于文本的问题、起草内容、解释概念和提供建议。提示(prompt)始终是自包含的,包含回答问题所需的任何文档文本,因此在训练过程中不需要外部上下文。数据集包含2965个示例,分为训练集(2816个)和测试集(149个),指令类型包括SUMMARIZE、QA、EXPLAIN、ADVISE、ANALYZE和DRAFT,其中SUMMARIZE和QA占主导(分别占62%和31%)。数据集主要用于丹麦语言模型的监督微调(SFT),兼容trl、SFTTrainer等框架。
Autodata DA is a Danish synthetic dataset for instruction tuning, generated from the Dynaword corpus via a two-step LLM workflow. Each example consists of authentic Danish chatbot dialogues covering diverse daily use cases: document summarization, question answering about texts, content drafting, concept explanation, and advice provision. The prompt is always self-contained, containing any document text required for answering questions, so no external context is needed during training. The dataset contains 2965 examples in total, split into a training set (2816 examples) and a test set (149 examples). The instruction types include SUMMARIZE, QA, EXPLAIN, ADVISE, ANALYZE and DRAFT, among which SUMMARIZE and QA are the dominant categories, accounting for 62% and 31% of the total respectively. This dataset is primarily intended for supervised fine-tuning (SFT) of Danish large language models, and is compatible with frameworks such as trl and SFTTrainer.




