afrizalha/Gatra-2-Javanese

该数据集包含36870个Krama Javanese指令调优示例的prompt-response对，数据几乎完全是合成的，只有少量人工整理。当前数据集仅支持单轮问答，尽管在指令调优模型上进行微调可能允许转移多轮能力。prompt由GPT-4o生成，response由Claude 3 Haiku生成。数据生成的方式可能导致prompt中包含用户询问模型的英语术语，但这不应影响调优模型学习将Krama Javanese响应与prompt中的Krama Javanese方面关联起来。

The dataset comprises 36870 prompt-response pairs of Krama Javanese instruction-tuning examples. The data is almost entirely synthetic with minimal human curation. The current dataset supports only single-turn QA, although fine-tuning on instruction-tuned models may allow for transfer of multi-turn capabilities. The prompts are generated by GPT-4o, while the responses are generated by Claude 3 Haiku. The way the data set was generated, the prompt may contain terms in the English language that the user is asking the model. However, this should not matter as tuned models would learn to associate Krama Javanese responses with the Krama Javanese aspects of the prompt.