sungkwang2/khalani-en-kha-translation
收藏资源简介:
Khalani与英语翻译数据集,专注于Khalani(《星际争霸》中神族构造语言)与英语之间的双向翻译。数据集包含55个翻译对,采用5折交叉验证(例如fold0:44个训练样本和11个测试样本)。文件格式为JSONL,其中`*.jsonl`表示英语到Khalani的翻译,`*_kha2en.jsonl`表示Khalani到英语的翻译。每个JSONL行包含以下字段:指令(instruction)、输出(output)、英语文本(en)和Khalani文本(khalani)。需要注意的是,由于数据集规模较小(仅55对),全微调可能导致过拟合(即模型记忆训练数据而无法泛化到未见数据)。该数据集旨在用于探索性研究,特别是关于自回归与扩散模型在构造语言处理中的比较,不建议用于得出一般性结论。相关模型可在huggingface.co/sungkwang2找到。
Khalani to English translation dataset, focusing on bidirectional translation between Khalani (a constructed language for the Protoss race in StarCraft) and English. The dataset consists of 55 translated pairs, using 5-fold cross-validation (e.g., fold0: 44 training samples and 11 test samples). The file format is JSONL, where `*.jsonl` represents English to Khalani translation, and `*_kha2en.jsonl` represents Khalani to English translation. Each JSONL line contains the following fields: instruction, output, English text (en), and Khalani text (khalani). Note that due to the small dataset size (only 55 pairs), full fine-tuning may lead to overfitting (i.e., the model memorizes training data and fails on held-out data). This dataset is intended for exploratory research, particularly in the context of comparing autoregressive versus diffusion models for constructed language processing, and is not suitable for drawing general conclusions. Related models can be found at huggingface.co/sungkwang2.




