klusai/ds-tf3-ro-3m
收藏资源简介:
ds-tf3-ro-3m数据集包含300万条罗马尼亚语的合成道德寓言,由TF3学生模型(klusai/tf3-26m-student)生成。该模型是一个2290万参数的LLaMA风格语言模型,专门针对罗马尼亚寓言数据进行训练。数据集的目标受众为4-7岁的儿童(Grupa B),平均每则寓言约256字。生成方法是通过结构化实体组合(角色、背景、挑战、解决、道德)从2120万种可能的组合中采样生成。每条记录包含一个fable字段,存储生成的寓言文本。
The ds-tf3-ro-3m dataset contains 3 million Romanian-language synthetic moral fables generated by the TF3 student model (`klusai/tf3-26m-student`), a 22.9M-parameter LLaMA-style language model trained from scratch on Romanian fable data. The target audience is children aged 4-7 (Grupa B), with an average length of ~256 words per fable. The generation method involves prompting with structured entity combinations (character, setting, challenge, resolution, moral) sampled from a pool of 21.2M possible combinations. Each example contains a single field: fable (string), which stores the generated fable text in Romanian.




