zerofata/Anime-AMA-Prose

该数据集包含动漫、游戏角色和VTuber针对其维基页面相关问题的对话。 prompts、场景和初始响应是由GLM 4.6生成的，而重写（如果发现过多的重复词汇）是由Gemma 3 27b完成的。重写的质量通常低于GLM提供的响应，但提供了不同的措辞和词汇选择。此外，有些重写包含小的ASCII表情反应。数据集大部分是适合所有年龄段的，但没有进行广泛的过滤。原本希望使用Kimi K2 Thinking进行提示和场景生成，甚至可能间歇性地用于计划和响应，但由于其不可用，所以使用了GLM 4.6。尽管尝试避免，但数据集中可能仍然存在重复说话的情况。另外，Gemma 3 27B使用不同的引号，这些引号没有被清理。

This dataset includes dialogues of anime, game characters, and VTubers answering questions related to their wiki pages. Prompts, scenes, and initial responses are generated by GLM 4.6, while rewrites (if too many overused words are found) are done by Gemma 3 27b. The rewrites are generally of lower quality than the responses provided by GLM but offer different prose and word choices. Some rewrites include small ASCII face reactions. Most of the dataset is SFW but not extensively filtered. Kimi K2 Thinking was preferred for prompt and scene generation, and possibly for planning/responses, but it was unavailable at the time. GLM 4.6 is used and can be very deterministic in such tasks. The dataset may contain parroting despite attempts to avoid it. Additionally, Gemma 3 27B uses different quotation marks which were not cleaned.