I created an open source LLM dataset for finetuning models to roleplay as your favorite anime waifus (I spent $80 worth of Hack Club AI credits generating this dataset). Yes you can train your anime AI waifus using this dataset :D
- AI generated the dataset
No followers yet
Once you ship this you can't edit the description of the project, but you'll be able to add more devlogs and re-ship it as you add new features!
Uploaded the dataset to Huggingface! Apache 2.0 license cuz open source is goated :D
So. Much. Work!!! So basically I scrambled to code a dataset generator for some anime roleplay today using the Anthropic API credits I bought a year ago during Arcade. They expire after 1 year for some reason, and today was literally the last day I had to use them. I've basically spent the entire day today scrambling to cobble together a customized synthetic dataset generator for anime roleplaying. The main features include optimization for TTS (text-to-speech) model dialogue (there's no actions or thoughts generated like in common datasets), and each line of dialogue includes the emotion that the character is currently feeling so that the TTS can sound more realistic. I pretty much scraped MyAnimeList to grab the top 500 characters and their descriptions, reformatted them into character cards with the free ai.hackclub.com api, then finally generated full conversations using the Opus 4.1 and Sonnet 4 models in the Anthropic API. Honestly, the final quality isn't as good as I'd like, but Opus 4.1 really surprised me with some of the conversations. In the end, the dataset consists of 239 conversations between different anime characters. The first 47 were generated by Sonnet 4, and onwards the conversations were generated by Opus 4.1. I've attached an image of the .jsonl file, and I'll be open-sourcing the dataset on Huggingface once I've applied some post-processing.