Scrap turns EPUBs into hobbyist LLM training text

Scrap captures the scrappy, low-budget side of LLM experimentation: the bottleneck is often data preparation, not model architecture. The cleverness is real, but the legal and curation risks are impossible to ignore.

• –EPUB-to-text conversion removes formatting noise and creates a corpus that is easy to tokenize and process
• –A large book dump provides volume, but not necessarily clean, diverse, representative training data
• –Copyright status and usage rights make LibGen a risky foundation for anything beyond private experimentation
• –Hobbyists can pursue safer alternatives through public-domain books, openly licensed datasets, and opt-in content
• –The project’s broader value is its reproducible Unix-style tooling for turning messy files into machine-readable data