Personal taste tests of efforts to improve Qwen 3.8 27B’s writing
The user says they've labeled 17 books with 81,837 annotations across 2,509 context windows in an effort to improve the model's writing.
TLDR
Frustrated with AI-generated writing, a user says they're training Qwen 3.8 27B to write more to their taste. They report an 86% improvement in writing quality over the base model, based on blinded tests judged against their personal preferences. They say the training run is still underway.
Combined views
2.4K
1 Source, first seen 18d ago