Qwen writing quality in tests judged by personal taste
The user says they've labeled 17 books with 81,837 annotations to help train Qwen 3.8 27B to write better, with a preference-training run underway.
TLDR
Frustrated with AI-generated writing, a user says they're trying to improve Qwen 3.8 27B using 81,837 annotations across 17 books and 2,509 context windows. They report an 86% writing-quality improvement over the base model in blinded tests judged against their own preferences. The user says training and testing are ongoing.
Combined views
555
1 Source, first seen 18d ago