Dev fine-tunes Qwen 3.8 on 81,837 book annotations, writing quality up 86% in blind tests
Scobleizer · x · 2026-09-13
Frustrated by generic "AI slop" prose, woodchipdaddy is fine-tuning Qwen 3.8 27B to write well: 17 books labeled with 81,837 annotations across 2,509 context windows, now in a DPO run. Blinded tests against his own preferences show +86% writing quality over base — and counting.
More from coding & agent
- Agent swarm leaderboard 'turkish-delight' dethroned; anyone can submit an agent in 2 minutes — mervenoyann · 2026-09-13
- Dev open-sources MnemoBrain, an agent memory system that works without prompt-based recall — AxelFooley · 2026-09-13
- Dev seeks validation for a control gateway managing agent credentials, permissions and auditing — Decent_Progress7631 · 2026-09-13
- Dev builds 'recursive self-improvement': scheduled automations with MCP access that fix codebases — DanielLockyer · 2026-09-13
- WTMemory: a 794KB Mac utility that hunts AI model weights and dev servers eating your RAM — saibharadwaj · 2026-09-13
- "AI won't kill me—but I'll be buried under an ever-growing review pile" — tokenbender · 2026-09-13