0.8B Qwen 3.5 full finetune transcribes 2,000+ medieval manuscripts across 10,000 pages
wjb_mattingly · x · 2026-09-04
wjbmattingly shares a full finetune of 0.8B Qwen 3.5 trained on 2,000+ medieval manuscripts (10,000 pages total) from the Comma dataset (0.09 CER), to be presented with models in Vienna. He acknowledges CER is no longer a good default eval and plans to compare other approaches.
Related event: 0.8B Qwen Fine-Tuned on Medieval Manuscripts Sparks CER Debate(4 posts)→
More from Research
- Bug Hunt Bench: 105 real bugs stress-test GPT-6, Claude, Grok, Gemini and more coding agents — PawelHuryn · 2026-09-05
- Many mathematicians value prestige over truth, discussion on AI proofs notes — avt_im · 2026-09-05
- eyebench author says no v4, moving on to harder benchmarks — adonis_singh · 2026-09-05
- After 8 months of digging, researcher says persona models fail in RL — BronsonSchoen · 2026-09-05
- Full Fruit Fly Connectome With 166,700 Neurons Runs Inside Minecraft, Driving a Fly's Movement — Dan_Jeffries1 · 2026-09-05
- Declarative Attention lets LLMs declare their own focus, cutting 52% of KV cache reads — eigenlaplace · 2026-09-05