News: add Jev-Style v3 (0.8B full fine-tune, 25,600-token input)
Adds one entry to the trained section of ITEMS in news.html: Jev-Style-0.8B-Decision-v3, a full fine-tune of Qwen3.5-0.8B (Apache-2.0) released 2026-09-24.
- One verdict slot per option (no letter cap; 77 options scored in one pass), up to 25,600 input tokens, many choice / score / noul questions per state in one call.
- Builds: transformers, GGUF (Q4_K_M 0.53 GB) and MLX; a ZeroGPU Space.
- Links: model, GGUF repo, Space, GitHub mirror, the Hub post and jevstyle.com. There is no post on X, so
by.handleis null. - Metrics as of 2026-09-25: 0 GitHub stars, 0 likes on the main Hub repo.
- The one number in the description (79.2% vs Laya typed 76.6% on the 2,000 typed decisions) is from the model card and says that both were trained on that dataset's train split.
Only this entry is added; nothing else in the file changes (checked by parsing ITEMS with node: 64 entries). Not on the Decision Index yet. Happy to change the wording or drop fields.
Not affiliated with TypeSafe AI.
Hello, can you please score it yourself (following the instructions here: https://github.com/apolinario/decision-index ) and submit as a PR to this same repo https://github.com/apolinario/decision-index ?
Done: self-scored with the kit (edition 0.2.1, complete runs) and submitted as https://github.com/apolinario/decision-index/pull/22.
- Jev-Style-2B-Decision-v3: 34.83
- Jev-Style-0.8B-Decision-v3: 16.64
Results, engine and procedure: https://huggingface.co/datasets/chaoliangUNSW/decision-index-results
Happy to update this news entry with the score, or close it if you prefer to handle news separately.