CoCo-Decision-4B-Ko
A Korean and English decision adapter on Qwen3.5-4B. A state, a typed question and lettered options go in; probabilities are read from the option-letter logits of one forward pass, with no generation. Jev-inspired, trained with a Brier term for calibration.
A rank-16 LoRA (39 MB) on Qwen3.5-4B, published by Minyoung Yu as Corners AI and built with his oh-my-jev toolkit. It reads option-letter logits, so it is Jev-inspired, like Tev1. Intended for routing and gating in agent systems. The card warns that JevBench numbers may overstate results, because synthetic training data was designed after studying the public task types, and that calibration is weaker than Jev's in its own comparison. Weak on long policy documents (0.63) and date arithmetic (0.40, overconfident). On MASSIVE ko-KR the maker reports 0.915 (ECE 0.023).
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 4B |
| Base model | Qwen/Qwen3.5-4B |
| Maker | Corners AI |
| Released | 2026-09-29 |
| License | apache-2.0 |
| Reported accuracy | 87.0% |
Get the weights
pip install systemonemodels
systemone pull corners-ai/coco-decision
The files are served from the maker's Hugging Face repository, corners-ai/CoCo-Decision-4B-Ko, and verified against the checksums recorded here.
Read more
This page was opened by System One for Corners AI, who can claim the organisation and take it over at any time.