Jev-Style-2B-Decision-v3
An independent 1.88B full fine-tune of Qwen3.5-2B that reads up to 25,600 tokens of state and options and returns a calibrated probability for every option of choice, yes/no (noul) and score questions, without generating text. Needs the maker's runtime. Also in 0.8B.
An independent project by Chaoliang Yan, not affiliated with TypeSafe. Each option is scored at its own slot and turned into probabilities with one global temperature. Decisions need the runtime shipped in the repo (PyTorch, a GGUF scorer for llama.cpp, or MLX) or the jev-style pip package, which serves a local /v1/systemone API; stock llama.cpp, Ollama and mlx-lm cannot produce the scores. The maker's own run of the third-party JevBench v1.4.1 public items gives 73.6% (170/231), not a board entry; the training pool includes the train splits of several public benchmarks. Sibling Jev-Style-0.8B-Decision-v3 and a cascade configuration.
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
At a glance
| Parameters | 1.88B |
| Base model | qwen/qwen3.5-2b |
| Maker | Chaoliang Yan |
| Released | 2026-09-27 |
| License | apache-2.0 |
| Reported accuracy | 73.6% |
Get the weights
pip install systemonemodels
systemone pull chaoliang-yan/jev-style
The files are served from the maker's Hugging Face repository, chaoliangUNSW/Jev-Style-2B-Decision-v3, and verified against the checksums recorded here.
Read more
This page was opened by System One for Chaoliang Yan, who can claim the organisation and take it over at any time.