ezjev-4b-s3
An independent developer's Jev-style decision model: a LoRA fine-tune of Qwen3.5-4B that answers choice, yes/no and score questions with a probability per option from a single prefill. Weights are open; the maker asks that they be treated as non-commercial.
A merged LoRA; each question takes one prefill and the model's option-letter log-probabilities after "Answer:" give the probabilities, with one global temperature fitted on the maker's held-out dev split. s3 is fine-tuned from s2 (both public). On JevBench's public items (a third-party benchmark) the maker's own run gives 187/231 (easy 48/48, original 71/72, hard 68/111); the maker says no JevBench item was used in training. No licence is declared for the weights and the maker advises research or non-commercial use; the code is MIT.
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 | 4B |
| Base model | qwen/qwen3.5-4b |
| Maker | Everett (everettjf) |
| Released | 2026-10-05 |
| License | No weights licence; maker advises research/non-commercial use (training data includes ANLI, CC BY-NC 4.0); code MIT |
| Reported accuracy | 81.0% |
| Reported latency | 34 ms p50 / 99 ms p95 on one RTX PRO 6000 |
Get the weights
pip install systemonemodels
systemone pull everettjf/ezjev
The files are served from the maker's Hugging Face repository, everettjf/ezjev-4b-s3, and verified against the checksums recorded here.
Read more
This page was opened by System One for Everett (everettjf), who can claim the organisation and take it over at any time.