krino
Typed decision models that output calibrated probabilities instead of free text. Open replication of TypeSafe's Jev.
95.2%
Best Banking77
Ettin-150m r128, trained heads
75.0%
Jev Reference
Banking77 (API benchmark)
23
Models Swept
Causal + encoder + reranker
19
Benchmarks
Choice + noul + score types
Top Results
| Model | Params | Method | Benchmarks | Avg Accuracy |
|---|---|---|---|---|
| Jev (TypeSafe API) | unknown | jev_api | 18 | 87.0% |
| OpenJev Qwen3.5-4B | 4B | openjev | 1 | 79.2% |
| Qwen3.5-9B | 9B | logit | 19 | 67.3% |
| Qwen2.5-7B-Instruct | 7B | logit | 19 | 67.1% |
| Qwen3-4B | 4B | logit | 19 | 66.3% |
| Qwen3-8B | 8B | logit | 19 | 65.3% |
| Phi-4-mini-instruct | 3.8B | logit | 19 | 64.4% |