pjev-2b
A 2.5B typed-decision model: a LoRA merged into MiniCPM5-2B that answers choice, score and yes/no questions in one forward pass by reading the answer-letter logits, giving a probability per option and no generated text. Apache-2.0; shipped uncalibrated.
A personal project. A rank-64 LoRA folded into MiniCPM5-2B and published as one stock checkpoint; no head is added and nothing is decoded. Options are read as letters at a single position, and several questions about one state share a prefix (8.0 ms per question when six share a state). Includes serve.py, a TypeSafe-compatible /v1/systemone server. Not calibrated (temperature 1, hard-tier ECE 0.120 by the maker's measure); no abstention; English only; more than 26 options fall back to a tournament approximation. The maker's run on JevBench's 231 public items is not an official score: easy 48/48, standard 59/72, hard 68/111; JevBench has it in its queue. Trained on Caio Vicentino's CC BY 4.0 eikos-decisions dataset, with MIT serving code from his Eikos project; the maker discloses that an earlier data snapshot included rows later withdrawn for overlapping the Decision Index test set, and reports no Decision Index result.
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 | 2.5B |
| Base model | openbmb/MiniCPM5-2B |
| Maker | Kailune-AI |
| Released | 2026-09-28 |
| License | apache-2.0 |
| Reported accuracy | 75.8% |
| Reported latency | 20.5 ms per single question on one H100, in-process, batch 1 |
Get the weights
pip install systemonemodels
systemone pull kailune-ai/pjev-2b
The files are served from the maker's Hugging Face repository, Kailune-AI/pjev-2b, and verified against the checksums recorded here.
Read more
This page was opened by System One for Kailune-AI, who can claim the organisation and take it over at any time.