MiniCPM5-2B-Jev — Calibrated Single-Pass Decision Model
Published:
MiniCPM5-2B adapted into a structured decision model that returns calibrated probabilities for categorical, binary, and ordinal-score questions in a single forward pass. #1 open-weight model ≤2B on JevBench; LoRA weights open-sourced on Hugging Face (Apache-2.0).
MiniCPM · PyTorch · LoRA · PEFT · Transformers · Hugging Face · Calibration · KV Cache · FastAPI · LLM Evaluation
- Uses an lm_head letter-readout head instead of autoregressive decoding.
- Rank-16 LoRA on all 42 layers and 7 projections (25.3M trainable parameters), trained on an 18,604-record multi-source dataset with a hybrid KL + Brier + ordinal-distance loss, followed by class-balanced prior debiasing and per-type temperature scaling.
- Shared-prefix KV-cache branching lets multiple questions reuse one encoded context with exact per-question isolation, served through a FastAPI
/v1/systemoneendpoint. - Unified evaluator across 7 benchmark suites: 78.8% on JevBench and 86.4% accuracy / 0.031 ECE on the 10-source decision-v7 set, up from 62.5% / 0.272 for the untuned base model.