ksonny4/modernbert-jev-v2

modernbert-jev v2 (fine-tuned, x86 QUInt8)

150M ModernBERT-base NLI (tasksource/ModernBERT-base-nli), fine-tuned with KL distillation on 50k JevDistillCorpus-v3 rows + MNLI aux, exported to ONNX opset 17, dynamically quantized with unsigned int8 weights (signed mis-executes on x86 ORT).

• model.onnx — int8 model, ~150 MB, 1 CPU thread
• tokenizer.json, config.json — Rust tokenizer + head config (GLUE label order)
• calibration.json — per-head temperatures (choice 0.9, score 0.9, noul 2.5)

Run with https://github.com/KSonny4/modernbert-jev: modernbert-jev run --model <this-dir> --request ... or serve for POST /v1/systemone.

Public JevBench accuracy (self-measured, 231 items): 0.554.