wl-ko-ner-v1

wl-ko-ner-v1

Korean NER (KLUE-NER 6 entity: PS/LC/OG/DT/TI/QT) — KoELECTRA-base-v3 finetune, ONNX fp16 for browser inference (onnxruntime-web, WebGPU/WASM).

• KLUE-NER dev entity F1: 0.8180 (KLUE 공식 character-level 과 수학적 동일).
• Base: monologg/koelectra-base-v3-discriminator (Apache-2.0).
• Training data: KLUE-NER (CC-BY-SA-4.0) — Park et al. 2021, NeurIPS.
• Weights license: CC-BY-SA-4.0 (KLUE 상속). 코드 Apache-2.0.

WygLore Leaf (RisuAI V3 플러그인) 의 온디바이스 cold-start NER 용. 자세히: NOTICE, metric-report.md.

Files

• wl-ko-ner-v1.onnx — fp16 ONNX (~215MB). int8 PTQ 는 이 모델에서 붕괴 → fp16 채택.
• tokenizer.json / config.json / vocab.txt / ... — KoELECTRA tokenizer.

Usage (onnxruntime-web)

const sess = await ort.InferenceSession.create(modelUrl, { executionProviders: ['webgpu','wasm'] });