NGA-KR/ko-embed-cls
Korean text embedding model (568M) fine-tuned from BAAI/bge-m3 for topic classification.
Training data
- KLUE-TC (ynat) train split only (45,678 news titles, 7 topics), label-aware contrastive pairs (same-topic titles as positives, other-topic titles as hard negatives).
- No evaluation splits were used. Declared as
KLUE-TCin the MTEB model metadata; base-model training data is inherited from bge-m3.
Evaluation (MTEB(kor, v2))
| Task | Score |
|---|---|
| KLUE-TC (Classification, accuracy) | 84.39 |
Training
CachedMultipleNegativesRankingLoss · batch 512 · lr 2e-5 · 1 epoch · max_len 128 · CLS pooling · bf16
Usage
from sentence_transformers import SentenceTransformer
m = SentenceTransformer("NGA-KR/ko-embed-cls")
e = m.encode(["삼성전자, 2분기 영업이익 10조 돌파", "손흥민 시즌 20호골"])
License
MIT (same as BAAI/bge-m3).