NGA-KR/ko-embed-nli

Korean text embedding model (568M) fine-tuned from BAAI/bge-m3 for sentence-pair classification (NLI / paraphrase).

Training data (train splits only)

• KLUE-NLI, KorNLI (MNLI + SNLI translations): premise → entailment hypothesis as positive, contradiction hypotheses as hard negatives
• PAWS-X (ko): sentence1 → paraphrase (label 1) as positive, non-paraphrase (label 0) of the same sentence1 as hard negatives

Declared as KLUE-NLI, KorNLI, PawsXPairClassification in the MTEB model metadata; base-model training data is inherited from bge-m3.

Evaluation (MTEB(kor, v2), PairClassification)

Training

CachedMultipleNegativesRankingLoss · batch 512 · lr 1.5e-5 · 2 epochs · max_len 128 · CLS pooling · bf16

License

MIT (same as BAAI/bge-m3).