NGA-KR/ja-embed-sts

Korean dense text embedding model (337M parameters), fine-tuned from cl-nagoya/ruri-large-v2.

Training data (zero-shot on MTEB(kor))

This model was trained only on public English retrieval datasets — no Korean benchmark train splits were used:

• JSTS
• JSICK

Consequently, the model is 100% zero-shot on MTEB(kor, v1) and MTEB(kor, v2) (declared via training_datasets in the MTEB model metadata).

Usage

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("NGA-KR/ja-embed-sts")
q = model.encode(["전세 계약서에 기간을 안 썼으면 얼마 동안 유효해?"], prompt_name="query")
d = model.encode(["임대차 계약 기간을 정하지 않은 경우 2년으로 본다..."], prompt_name="document")
print(q @ d.T)

Prompts are stored in the model config: query: for queries, passage: for documents.

Evaluation

Evaluated with mteb on MTEB(kor, v2) (20 tasks). See the MTEB leaderboard entry for full per-task scores.

Training details

• Base: cl-nagoya/ruri-large-v2 (ModernBERT-style Korean encoder, mean pooling, L2-normalised)
• Objective: contrastive (CachedMultipleNegativesRankingLoss), in-batch + 1 hard negative
• Batch 2048, lr 2e-5, cosine schedule, 1 epoch, bf16
• Max sequence length: 512

License

apache-2.0. Base model license: see cl-nagoya/ruri-large-v2.