yes24-task-targeted-embedding-0.6B

YES24 도서 커머스 검색에 특화된 0.6B 한국어 임베딩 모델입니다.

Qwen3-0.6B를 기반으로 2단계 학습을 거쳤습니다:

1. Stage 1: MAKI 검색-클릭 로그(4.5M pairs)로 query-document distillation
2. Stage 2 (v7): NV-Retriever TopK-PercPos로 마이닝한 584k hard negative triplet으로 retrieval LoRA 학습 후 merge

Performance

MAKI holdout 1,000 queries, 13,476 docs 기준:

Model	MRR	Recall@1	nDCG@1	nDCG@10
yes24-task-targeted-0.6B (v7)	0.6746	0.558	0.7638	0.8644
bge-m3-yes24-ft (568M)	0.6692	0.548	0.7585	0.8612
yes24-task-targeted-0.6B (v6)	0.6635	0.534	0.7507	0.8631
stage1-base (no LoRA)	0.6557	0.523	0.7438	0.8603

v7 vs v6 개선

Metric	v6	v7	Delta
MRR	0.6635	0.6746	+0.0111
Recall@1	0.534	0.558	+0.024
nDCG@1	0.7507	0.7638	+0.0131
nDCG@10	0.8631	0.8644	+0.0013

63% 데이터량(584k vs 927k)으로 전 지표 개선. TopK-PercPos 마이닝이 false negative를 효과적으로 제거한 결과.

Model Details

• Architecture: Qwen3-0.6B + LoRA (rank=32) merged
• Embedding Dimension: 1024
• Max Sequence Length: 384 tokens
• Pooling: Last token
• Normalization: L2 normalize
• Similarity Function: Cosine similarity (= dot product after normalization)

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Ja-ck/yes24-task-targeted-embedding-0.6B")

queries = ["Query: 개브리얼 제빈"]
documents = [
    "Document: Title: 내일 또 내일 또 내일\nCategory: 영미소설",
    "Document: Title: 파친코\nCategory: 영미소설",
]

q_emb = model.encode(queries)
d_emb = model.encode(documents)

similarities = model.similarity(q_emb, d_emb)
print(similarities)

Note: Query에는 Query: prefix, Document에는 Document: prefix를 붙여야 합니다.

Training Details

Stage 2 (v7) Training

• Hard Negative Mining: NV-Retriever TopK-PercPos (relative_margin=0.95, absolute_threshold=0.90)
• Teacher Model: nlpai-lab/KURE-v1 (1024d, Title+Author+Category embedding)
• Triplets: 583,923 (hard 84%, medium 16%)
• Loss: InfoNCE + 2×Cosine Distillation (Qwen3-Embedding-4B teacher) + GOR
• LoRA: rank=32, alpha=32, target=q/k/v/o/gate/up/down_proj
• Training: 2 epochs, batch=160×4 grad accum, lr=1.5e-4, bf16
• Hardware: RunPod H100 SXM 80GB

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 384, 'do_lower_case': False, 'architecture': 'Qwen3Model'})
  (1): Pooling({'pooling_mode_lasttoken': True})
  (2): Normalize()
)

Framework Versions

• Sentence Transformers: 5.x
• Transformers: 5.x
• PyTorch: 2.10+
• PEFT: 0.18+