qwen3-embedding

qwen3-embedding-h100

H100 fine-tuned Korean embedding model based on Qwen/Qwen3-Embedding-4B.

This is the stronger of the two uploaded models for public Korean retrieval. Compared with bge-m3-ko-h100, it is the better model to highlight when you want the strongest general Korean retrieval story.

Training

Fine-tuned from Qwen/Qwen3-Embedding-4B with a contrastive (in-batch-negative) objective.

• Base model: Qwen/Qwen3-Embedding-4B
• Training data: nvidia/Nemotron-Personas-Korea — synthetic Korean personas (demographic / geographic / personality), used as (query, positive-document) pairs
• Objective: sentence-transformers MultipleNegativesRankingLoss (in-batch negatives, InfoNCE-style)
• Hardware: 4×H100 (Slurm)
• Hyperparameters: 3 epochs · per-device batch size 1 · gradient accumulation 8 · learning rate 1e-5 · warmup ratio 0.1 · weight decay 0.01 · bf16
• Finetune run: 111677_20260506_114341_both_2gpu

Benchmark Results

AutoRAG

• Corpus size: 720
• Queries evaluated: 114
• MRR: 0.7677
• MAP: 0.7677
• Hit@1: 0.6579
• Hit@5: 0.9035
• Hit@10: 0.9298
• Hit@50: 0.9649

MIRACL

• Task: MIRACLRetrieval
• Dataset subset: ko
• Corpus size: 1,486,752
• Queries evaluated: 213
• mRR: 0.5773
• mAP: 0.4328

Artifacts

• Model: output/111677_20260506_114341_both_2gpu/qwen
• AutoRAG benchmark: benchmark_results/autorag_benchmark.json
• MIRACL summary: benchmark_results/qwen_miracl_fast4/miracl_benchmark.txt
• MIRACL details: benchmark_results/qwen_miracl_fast4/miracl_benchmark.json

Comparison note

• Strongest public Korean retrieval story among the open Korean embedding models I checked
• Ahead of dragonkue/snowflake-arctic-embed-l-v2.0-ko 0.740433, dragonkue/BGE-m3-ko 0.729993, nlpai-lab/KURE-v1 0.727739, and nlpai-lab/KoE5 0.711356 on the Korean retrieval leaderboard cited by the model card
• Public Qwen3-Embedding-4B leaderboard result on Korean retrieval is average NDCG@10 0.7484