bge-m3-ko

bge-m3-ko-h100

H100 fine-tuned Korean embedding model based on dragonkue/BGE-m3-ko.

This is the better fit for the local AutoRAG corpus among the two uploaded models. Compared with qwen3-embedding-h100, it is the stronger model to highlight for the domain-specific benchmark.

Training

Fine-tuned from dragonkue/BGE-m3-ko (the official BAAI/bge-m3 continued-trained on Korean MIRACL) with a contrastive (in-batch-negative) objective.

• Base model: dragonkue/BGE-m3-ko — Korean-specialized BGE-M3
• 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 8 · gradient accumulation 4 · learning rate 2e-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: 0.7773
• MAP: 0.7773
• Hit@1: 0.6754
• Hit@5: 0.9123
• Hit@10: 0.9474
• Hit@50: 0.9825

MIRACL

• Task: MIRACLRetrieval
• Main score: 0.50
• Dataset revision: 9c09abc13478308c27598f350e31d8f06b9b5481

Artifacts

• Model: output/111677_20260506_114341_both_2gpu/bge
• AutoRAG benchmark: benchmark_results/autorag_benchmark.json
• MIRACL summary: benchmark_results/miracl_benchmark.txt
• MIRACL details: benchmark_results/mteb/MIRACLRetrieval.json

Comparison note

• Strongest AutoRAG result of the two uploaded models
• Weaker MIRACL-KO result than rtfin-qwen3-embedding-h100
• Below dragonkue/snowflake-arctic-embed-l-v2.0-ko and rtfin-qwen3-embedding-h100 on the public Korean retrieval leaderboard, but still a strong local-domain model
• Good to cite when the audience cares about the local domain corpus rather than the widest public Korean retrieval leaderboard