bge-m3-bank-it

A fine-tuned version of BAAI/bge-m3 for domain-specific retrieval and reranking.

Produces dense, sparse (lexical), and ColBERT embeddings simultaneously.

Benchmark Results

Evaluation on held-out test set (20% split, queries never seen during training):

Dense Retrieval

Metric	Base (bge-m3)	Fine-tuned	Delta
Recall@1	29.5%	85.0%	↑ 55.5%
Recall@5	63.0%	96.5%	↑ 33.5%
Recall@10	76.0%	99.0%	↑ 23.0%
MRR	45.0%	90.4%	↑ 45.4%
NDCG@10	51.8%	92.5%	↑ 40.6%

Multi-Mode Reranking

Metric	Base (bge-m3)	Fine-tuned	Delta
Accuracy	37.5%	94.0%	↑ 56.5%
MRR	57.3%	96.6%	↑ 39.4%

Usage

from FlagEmbedding import BGEM3FlagModel

model = BGEM3FlagModel("Sophia-AI/bge-m3-bank-it", device="cuda", use_fp16=True)

# Embeddings
output = model.encode(["Your query here"], return_dense=True, return_sparse=True)

# Reranking
scores = model.compute_score(
    [["query", "document"]],
    weights_for_different_modes=[0.30, 0.65, 0.05],
)

Fine-Tune Your Own

This model was fine-tuned using bge-auto-tune:

pip install bge-auto-tune

bge-auto-tune generate --collection your_collection --min-pairs 2000
bge-auto-tune finetune --dataset bge_m3_training.jsonl --epochs 4
bge-auto-tune test --model ./bge-m3-finetuned
bge-auto-tune publish --repo your-user/your-model-name