You can deploy this model and start earning money today!

sampoorna42/
bge-reranker-base-onnx

Name: bge-reranker-base-onnx
Availability: InStock

sentence-transformers

information-retrieval

BAAI/bge-reranker-base

quantized

mit

Views

No views yet

BGE Reranker Base — ONNX

ONNX export of BAAI/bge-reranker-base, converted using Hugging Face Optimum for faster CPU inference with ONNX Runtime. Drop-in replacement for the original PyTorch model with no code changes beyond the model loader.

Why use this?

The original PyTorch model typically takes 4–5 seconds per batch on CPU. This ONNX export runs ~2x faster with identical ranking quality — verified across multiple query/passage sets with zero ranking order degradation.

Usage

python

from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer
import torch

model = ORTModelForSequenceClassification.from_pretrained("sampoorna42/bge-reranker-base-onnx")
tokenizer = AutoTokenizer.from_pretrained("sampoorna42/bge-reranker-base-onnx")

query = "What causes climate change?"
passage = "Climate change is primarily driven by greenhouse gas emissions from human activities."

inputs = tokenizer(query, passage, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
    score = float(model(**inputs).logits[0][0])

print(f"Relevance score: {score:.4f}")

Download at Build Time (Docker / CI)

python

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="your-username/bge-reranker-base-onnx",
    local_dir="./onnx_reranker"
)

No authentication needed — public repo.

Model Details

Property	Value
Base model	BAAI/bge-reranker-base
Conversion tool	Hugging Face Optimum 2.1.0
Format	ONNX (FP32)
Max sequence length	512
Task	Text pair classification / passage reranking
License	MIT

Performance (CPU benchmark)

Tested on Intel CPU, no GPU, 3 independent query sets with 5–6 passages each:

Metric	PyTorch	ONNX
Total latency (3 samples)	1.88s	1.00s
Average speedup	—	~1.9x
Ranking order match	—	100% ✅

Scores differ slightly in absolute value (FP32 precision) but relative ranking is always preserved.

License

MIT — same as the original BAAI/bge-reranker-base.