bge-m3-zeroshot-v2.0-onnx-int8

23donge/bge-m3-zeroshot-v2.0-onnx-int8

bge-m3-zeroshot-v2.0-onnx-int8 is a zero-shot classification model from 23donge. Use it when you need labels you did not train the model on. It is set up for transformers. The card lists the license as mit.

MoritzLaurer/bge-m3-zeroshot-v2.0 的 INT8 动态量化 ONNX 版本。多语言零样本分类模型，基于 BAAI/bge-m3-retromae (XLM-RoBERTa-large)。

Downloads · 30 days

181

30% of all-time downloads

All-time downloads

613

Public

Repo size

593 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.onnx571 MB · 96%

At a glance

Task: Zero-Shot Classification
Library: transformers
License: mit
Model type: xlm-roberta
Access: Public
Created: Jun 30, 2026
Updated: Jun 30, 2026
SHA: 43aa9c80

Try a prompt

Model card

From the Hugging Face model README

bge-m3-zeroshot-v2.0-onnx-int8

MoritzLaurer/bge-m3-zeroshot-v2.0 的 INT8 动态量化 ONNX 版本。多语言零样本分类模型，基于 BAAI/bge-m3-retromae (XLM-RoBERTa-large)。

INT8 dynamic quantized ONNX version of MoritzLaurer/bge-m3-zeroshot-v2.0, a multilingual zero-shot classifier.

原模型 / Original model: https://huggingface.co/MoritzLaurer/bge-m3-zeroshot-v2.0

使用 / Usage

from transformers import pipeline

classifier = pipeline("zero-shot-classification",
                      model="23donge/bge-m3-zeroshot-v2.0-onnx-int8")
result = classifier(
    "Angela Merkel is a politician in Germany",
    candidate_labels=["politics", "sports", "science"],
)
print(result)

性能 / Performance

环境: CPU 4核 (taskset -c 0-3), ONNX Runtime 1.24.4

指标	fp32 ONNX	INT8 ONNX	加速
模型大小	2,166 MB	544 MB	4.0x
10类单条推理	1,535 ms	363 ms	4.2x
5类中文单条推理	887 ms	223 ms	4.0x
10条 batch 推理	20,194 ms	4,888 ms	4.1x

精度验证 / Accuracy

结论: INT8 与 fp32 的 top-1 标签 100% 一致（全部 15 条测试），平均分数差异 0.011。

Conclusion: INT8 preserves 100% top-1 label accuracy across all 15 test prompts. Mean score difference: 0.011.

English (10 classes)

#	Text	Label	fp32↑	INT8↑	Diff
1	How do I implement binary search in Go?	programming	0.942	0.955	0.014
2	The weather in Tokyo will be rainy tomorrow with highs of 22 degrees.	weather	0.994	0.996	0.002
3	To make authentic Italian carbonara, you need eggs, pecorino cheese, guanciale, and black pepper.	cooking	0.971	0.974	0.003
4	A zero-day vulnerability was discovered in the OpenSSL library affecting TLS handshakes.	security	0.993	0.996	0.003
5	The Eiffel Tower in Paris attracts over 7 million visitors annually.	travel	0.810	0.873	0.063
6	Quantum entanglement allows particles to be correlated across vast distances.	science	0.985	0.983	0.002
7	The S&P 500 index rose 2.3% following the Federal Reserve's interest rate decision.	finance	0.928	0.909	0.020
8	Active learning techniques improve student engagement and knowledge retention by up to 40%.	education	0.997	0.996	0.001
9	The new Marvel movie broke box office records in its opening weekend.	entertainment	0.998	0.996	0.002
10	Regular exercise and a balanced diet are key to maintaining cardiovascular health.	health	0.995	0.997	0.001

中文 / Chinese (5 classes)

#	Text	Label	fp32↑	INT8↑	Diff
1	如何在Go语言中实现二分查找？	编程	0.957	0.914	0.044
2	明天东京天气有雨，最高气温22度。	天气	0.997	0.992	0.004
3	制作正宗意大利培根蛋酱意面需要鸡蛋、佩科里诺奶酪、猪脸肉和黑胡椒。	烹饪	0.986	0.989	0.003
4	在OpenSSL库中发现了一个影响TLS握手的零日漏洞。	安全	0.920	0.903	0.017
5	巴黎埃菲尔铁塔每年吸引超过700万游客。	旅游	0.984	0.991	0.007

License: MIT

Task: Zero-Shot Classification
Library: transformers
Type: xlm-roberta
License: mit
Created: Jun 30, 2026
Updated: Jun 30, 2026