aspire/acge_text_embedding

acge model

acge模型来自于合合信息技术团队，对外技术试用平台TextIn, github开源链接为github。合合信息是行业领先的人工智能及大数据科技企业，致力于通过智能文字识别及商业大数据领域的核心技术、C端和B端产品以及行业解决方案为全球企业和个人用户提供创新的数字化、智能化服务。

技术交流请联系[email protected]，商务合作请联系[email protected]，可以点击图片，扫面二维码来加入我们的微信社群。想加入合合信息，做“文档解析”、“文档检索”、“文档预研”的同学可以投简历给[email protected]，也可直接添加HR微信详聊岗位内容。

acge是一个通用的文本编码模型，是一个可变长度的向量化模型，使用了Matryoshka Representation Learning，如图所示：

建议使用的维度为1024或者1792

Model Name	Model Size (GB)	Dimension	Sequence Length	Language	Need instruction for retrieval?
acge-text-embedding	0.65	[1024, 1792]	1024	Chinese	NO

Metric

C-MTEB leaderboard (Chinese)

测试的时候因为数据的随机性、显卡、推理的数据类型导致每次推理的结果不一致，我总共测试了4次，不同的显卡(A10 A100)，不同的数据类型，测试结果放在了result文件夹中，选取了一个精度最低的测试作为最终的精度测试。 根据infgrad的建议，选取不用的输入的长度作为测试，Sequence Length为512时测试最佳。

Model Name	GPU	tensor-type	Model Size (GB)	Dimension	Sequence Length	Average (35)	Classification (9)	Clustering (4)	Pair Classification (2)	Reranking (4)	Retrieval (8)	STS (8)
acge_text_embedding	NVIDIA TESLA A10	bfloat16	0.65	1792	1024	68.91	72.76	58.22	87.82	67.67	72.48	62.24
acge_text_embedding	NVIDIA TESLA A100	bfloat16	0.65	1792	1024	68.91	72.77	58.35	87.82	67.53	72.48	62.24
acge_text_embedding	NVIDIA TESLA A100	float16	0.65	1792	1024	68.99	72.76	58.68	87.84	67.89	72.49	62.24
acge_text_embedding	NVIDIA TESLA A100	float32	0.65	1792	1024	68.98	72.76	58.58	87.83	67.91	72.49	62.24
acge_text_embedding	NVIDIA TESLA A100	float16	0.65	1792	768	68.95	72.76	58.68	87.84	67.86	72.48	62.07
acge_text_embedding	NVIDIA TESLA A100	float16	0.65	1792	512	69.07	72.75	58.7	87.84	67.99	72.93	62.09

Reproduce our results

C-MTEB:

import torch
import argparse
import functools
from C_MTEB.tasks import *
from typing import List, Dict
from sentence_transformers import SentenceTransformer
from mteb import MTEB, DRESModel


class RetrievalModel(DRESModel):
    def __init__(self, encoder, **kwargs):
        self.encoder = encoder

    def encode_queries(self, queries: List[str], **kwargs) -> np.ndarray:
        input_texts = ['{}'.format(q) for q in queries]
        return self._do_encode(input_texts)

    def encode_corpus(self, corpus: List[Dict[str, str]], **kwargs) -> np.ndarray:
        input_texts = ['{} {}'.format(doc.get('title', ''), doc['text']).strip() for doc in corpus]
        input_texts = ['{}'.format(t) for t in input_texts]
        return self._do_encode(input_texts)

    @torch.no_grad()
    def _do_encode(self, input_texts: List[str]) -> np.ndarray:
        return self.encoder.encode(
            sentences=input_texts,
            batch_size=512,
            normalize_embeddings=True,
            convert_to_numpy=True
        )


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('--model_name_or_path', default="acge_text_embedding", type=str)
    parser.add_argument('--task_type', default=None, type=str)
    parser.add_argument('--pooling_method', default='cls', type=str)
    parser.add_argument('--output_dir', default='zh_results',
                        type=str, help='output directory')
    parser.add_argument('--max_len', default=1024, type=int, help='max length')
    return parser.parse_args()


if __name__ == '__main__':
    args = get_args()
    encoder = SentenceTransformer(args.model_name_or_path).half()
    encoder.encode = functools.partial(encoder.encode, normalize_embeddings=True)
    encoder.max_seq_length = int(args.max_len)

    task_names = [t.description["name"] for t in MTEB(task_types=args.task_type,
                                                      task_langs=['zh', 'zh-CN']).tasks]
    TASKS_WITH_PROMPTS = ["T2Retrieval", "MMarcoRetrieval", "DuRetrieval", "CovidRetrieval", "CmedqaRetrieval",
                          "EcomRetrieval", "MedicalRetrieval", "VideoRetrieval"]
    for task in task_names:
        evaluation = MTEB(tasks=[task], task_langs=['zh', 'zh-CN'])
        if task in TASKS_WITH_PROMPTS:
            evaluation.run(RetrievalModel(encoder), output_folder=args.output_dir, overwrite_results=False)
        else:
            evaluation.run(encoder, output_folder=args.output_dir, overwrite_results=False)

Usage

acge 中文系列模型

在sentence-transformer库中的使用方法：

from sentence_transformers import SentenceTransformer

sentences = ["数据1", "数据2"]
model = SentenceTransformer('acge_text_embedding')
print(model.max_seq_length)
embeddings_1 = model.encode(sentences, normalize_embeddings=True)
embeddings_2 = model.encode(sentences, normalize_embeddings=True)
similarity = embeddings_1 @ embeddings_2.T
print(similarity)

在sentence-transformer库中的使用方法，选取不同的维度：

from sklearn.preprocessing import normalize
from sentence_transformers import SentenceTransformer

sentences = ["数据1", "数据2"]
model = SentenceTransformer('acge_text_embedding')
embeddings = model.encode(sentences, normalize_embeddings=False)
matryoshka_dim = 1024
embeddings = embeddings[..., :matryoshka_dim]  # Shrink the embedding dimensions
embeddings = normalize(embeddings, norm="l2", axis=1)
print(embeddings.shape)
# => (2, 1024)