DMetaSoul/sbert-chinese-qmc-finance-v1-distill

此模型是之前开源金融问题匹配模型的蒸馏轻量化版本（仅4层 BERT），适用于金融领域的问题匹配场景，比如：

• 8千日利息400元? VS 10000元日利息多少钱
• 提前还款是按全额计息 VS 还款扣款不成功怎么还款？
• 为什么我借钱交易失败 VS 刚申请的借款为什么会失败

离线训练好的大模型如果直接用于线上推理，对计算资源有苛刻的需求，而且难以满足业务环境对延迟、吞吐量等性能指标的要求，这里我们使用蒸馏手段来把大模型轻量化。从 12 层 BERT 蒸馏为 4 层后，模型参数量缩小到 44%，大概 latency 减半、throughput 翻倍、精度下降 5% 左右（具体结果详见下文评估小节）。

Usage

1. Sentence-Transformers

通过 sentence-transformers 框架来使用该模型，首先进行安装：

pip install -U sentence-transformers

然后使用下面的代码来载入该模型并进行文本表征向量的提取：

from sentence_transformers import SentenceTransformer
sentences = ["到期不能按时还款怎么办", "剩余欠款还有多少？"]

model = SentenceTransformer('DMetaSoul/sbert-chinese-qmc-finance-v1-distill')
embeddings = model.encode(sentences)
print(embeddings)

2. HuggingFace Transformers

如果不想使用 sentence-transformers 的话，也可以通过 HuggingFace Transformers 来载入该模型并进行文本向量抽取：

from transformers import AutoTokenizer, AutoModel
import torch


#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)


# Sentences we want sentence embeddings for
sentences = ["到期不能按时还款怎么办", "剩余欠款还有多少？"]

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('DMetaSoul/sbert-chinese-qmc-finance-v1-distill')
model = AutoModel.from_pretrained('DMetaSoul/sbert-chinese-qmc-finance-v1-distill')

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling. In this case, mean pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

print("Sentence embeddings:")
print(sentence_embeddings)

Evaluation

这里主要跟蒸馏前对应的 teacher 模型作了对比：

性能：

	Teacher	Student	Gap
Model	BERT-12-layers (102M)	BERT-4-layers (45M)	0.44x
Cost	23s	12s	-47%
Latency	38ms	20ms	-47%
Throughput	418 sentence/s	791 sentence/s	1.9x

精度：

	csts_dev	csts_test	afqmc	lcqmc	bqcorpus	pawsx	xiaobu	Avg
Teacher	77.40%	74.55%	36.00%	75.75%	73.24%	11.58%	54.75%	57.61%
Student	75.02%	71.99%	32.40%	67.06%	66.35%	7.57%	49.26%	52.80%
Gap (abs.)	-	-	-	-	-	-	-	-4.81%

基于1万条数据测试，GPU设备是V100，batch_size=16，max_seq_len=256

Citing & Authors

E-mail: xiaowenbin@dmetasoul.com