预训练模型
Hugging Face Hub 上已公开发布了多个稀疏编码器模型。
社区模型：Hugging Face 上的所有稀疏编码器模型。
模型与此简单接口无缝集成。
from sentence_transformers import SparseEncoder
# Download from the 🤗 Hub
model = SparseEncoder("naver/splade-v3")
# Run inference
queries = ["what causes aging fast"]
documents = [
"UV-A light, specifically, is what mainly causes tanning, skin aging, and cataracts, UV-B causes sunburn, skin aging and skin cancer, and UV-C is the strongest, and therefore most effective at killing microorganisms. Again â\x80\x93 single words and multiple bullets.",
"Answers from Ronald Petersen, M.D. Yes, Alzheimer's disease usually worsens slowly. But its speed of progression varies, depending on a person's genetic makeup, environmental factors, age at diagnosis and other medical conditions. Still, anyone diagnosed with Alzheimer's whose symptoms seem to be progressing quickly â\x80\x94 or who experiences a sudden decline â\x80\x94 should see his or her doctor.",
"Bell's palsy and Extreme tiredness and Extreme fatigue (2 causes) Bell's palsy and Extreme tiredness and Hepatitis (2 causes) Bell's palsy and Extreme tiredness and Liver pain (2 causes) Bell's palsy and Extreme tiredness and Lymph node swelling in children (2 causes)",
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 30522] [3, 30522]
# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[11.3768, 10.8296, 4.3457]])
核心 SPLADE 模型
MS MARCO 段落检索作为黄金标准数据集，包含来自 Bing 搜索引擎的真实用户查询与专家标注的相关文本段落。在此基准上训练的模型在生产搜索系统中作为嵌入模型表现出卓越的有效性。性能分数反映了在此数据集上的评估结果，这是一个很好的指标，但不应是唯一考虑的参数。
BEIR (Benchmarking IR)提供了一个异构基准，用于评估信息检索模型，在此处涵盖了 13 个不同的数据集。平均 nDCG@10 分数代表了所有 13 个数据集上的平均性能。
请注意，以下所有数字均从不同论文中提取。这些模型代表了稀疏神经检索的骨干。
| 模型名称 | MS MARCO MRR@10 | BEIR-13 平均 nDCG@10 | 参数量 |
|---|---|---|---|
| opensearch-project/opensearch-neural-sparse-encoding-v2-distill | NA | 52.8 | 67M |
| opensearch-project/opensearch-neural-sparse-encoding-v1 | NA | 52.4 | 133M |
| naver/splade-v3 | 40.2 | 51.7 | 109M |
| ibm-granite/granite-embedding-30m-sparse | NA | 50.8 | 30M |
| naver/splade-cocondenser-selfdistil | 37.6 | 50.7 | 109M |
| naver/splade_v2_distil | 36.8 | 50.6 | 67M |
| naver/splade-cocondenser-ensembledistil | 38.0 | 50.5 | 109M |
| naver/splade-v3-distilbert | 38.7 | 50.0 | 67M |
| prithivida/Splade_PP_en_v2 | 37.8 | 49.4 | 109M |
| naver/splade-v3-lexical | 40.0 | 49.1 | 109M |
| prithivida/Splade_PP_en_v1 | 37.2 | 48.7 | 109M |
| naver/splade_v2_max | 34.0 | 46.4 | 67M |
免推理 SPLADE 模型
免推理 Splade 在文档部分使用传统的 Splade 架构，而查询部分则是一个 SparseStaticEmbedding 模块，该模块仅返回每个查询 token 的预计算分数。因此，对于这些模型，我们失去了查询扩展能力，但查询推理变得几乎即时，这对于速度优化非常有价值。
| 模型名称 | BEIR-13 平均 nDCG@10 | 参数量 |
|---|---|---|
| opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte | 54.6 | 137M |
| opensearch-project/opensearch-neural-sparse-encoding-doc-v3-distill | 51.7 | 67M |
| opensearch-project/opensearch-neural-sparse-encoding-doc-v2-distill | 50.4 | 67M |
| opensearch-project/opensearch-neural-sparse-encoding-doc-v2-mini | 49.7 | 23M |
| opensearch-project/opensearch-neural-sparse-encoding-doc-v1 | 49.0 | 133M |
| naver/splade-v3-doc | 47.0 | 109M |
模型集合
这些是 Hugging Face Hub 上可用的模型集合