sbert_raw_finmodern_pair_merged
financial SBERT model for document search
2 min readupdated
This is a fine-tuned version of sentence-transformers/bert-base-nli-mean-tokens adapted for the financial domain. It maps sentences and paragraphs into a 768-dimensional dense vector space for semantic search (retrieval). Trained with the sentence-transformers library on synthetic text pairs from SEC (Securities and Exchange Commission) filings, it is built for developers and analysts building search engines or RAG (retrieval-augmented generation) pipelines for financial reports and corporate documents.
Key features
- Specialized for financial texts and SEC filings
- Maps text into 768-dimensional vectors (Output Dimensionality)
- Maximum sequence length of up to 512 tokens
- Uses Cosine Similarity to measure vector proximity
What you can do
- Set up semantic search across annual and quarterly corporate financial reports
- Plug the model into a RAG pipeline for Q&A over corporate documentation
- Match financial search queries with relevant text passages
Example
from sentence_transformers import SentenceTransformer
# Загрузка модели
model = SentenceTransformer("finvector/sbert_raw_finmodern_pair_merged")
# Запуск инференса
queries = [
'What is the reporting currency used for the Ares Management Corporation financial statements?',
]
documents = [
'Financial Statements (Continued) (Dollars in Thousands, Except Share Data and As Otherwise Noted)',
'Factors contributing to a decrease in Genco’s electric margin were as follows...',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
# Получение оценок сходства
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
Read more
DetailsMetrics on the finance-eval benchmark: cosine_ndcg@10 is 0.9754, and cosine_accuracy@1 is 0.9333.
Model size: 0.11B parameters.
Trained on the fin_modern_bert-pairs-sec-synthetic-v1 dataset using CachedMultipleNegativesRankingLoss.
License: apache-2.0.










Comments