sbert_raw_investo-finmodern_tri_merged
fintech model for semantic search
2 min read
This model is a specialized fine-tuned version of the base sentence-transformers/bert-base-nli-mean-tokens model, trained on the investopedia-triplets dataset for financial search and text matching tasks. It converts sentences and paragraphs into dense 768-dimensional vector representations (embeddings) for efficient document retrieval. It differs from the base version through its narrow specialization in finance and investment topics. It is useful for developers of financial applications, search engines, and knowledge bases for analyzing corporate reports and texts.
Key features
- Based on the BERT architecture with mean pooling to generate 768-dimensional vectors.
- Trained on 11,687 triplet examples from the investopedia-triplets dataset.
- Supports a maximum sequence length of up to 512 tokens.
- Uses cosine similarity to evaluate text proximity.
What you can do
- Set up semantic search across financial documents, glossaries, and investment articles.
- Build a question-answering retrieval system for bank and broker knowledge bases.
- Cluster and rank financial content based on vector similarity.
Example
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("sachinyerrawar/sbert_raw_investo-finmodern_tri_merged")
queries = ['Who is the principal of Sona Financial?']
documents = ['Mark Struthers']
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
Read more
DetailsMetrics (investo_test): cosine_ndcg@10 is 0.9187, cosine_accuracy@1 is 0.868.
Hardware: lightweight model (0.11 billion parameters / ~110 million), runs on a standard CPU or GPU.
Limitations: maximum sequence length of 512 tokens, English language.
License: apache-2.0.










Comments