This model, fpml-semantic-model, is a specialized version of the base all-MiniLM-L6-v2 Sentence Transformer, fine-tuned on proprietary data derived from FpML XSD schema definitions and documentation. It is engineered to provide superior semantic relevance for financial terminology, significantly outperforming general-purpose models when searching FpML data structures.

The fpml-semantic-model provides a robust, offline-capable utility for performing semantic searches across the Financial products Markup Language (FpML) XSD schema definitions. Instead of relying on exact keyword matches, this model allows users to search for complex FpML elements using natural language, helping developers and analysts quickly locate the correct XML structure for derivatives and financial transaction data.

💡 Model Overview

This model maps complex natural language queries about financial concepts directly to the specific FpML tags and definitions within the schema.

Key Use Cases

Schema Exploration: Quickly locate the exact FpML element (e.g., tradeId, floatingRateIndex) needed for a given financial scenario.
Data Mapping: Improve accuracy when mapping regulatory or proprietary data fields to FpML standards.-
Validation: Use semantic similarity to suggest relevant documentation for validation failures.

🛠️ Usage with fpml-semantic-search Package

If you are using the official fpml-semantic-search Python package, this model is automatically loaded via the SemanticSearchModel class.

from fpml_semantic_search import StructuralModel, SemanticSearchModel
from fpml_semantic_search.fpml_structural_model import MOCK_STRUCTURAL_DATA 

# 1. Initialize the Structural Data (Mocking data for demonstration)
structural_model = StructuralModel(MOCK_STRUCTURAL_DATA)

# 2. Initialize the Semantic Search Model
# NOTE: This class automatically loads the 'your_username/fpml-semantic-model' 
# from the Hugging Face Hub (as defined in the source code).
search_model = SemanticSearchModel(structural_model.data)

# 3. Perform a natural language search
query = "What element defines the reference rate used for calculating interest payments in a swap?"

results = search_model.semantic_search(query, top_k=3)

print(f"\n--- Semantic Search Results for: '{query}' ---\n")
for result in results:
    print(f"[{result['Score']}] Tag: {result['Tag Name']} | Description: {result['Description']}")

# Expected Output (High relevance scores due to fine-tuning):
# [0.9125] Tag: floatingRateIndex | Description: The benchmark rate (e.g., 'USD-LIBOR-BBA', 'EUR-EONIA-OIS') used for the floating stream.
# [0.8850] Tag: rateTreatment | Description: Specifies how a mid-market rate is treated, such as 'mid', 'average', or 'interpolated'.

🚀 Installation

This package requires Python (3.8+).

pip install fpml-semantic-model sentence-transformers numpy scipy

Note: The sentence-transformers and numpy/scipy libraries are required for embedding the query and calculating similarity scores, respectively.

💡 How It Works

-Embeddings: The model uses pre-calculated embeddings for FpML element names and their documentation, generated using the highly efficient all-MiniLM-L6-v2 Sentence Transformer model. These embeddings are stored locally in the package's default_embeddings.json file.

-Search: When a user provides a natural language query (e.g., "details about when the interest rate will be paid"), the model embeds the query and calculates the Cosine Similarity against all stored FpML element vectors.

-Results: It returns the elements with the highest semantic relevance, ranked by score.

Direct Sentence-Transformer Loading

You can also load this model directly into any application using the SentenceTransformer library:

from sentence_transformers import SentenceTransformer

MODEL_ID = 'thelocalhost/fpml-semantic-model' 
model = SentenceTransformer(MODEL_ID)

# Example: Embed a query and compare it to known FpML elements
query_embedding = model.encode("currency of the notional amount", normalize_embeddings=True)

Model Details

Attribute	Value

Model ID: thelocalhost/fpml-semantic-model
Backbone Model: Fine-tuned all-MiniLM-L6-v2 (Sentence Transformers)
Training Dataset: FpML Element Names and corresponding XSD Documentation Pairs
Loss Function: Multiple Negatives Ranking Loss (MNR)
Embedding Dimension: 384
Similarity Metric: Cosine Similarity
Performance: Optimized for FpML-specific semantic alignment.


Default Data Source	Elements and documentation extracted from core FpML XSD files (provided via default_embeddings.json).
Use Case	Semantic indexing and retrieval, FpML schema exploration.

📚 Training Data Generation

The training dataset was generated using the following methodology:

Corpus: All element names and their full XSD documentation strings were extracted from the core FpML schema files.
Positive Pairs: A positive pair was defined as (Element Name, Documentation String).

Example: (tradeId, A unique identifier assigned to the trade by one of the parties.)
Fine-Tuning: The model was trained to ensure the vector distance between these positive pairs was minimized, thereby aligning the vector space specifically toward FpML concepts.