Ranjit0034/finance-entity-extractor

TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:4kPublished:Dec 27, 2025License:mitArchitecture:Transformer Open Weights Cold

Ranjit0034/finance-entity-extractor is a 4 billion parameter hybrid model developed by Ranjit Behera, combining regex with a Phi-3 LLM for production-grade financial entity extraction from Indian bank messages. It achieves 94.5% accuracy with a unique architecture that uses regex for common cases (<1ms latency) and optionally deploys the 3.8B LLM for complex edge cases. This model is specifically optimized for high-accuracy, low-latency financial data parsing, providing a guaranteed JSON output schema.

Loading preview...

Finance Entity Extractor (FinEE) v1.0

FinEE is a specialized, production-grade financial entity extraction model designed for Indian bank messages. Developed by Ranjit Behera, it employs a hybrid architecture that prioritizes efficiency and accuracy. By default, it operates with a highly optimized regex engine, delivering sub-millisecond latency and 87% accuracy without requiring any model download. For more complex or ambiguous cases, it can optionally engage a 3.8 billion parameter Phi-3 LLM, boosting accuracy to 94.5% with a one-time 7GB download.

Key Capabilities

  • Hybrid Performance: Achieves <1ms latency with regex for 87% accuracy, and ~50ms with the LLM for 94.5% accuracy.
  • Offline Operation: The regex component runs 100% offline, with the LLM also running locally after initial download.
  • Guaranteed Output Schema: Provides a consistent JSON output for extracted entities like amount, currency, type, merchant, and category.
  • Optimized for Indian Banks: Specifically trained and tested for messages from major Indian banks (HDFC, ICICI, SBI, Axis, Kotak).
  • Robust Edge Case Handling: Designed to correctly parse challenging inputs, including varied currency formats, missing spaces, and multiple values.
  • Multi-platform Support: Compatible with Apple Silicon (MLX), NVIDIA GPUs (PyTorch/CUDA), and CPUs (llama.cpp/GGUF).

Good For

  • Extracting structured data from unstructured financial SMS or push notifications.
  • Applications requiring high-speed, accurate parsing of transaction details.
  • Developers needing a reliable, locally deployable solution for financial entity recognition in an Indian context.

Finetunes
1 models