finance-entity-extractor / CHANGELOG.md
Ranjit Behera
docs: Update CHANGELOG for v1.1.0 release
438d5f9
|
Raw History Blame Contribute Delete
3.4 kB

Changelog

All notable changes to FinEE will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Added

  • (Next features go here)

[1.1.0] - 2026-01-12

Added

  • Complete Data Pipeline (scripts/data_pipeline/)

    • step1_unify.py: Unifies MBOX, JSON, CSV, XML sources
    • step2_filter.py: Removes OTPs, spam, marketing messages
    • step3_baseline.py: Tests regex extractor accuracy
    • step4_label.py: Creates labeled training data with ground truth
  • Synthetic Data Generator

    • generate_synthetic.py: Production-grade grammar-based generator
      • 100K+ realistic Indian bank transactions
      • All major banks (HDFC, ICICI, SBI, Axis, Kotak, PNB, BOB, etc.)
      • Brokerages (Zerodha, Groww, Upstox, Angel One, 5Paisa, etc.)
      • E-commerce, food, travel, utilities, entertainment categories
    • generate_advanced.py: Advanced features
      • Markov Chain for realistic message flow
      • Real data calibration from actual samples
      • Multilingual support (Hindi, Tamil, Telugu, Bengali, Kannada)
      • Data augmentation and edge case oversampling
  • LLM Fine-tuning Pipeline (scripts/finetune.py)

    • Supports MLX (Apple Silicon) and PyTorch backends
    • LoRA fine-tuning with automatic data preparation
    • Model fusion and evaluation utilities

Performance

  • Trained on 152,519 records (2,419 real + 100K synthetic + 50K multilingual)
  • Val loss: 2.42 → 0.46 (81% reduction)
  • 100% JSON parsing accuracy on test cases
  • Multilingual extraction working (Hindi, Tamil, Telugu, Bengali, Kannada)
  • Fine-tuned model: 7.6GB (Phi-3-mini + LoRA fused)

Models

  • Fine-tuned model: finetuned-v1/ on Hugging Face
  • LoRA adapters: lora-adapters/ on Hugging Face

[1.0.3] - 2025-01-11

Added

  • Lakhs notation support (e.g., "1.5 Lakh" → 150000)
  • benchmark.py script for accuracy verification
  • Torture test suite for edge cases
  • CONTRIBUTING.md with Git Flow guidelines
  • Professional branching: main, develop, feature/*

Changed

  • Moved Jupyter notebooks to experiments/ folder
  • Default use_llm=False for instant usage (no model download)
  • Updated README with edge case examples

Fixed

  • Double-escaped regex patterns in amount extraction

[1.0.2] - 2025-01-11

Changed

  • Default to regex-only mode (use_llm=False)
  • Package works instantly without downloading 5GB model

Fixed

  • Package build configuration (hatch sources)

[1.0.1] - 2025-01-11

Fixed

  • Package did not include source files (only 5KB)
  • Fixed pyproject.toml build configuration

[1.0.0] - 2025-01-11

Added

  • Initial PyPI release
  • 5-tier additive extraction pipeline (Cache/Regex/Rules/LLM/Validate)
  • Multi-backend support (MLX, PyTorch, GGUF)
  • CLI with finee command
  • 88 unit tests
  • Colab demo notebook
  • JSON schema contract documentation
  • Support for 5 banks: HDFC, ICICI, SBI, Axis, Kotak

Performance

  • 94.5% field accuracy on multi-bank benchmark
  • <1ms latency in regex-only mode
  • 50,000+ messages/second throughput

Links