ratulsur/multi-format-finance-parser

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 16, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The ratulsur/multi-format-finance-parser is a 7.6 billion parameter financial document parser, fine-tuned by ratulsur on Qwen2.5-7B-Instruct using QLoRA (4-bit NF4 quantization). This model is specifically optimized to extract structured JSON data from various financial documents, including invoices, SAP reports, income statements, balance sheets, and bank statements. It excels at transforming raw financial text into a standardized JSON output for integration and analytics.

Loading preview...

Multi-Format Finance Document Parser

This model, ratulsur/multi-format-finance-parser, is a specialized 7.6 billion parameter language model fine-tuned on Qwen2.5-7B-Instruct using QLoRA (4-bit NF4 quantization). Its core function is to parse raw text from diverse financial documents and output structured JSON data, making it ideal for automated financial processing and data integration.

Key Capabilities

  • Structured Data Extraction: Converts unstructured financial text into a predefined JSON schema, including fields like document_type, vendor, total_amount, line_items, and more.
  • Multi-Format Support: Capable of processing a wide range of financial documents, such as:
    • Invoices (vendor invoices, GST bills)
    • SAP Reports (ALV exports, FI vendor payment reports)
    • Income Statements (P&L, earnings reports)
    • Balance Sheets (assets, liabilities, equity)
    • Bank Statements (transaction records)
    • Purchase Orders, SQL Results, and tabular CSV/Excel data.
  • Efficient Fine-tuning: Utilizes QLoRA with 4-bit NF4 quantization and double quantization, significantly reducing model size and memory footprint while maintaining performance.
  • Production-Grade Output: Designed to provide clean, parseable JSON suitable for downstream processing, ERP systems, and analytics pipelines.

Training Details

The model was trained on a curated dataset of 954 samples, including the CORD-v2 dataset (real receipt images with JSON) and synthetically generated invoices, SAP reports, and income statements. Training was performed on an L40S 48GB GPU for approximately 1 hour.

Limitations

Primarily trained on English financial documents, with best performance on structured text rather than handwritten documents. Accuracy for scanned documents is dependent on OCR quality, and SAP report parsing is optimized for ALV-style exports.

Finetunes
1 models