dhanushmekaka/qwen25-1.5b-invoice-extraction-v3

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 9, 2026Architecture:Transformer Featherless Exclusive Cold

The dhanushmekaka/qwen25-1.5b-invoice-extraction-v3 is a 1.5 billion parameter QLoRA fine-tune of Qwen2.5-1.5B-Instruct, designed for extracting structured financial terms from legal-financial documents like SEC contracts and commitment letters. This model specializes in identifying financial values and labeling them with a fixed vocabulary of term types, outputting the results as strict JSON. It is optimized for drafting and first-pass extraction as an aid for human reviewers, with a context length of 32768 tokens.

Loading preview...

Overview

This model, dhanushmekaka/qwen25-1.5b-invoice-extraction-v3, is a QLoRA fine-tune of the Qwen/Qwen2.5-1.5B-Instruct base model, specifically engineered for financial term extraction from legal documents. It processes contract excerpts and outputs a JSON object containing identified financial values and their corresponding term_type from a predefined vocabulary.

Key Capabilities

  • Structured Financial Extraction: Identifies and labels financial terms (e.g., interest_rate, credit_facility, salary) from text into a fixed-schema JSON format.
  • QLoRA Fine-tuning: Utilizes 4-bit NF4 QLoRA with Unsloth for efficient training, applied to all attention and MLP projection layers.
  • Specialized Training Data: Trained on 3,430 examples from TheTokenFactory/sec-contracts-financial-extraction-instructions, using a two-stage curriculum focusing on core financial fields.
  • JSON Output: Aims for strict JSON output, achieving 94.0% JSON Output Validity in evaluations.

Performance & Limitations

While still a converging checkpoint (v3), the model shows a Field-Level Extraction Accuracy of 44.9% on a 50-example held-out split. Exact Match Accuracy for the full JSON object is 16.0%. Limitations include lower precision/recall for rare term_types due to limited training coverage and the need for potential parsing fallbacks due to occasional invalid JSON output. Further training and an expanded evaluation set are recommended for production use.

Intended Use

This model is designed for drafting and first-pass extraction of financial terms from documents such as contracts and commitment letters. It serves as an aid for human reviewers and is not intended as a standalone source of truth for critical financial or legal decisions.

Finetunes
2 models