manuelaschrittwieser/invoice-extraction-dataset-v2

📑 Invoice Extraction Dataset v2

This dataset is specifically designed for fine-tuning Large Language Models (LLMs) to perform Structured Data Extraction. It contains 601 unique examples of messy, unstructured invoice and receipt descriptions paired with their clean, machine-readable JSON counterparts.

🚀 Dataset Purpose

General LLMs often include conversational filler when asked for JSON. This dataset was built to train models to:

1. 1.Ignore "Noise": Ignore conversational filler and irrelevant words.
2. 2.Strict Formatting: Output only valid JSON.
3. 3.Entity Recognition: Correctly identify vendors, totals, dates, and items regardless of sentence structure.

📊 Dataset Structure

The dataset is provided in JSONL (JSON Lines) format, which is the industry standard for LLM fine-tuning. Each row contains:

• —instruction: The task description ("Extract invoice details into JSON.")
• —input: A messy, natural language description of a transaction.
• —output: A stringified JSON object containing the extracted fields.

Target Schema

The model is trained to extract the following fields:

{
  "item": "string",
  "quantity": "integer",
  "date": "string",
  "vendor": "string",
  "total": "float",
  "currency": "string"
}

🛠️ Data Generation Methodology

This is a Synthetic Dataset generated using a custom Python pipeline.

• —Variety: Uses 20+ vendors, 20+ item categories, and 5 different currency types.
• —Robustness: 7 different sentence templates were used to ensure the model doesn't overfit to a single way of speaking.
• —Formats: Includes various date formats (e.g., "Oct 12", "12/10/23", "yesterday") to test the model's temporal reasoning.

💻 How to Use

You can load this dataset directly in your training script using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("manuelaschrittwieser/invoice-extraction-dataset-v2")
print(dataset["train"][0])

🎯 Use Cases

• —Fine-tuning Llama-3, Mistral, or Phi-3 for financial automation.
• —Training "Parser Agents" for accounting software.
• —Benchmarking small models on structured data tasks.

🏆 Project Context

This dataset was created as part of a Structured Data Extractor project. It serves as the training foundation for the llama-3-invoice-extractor model.

Developed by: manuelaschrittwieser