Invoice GL Code Classification

Explore 5,566 real invoices and predict GL codes using Aito predictive queries. Dataset from a public invoice classification benchmark with 9 GL codes, 1,253 vendors, and 36 product categories.

Database: aito-datasets

About This Dataset

Overview

Invoice GL Code Classification Dataset

This dataset contains 5,566 invoices from a public invoice classification benchmark. Each invoice has a natural language description and needs to be classified into one of 9 GL (General Ledger) codes.

Columns

  • Inv_Id — Invoice identifier
  • Vendor_Code — Vendor identifier (1,253 unique vendors)
  • GL_Code — Target GL code for classification (9 codes)
  • Inv_Amt — Invoice amount
  • Item_Description — Natural language description of the invoice item
  • Product_Category — Product category code (36 categories)

Key Properties

  • 60 vendors map to multiple GL codes — real classification ambiguity
  • 3 product categories are shared across GL codes
  • Descriptions are messy natural language with shuffled word order
  • Used in the Predictive Databases vs LLMs comparison experiment

Explore Data

Browse invoices

Interactive query — results displayed as table.

GL code distribution

Interactive query — results displayed as table.

Predict GL Code

Try predicting GL codes from invoice features

How to predict

Predicting GL Codes

Use the predict endpoint to classify an invoice into a GL code based on its features. The predictive database uses Bayesian inference over the full dataset to produce calibrated probability scores.

Try modifying the where clause with different vendor codes, descriptions, and amounts to see how predictions change.

Predict from description

Interactive predict — results displayed as table.

Predict with explainability

Interactive predict — results displayed as json.

Statistical Relationships

What relates to GL code?

Interactive relate — results displayed as table.

Evaluate Accuracy

About evaluation

Cross-Validation

The evaluate endpoint tests prediction accuracy on 20% of the data (every 5th row). This gives an unbiased estimate of how well the predictive database can classify unseen invoices.

The query below uses Item_Description, Vendor_Code, and Product_Category as input features to predict GL_Code, achieving 99.5% accuracy on 1,114 test samples.

GL Code prediction accuracy

Interactive evaluate — results displayed as json.