Data & licences

real data · open models
What this site shows, where the data comes from, and how the numbers were produced.

Explorer mode (this site)

Every reading, rule result and ranked page here is the real output of the evaluation runs reported in the README. Nothing is simulated. The models need a GPU, which free hosting does not provide, so this site shows the saved outputs. The accounting checker runs live in your browser, so edits are rechecked instantly.

Live mode: run the FastAPI reader on a GPU machine (bash scripts/wsl.sh bash scripts/serve.sh) and open the site locally; the Scan receipt page then reads uploaded photos with the fine-tuned model.

How the numbers were produced

  • Receipts: the 100-receipt CORD-v2 test split, never used in training. Fine-tuning used the 800-receipt train split.
  • Search: all 150 FinanceBench questions, scored against the evidence pages the analysts labelled.
  • Hardware: one RTX 4060 laptop GPU (8 GB), Windows 11 with WSL2 Ubuntu. Cost: $0.
  • Code and scripts that reproduce every number: github.com/WAYTECHG/SumProof

Datasets

  • CORD-v2 receipts · CC-BY-4.0

    Real Indonesian receipts with field labels, by NAVER Clova. The receipt photos on this site come from its test split.

  • FinanceBench · CC-BY-NC-4.0

    Questions, answers and evidence pages written by financial analysts about real SEC filings, by Patronus AI. Page text is shown here for non-commercial demonstration.

Models and code

  • Qwen3.5-4B, Qwen3-Embedding-0.6B, Qwen3-Reranker-0.6B · Apache-2.0

    Open models from the Qwen team, downloaded from Hugging Face.

  • SumProof code · MIT

    The licence covers the code only; datasets and model weights keep their own licences.