Tax Doc Classifier
Classify tax-document pages by IRS form and page kind with Jev.
What Jev decides
Which IRS form and which page kind best match the page text.
A text page is compared with descriptions of 261 IRS forms and seven page kinds; Jev returns option probabilities for code to gate.
- Jev role
- Classify a tax-document page against known form and page-kind candidates.
- Input
- Extracted English page text and descriptions of supported forms and page kinds.
- Output
- Choice distributions with form and page-kind confidence.
- Decision frequency
- Once per document page.
- Fallback
- Code gates low-confidence results; scanned pages need OCR before classification.
How it works
Overview
Tax Doc Classifier extracts text from a PDF page and asks Jev to select its IRS form and page kind from documented candidates. Form descriptions are stored in a JSON file derived from IRS PDFs; the project does not train or host a custom model. Code applies a confidence gate to the typed result.
The authors report strict accuracy on their own corpus and provide evaluation materials. That result should be read as author-reported, not as a guarantee for other documents. The classifier is English-only, identifies federal forms, and needs an OCR step for scanned pages.
- Language
- TypeScript
- License
- Apache-2.0
- Author
- kyotofin
Sources
Related projects
pg-jev
Use Jev decisions directly inside PostgreSQL filters, rankings, and classifications.
What Jev decides
Whether a row matches, how strongly it matches, or which supplied label best describes it.
jev-spam-eval
A reproducible email-classification study comparing zero-shot Jev decisions with trained TF-IDF baselines.
What Jev decides
Whether an email is ham, spam, or phishing in the main experiment.
jev-trader
A bounded buy-or-sell decision loop designed around Monad's roughly 300 ms blocks.
What Jev decides
Buy or sell for the next quote.