af123Af/clausify-models

Clausify models (CUAD)

Two fine-tuned distilbert-base-uncased checkpoints used by the Clausify app (https://github.com/afnanmz168/clausify-p3-app), trained on the 408-contract training split of CUAD (seed-42, contract-level 80/20 split). Research prototype — not legal advice.

Folder	Task	Held-out result (102 test contracts)
presence_mil/final	(category question, 2,000-char window) → does the window contain the clause? Max-pooled over windows.	micro-F1 0.692 alone; 0.779 in an AND-ensemble with a TF-IDF baseline
span/final	Extractive QA: clause span for a category question	token-F1 0.764 given an answer-bearing window

Details and full evaluation: https://github.com/afnanmz168/clausify-p3