evidenceprune-modernbert-large

evidenceprune is a lightweight context-pruning model fine-tuned for fact checking. Given a claim and a web page, it marks the sentences a fact-checker would need.

This checkpoint is a 0.4B encoder fine-tuned from ModernBERT-large as a token classifier.

Usage

pip install git+https://github.com/ofbread/evidenceprune

from evidenceprune import Pruner

kept = Pruner("ofbread/evidenceprune-modernbert-large").prune(claim, documents)

Input

Pruner(model, threshold=None, device=None) loads the model. prune(claim, documents, requirements=None, claim_date="", speaker="", cap_windows=16, stop_after_empty=2, doc_ceiling=12000) reads the pages.

parameter	type	meaning
claim	str	the claim being checked
documents	list of {"text": str, "title": str, "url": str}	the pages; title and url are optional but the model was trained with them
claim_date	str, optional	when the claim was made, YYYY-MM-DD
speaker	str, optional	who made the claim
requirements	list of str, optional	what must be established to check the claim
threshold	float, default 0.3458	a sentence is kept when its P(keep) is at or above this
cap_windows	int, default 16	read at most this many windows of 48 sentences (at most 10,000 characters each) per page
stop_after_empty	int, default 2	stop reading a page after this many consecutive windows with nothing kept
doc_ceiling	int, default 12000	a page contributes at most this many characters
device	str, optional	cuda or cpu

Output

prune returns one result per document, in order:

field	type	meaning
text	str	the kept sentences, in page order, gaps marked with […]
kept	list of (start, end)	offsets of each kept sentence in the document's text
sentences	list of str	the same sentences as strings
probs	dict, sentence index → float	P(keep) for every sentence the model read
title_kept	bool	the title is read as the page's first sentence
windows_read, n_windows	int	windows the model read, windows in the page
stopped_early	bool	whether the stop rule ended the read