Neurobiber authorship-verification model
A random forest for same-author verification over
Neurobiber style vectors.
Each text pair is represented as [A | B | |A-B|] of the two 96-dim binary
presence vectors; the forest predicts same vs. different author.
Trained on PAN 2020 (small) fanfiction pairs. Used by
biberplus:
from biberplus.neurobiber.authorship import compare_texts
compare_texts(text_a, text_b)
The CLI biberplus compare a.txt b.txt downloads this file on first use.
Not forensic-grade. Accuracy is useful signal, never sole evidence of authorship. This checkpoint reports weighted F1 ~0.60 on its held-out PAN 2020 split; document-level presence vectors saturate on long texts, so this is a floor, not the ceiling of the approach.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support