IndNumNLI-Classifier-v1

About Model

IndNumNLI Classifier is a scikit-learn ensemble model trained on the IndNumNLI v1.0 dataset, a benchmark of 92 premise-hypothesis pairs grounded in authoritative Indian government sources including Census of India 2011, MOSPI National Accounts Statistics 2023, RBI Annual Report 2022-23, National Family Health Survey 5 (2019-21), and MoSRTH Road Transport Yearbook 2021-22. The model classifies each premise-hypothesis pair into one of three Natural Language Inference labels: entailment (hypothesis is arithmetically supported by the premise), neutral (hypothesis makes a claim beyond the scope of the premise such as causal attribution or temporal extrapolation), or contradiction (hypothesis is numerically inconsistent with the premise). The classifier is a soft-voting ensemble of two pipelines: a Random Forest with 200 trees and balanced class weights, and a Gradient Boosting classifier with 150 estimators. Both pipelines use StandardScaler for feature normalisation. Input is a 16-dimensional feature vector capturing numerical entity counts, negation signals, scope signals, lexical overlap, difficulty, reasoning category, and domain encoding. The model achieves a test accuracy of 60.0 percent and macro F1 of 0.639 on the held-out test set, with a 5-fold cross-validation macro F1 of 0.671 plus or minus 0.134.