CureLink NLI Verifier 260623

This repository contains the lightweight JSON verifier artifact used to calibrate CureLink/curelink-biomed-nli-v2 outputs inside the CureLink backend.

It is not a standalone Transformer checkpoint. It is a small logistic-regression-style artifact consumed by the CureLink Node.js runtime in backend/nliVerifier.js.

Files

nli-verifier-260623.json: verifier weights and feature definitions
nli-verifier-260623.summary.json: offline training summary

What It Does

The CureLink runtime first gets a raw NLI label from the base model, then uses this verifier to adjust likely disagreements before hard guardrails run.

Typical inputs include:

raw entailment / neutral / contradiction scores
criterion section
negation and admin/consent flags
explicit-evidence features from patient answers
criterion-family indicators

Typical outputs are:

MET
UNCERTAIN
NOT_MET

Training Snapshot

training rows: 5812
eval rows: 1453
eval accuracy: 0.9780
eval macro F1: 0.9737

Runtime Use

The live backend points NLI_VERIFIER_MODEL_PATH at the JSON file and enables it with:

NLI_VERIFIER_ENABLED=1
NLI_VERIFIER_MODEL_PATH=/app/cure-link-site/backend/models/nli-verifier-260623.json

Notes

This verifier is intended to improve production trial-matching decisions.
It should be evaluated together with backend guardrails, not in isolation as a general-purpose public benchmark model.

Downloads last month: -; Downloads are not tracked for this model. How to track