PacketCourt
Collection
PacketCourt audits food-package claims against label evidence using small, transparent AI models. • 9 items • Updated • 1
A 4,386,565-parameter fine-tuned classifier used by PacketCourt's investigation agent to choose the next evidence tool for a packet claim.
Labels: ingredients, nutrition, license, dates, refuse_absolute.
Held-out validation accuracy: 1.000 on a small PacketCourt-specific routing set.
The router proposes an investigation tool; deterministic code remains responsible for final verdicts.
Base model
google/bert_uncased_L-2_H-128_A-2