grounded-multilingual-base

Repo: aghasalim/grounded-multilingual-base (published as siba/grounded-multilingual-base in the paper and leaderboard — same weights).

Token classifier: which spans of an assistant's answer are not supported by the sources it was given. Input <s> answer </s></s> sources </s>; label 1 = unsupported.

Trained on 2,844 synthetic labelled answers in 30 languages (business documents, 100 domains, errors planted and marked by open LLMs) plus 3,883 RAGTruth answers. Fine-tuned from xlm-roberta-base, 384 tokens, 3 epochs, Apple M4.

Benchmark (31 languages, 434 cases, github.com/aghasalim/groundedness): 352/372 planted errors caught, 32/62 false alarms, ~400 ms on a Raspberry Pi 5 (int8). HHEM-2.1-Open on the same cases: 265/372 but 22/62 false alarms.

Use: pip install groundedness then check(answer, sources, model="siba") (hosted), or load with transformers and the detect.py in the repo.

Limits: short-answer business domain; long documents are truncated to 384 tokens; not a fact checker — it compares against the sources you give it.