Two SetFit models (intent/ -> 12 routes, urgency/ -> 0-3) on paraphrase-multilingual-MiniLM-L12-v2, trained on the
dataset above. It is the System 1 of https://github.com/Leito2/smart-request-router: it answers when confident and
escalates otherwise (the escalation policy lives in the repo, config/policy.setfit.yaml).
Test split (3,080 rows, no escalation): accuracy 0.778, macro-F1 0.704, urgent recall 0.879. With the escalation policy (600-row test sample): selective accuracy 0.920 at 50% coverage, urgent recall 0.956.
from huggingface_hub import snapshot_download
path = snapshot_download("WhiteUniversal/smart-request-router-setfit")
# then, from the repo: SetFitRouter.load(path) (training/baselines/setfit_model.py)
Limits: English and Spanish only; 12 banking-support routes; Spanish training data is machine translated; urgency labels are rule-based from the intent, not human-annotated.
intent/model.safetensors 471 MBurgency/model.safetensors 471 MBsentence-transformers/paraphrase-multilingual-MiniLM-L12-v2