Nihilitybot666/shora-tfidf

Shora TF-IDF + LogReg (NaijaScam v0.1)

Tiny (~4 MB) scikit-learn baseline from Shora.

• binary.joblib: scam vs legit (word 1–2-gram + char_wb 2–5-gram TF-IDF → logistic regression, C=4, balanced)
• scam_type.joblib: same features → 17-way type (9 scam + 8 legit types)

split	accuracy	macro-F1
NaijaScam test (365, synthetic, in-distribution)	97.8	97.8
NaijaScam challenge (40, hand-written, OOD)	75.0	73.3

Scam-type accuracy on test: 89.3. Speed: ~0.5 ms per message on CPU.

Caveat: it misses about half of the short, conversational, hand-written scams (challenge scam recall 50%), so use it as a fast first pass alongside an LLM or human judgement, not as the only safeguard.

import joblib
clf = joblib.load("binary.joblib")
clf.predict_proba(["Your BVN has been suspended. Send the OTP to reactivate."])  # classes_: ['legit', 'scam']