sugiv/call-gate-risk-score
Text Classification • Updated
text stringlengths 42 101 |
|---|
A. Memorization check: score on training data vs. on unseen data |
train (fit on all, score all): AUC 0.9999 | max legit 0.53 | junk above max legit 399/404 |
held-out (5-fold): AUC 0.9999 | max legit 0.58 | junk above max legit 399/404 |
B. Stability: 10 different random 5-fold splits |
AUC mean 0.9999, min 0.9998, max 0.9999 |
C. Leakage check: shuffle the labels; a sound pipeline must fall to ~0.50 |
shuffled-label AUC mean 0.479 (range 0.410–0.559) over 20 shuffles |
D. Same-author test: synthetic only (real-style and junk both written by the same LLM) |
learned (5-fold): AUC 1.0000 | max legit 0.32 | junk above max legit 100/100 |
Jev verdict alone: AUC 0.9978 | max legit 0.94 | junk above max legit 78/100 |
E. Transfer test: train WITHOUT any real FTC robocalls, then score the 300 FTC robocalls |
FTC robocalls scored above every real-style alert: 293/300 (max legit 0.34); median FTC score 0.86 |
Jev verdict alone on the same: 207/300 above max legit 0.94 |
F. New-tactic test: hold out one entire junk category, train on the rest, score it |
bank_impersonation n= 25 above max legit 25/25 median 0.90 |
gov_impersonation n= 15 above max legit 15/15 median 0.97 |
retail_impersonation n= 10 above max legit 10/10 median 0.94 |
injected_scam n= 25 above max legit 25/25 median 0.97 |
sales_pitch n= 25 above max legit 25/25 median 0.98 |
ftc_robocall n=300 above max legit 293/300 median 0.86 |
G. Uncertainty: bootstrap 95% interval for the AUC gain over Jev's verdict |
AUC gain +0.0107, 95% interval [+0.0039, +0.0187] |
Call Gate Robocall Decisions
534 voicemail transcripts from automated phone calls, each labeled junk (scam or unwanted marketing) or legit (a genuine alert from a bank, pharmacy, doctor, school, utility or delivery company). Each one comes with the answers Jev, TypeSafe's System One decision model, gave to eight evidence questions.
It was built to train and test the decision layer of Call Gate, a personal system that keeps toll-free robocalls out of a phone's carrier voicemail. The companion model is sugiv/call-gate-risk-score.
What's inside
| File | Rows | Contents |
|---|---|---|
data/corpus.jsonl |
534 | id, source, category, label, transcript |
data/features.jsonl |
534 | id, label, category, source, model (served Jev snapshot), evidence (Jev's answers, below) |
data/oof_predictions.jsonl |
534 | Out-of-fold junk probability from 5-fold cross-validation, and the policy rule and action that would apply |
data/report.json |
— | Per-category results, baselines and the final model |
data/validation.txt |
— | Output of the seven overfitting, leakage and transfer tests (see the model card) |
Sources
| Source | Rows | Label | Notes |
|---|---|---|---|
ncsu-ftc |
300 | junk | Random sample of English transcripts from NC State's Robocall Audio Dataset: FTC enforcement-case robocalls, transcribed with Whisper. Public domain; please cite the original authors. |
synthetic-deepseek |
230 | 130 legit, 100 junk | Written by DeepSeek (deepseek-flash) to look like speech-to-text of automated voicemails. Organization names are real brands used in fictional messages. |
real-labeled |
4 | junk | Real "prime loan" robocalls received Oct 16, 2026, transcribed by Deepgram nova-3. Caller numbers and call IDs removed. |
Categories
Legit: bank_fraud_alert (25), bank_account_notice (15), vague_legit (10), pharmacy (15), medical_appointment (15), school (15), utility (10), delivery (10), legit_with_disclaimer (15, real notices that say "this is not a sales call").
Junk: ftc_robocall (300), bank_impersonation (25), gov_impersonation (15), retail_impersonation (10), injected_scam (25, scams that insist "this is not a scam"), sales_pitch (25), prime_loan (4).
Jev evidence fields
All values are probabilities from typesafe/jev-1.13 (snapshot jev-1.13-20260917). The state was the transcript alone.
| Field | Jev question |
|---|---|
sector, protected |
Which kind of organization the caller claims to be; protected = P(financial, about an existing account) + P(healthcare) + P(education) |
specific |
Concrete specifics: card digits, exact amount, merchant, date, name |
official |
Points to an official channel: the number on the card, the app, the website |
pressure |
Threats or urgency |
agent |
Pushes to a live agent, or to give or verify personal details |
sales |
Offers a loan, rate, warranty, insurance or deal |
selfAssert |
Insists it is legitimate, verified, "not a scam" |
catLegit, catScam, catMkt |
Jev's overall category guess |
The exact question wording is in code/worker/src/jev.ts in the model repo.
Intended use and limits
- For: research and teaching on decision layers for call screening, calibration of System One model outputs, and safety-first policy design.
- The legit side is synthetic. Real automated alerts are messier and more varied; treat the legit results as optimistic.
- English only. Short transcripts of about 1–4 sentences.
- Jev answers are version-specific. Re-featurize after any model change; answers vary by a few hundredths between runs.
- Not for: blocking calls without a human in the loop, or judging real organizations. The impersonation texts are fictional.
Citation
If you use the ncsu-ftc rows, cite NC State's Robocall Audio Dataset (Prasad and Reaves). For this compilation: Call Gate Robocall Decisions, sugiv, 2026.
- Downloads last month
- 18
Models trained or fine-tuned on sugiv/call-gate-robocalls