---
license: other
license_name: commercial
license_link: LICENSE
language:
- en
task_categories:
- text-classification
pretty_name: Attestable Scam Detection Corpus
tags:
- scam-detection
- phishing
- adversarial
- red-team
- telegram
- safety
- attestation
- prompt-injection
size_categories:
- n<1K
---

# Attestable Scam Detection Corpus v0.1

Matched-pair adversarial training data for AI safety classifiers. 153 human-crafted examples across 7 scam categories with published detection rates and full provenance.

## Published Detection Rates

| Pipeline | F1 | Precision | Recall | False Negatives |
|----------|-----|-----------|--------|-----------------|
| Heuristic-only | 0.703 | 0.889 | 0.582 | 23/55 |
| Full AI (DeepSeek V4 Pro) | 0.981 | 1.000 | 0.964 | 2/55 |

## Dataset Structure

| Split | Scam | Safe | Total |
|-------|------|------|-------|
| Green-team messages | 60 | 30 | 90 |
| Red-team adversarial | 25 | 2 | 27 |
| Green-team profiles | 15 | 13 | 28 |
| Red-team profiles | 8 | 0 | 8 |
| **Total** | **108** | **45** | **153** |

## Categories

- **Phishing**: Account suspension, billing scams, OTP harvest, delivery fee scams
- **Crypto Drainers**: Wallet connect, seed phrase harvest, airdrop bait, typosquat domains
- **Impersonation**: Telegram support, group admin, celebrity, exchange support
- **Job Scams**: Task scams, advance-fee, reshipping mule recruitment
- **Romance Scams**: Grooming, isolation tactics, gift card requests
- **Investment Scams**: Guaranteed returns, signal groups, presale rug pulls
- **Profile Screening**: Impersonation profiles, crypto recruiters, spam bots

## Adversarial Coverage

Red-team corpus tests evasion techniques: Unicode homoglyphs, zero-width injection, leetspeak, prompt injection, boundary probing, category blindspots.

## Data Format

JSONL. Each line:
```json
{
  "id": "unique identifier",
  "category": "phishing|crypto_drainer|impersonation|...",
  "text": "The message or profile text",
  "ground_truth_verdict": "safe|suspicious|dangerous",
  "ground_truth_scam_type": "...",
  "source": "public-pattern-reconstruction|synthetic-benign",
  "notes": "Human context about detection challenge"
}
```

## Licensing

Commercial license. Research use permitted with attribution. Contact for commercial terms.

Built by House Attestable. Methodology: green-team evaluation → red-team adversarial testing → public disclosure.