Links

• Live Space: https://huggingface.co/spaces/build-small-hackathon/jawbreaker
• Dataset/eval bundle: https://huggingface.co/datasets/build-small-hackathon/jawbreaker-scam-defense-data
• GitHub: https://github.com/gowtham0992/jawbreaker

Base + Adapter

• Base model: openbmb/MiniCPM5-1B
• Adapter: build-small-hackathon/jawbreaker-minicpm5-1b-lora-v4
• Runtime target: Hugging Face ZeroGPU via Gradio

Final Completed Evaluation

Guarded eval on eval/hard_v5_eval.jsonl:

• Cases: 394
• Risk-level accuracy: 96.19%
• Scam-type accuracy: 96.19%
• Mean tactic recall: 96.55%
• Dangerous classified as safe: 0
• Dangerous downgraded to needs-check: 0
• Suspicious classified as safe: 0
• Unsafe action violations: 0
• Invalid predictions: 0
• Model errors: 0

The larger 470-case v6 stress run timed out before completion, so it is retained as future evaluation material rather than final evidence.

Intended Behavior

The adapter is trained to produce a strict Jawbreaker JSON contract for consumer scam-defense analysis. The app validates model output and applies a safety guardrail before rendering a plain-English card.

Primary scenarios:

• Package and delivery phishing
• Bank, PayPal, Coinbase, and account-security scams
• Family impersonation and urgent money requests
• Fake recruiters and task/job scams
• Prize, lottery, and refund scams
• Benign or ambiguous messages that should not be over-escalated

Safety Notes

Jawbreaker is a hackathon safety assistant, not professional fraud, legal, financial, or cybersecurity advice. It should encourage safer next steps: do not click suspicious links, do not reply to pressure messages, and verify through official apps, websites, or known phone numbers.

The public dataset and eval files are synthetic/sanitized and do not include raw private chats, phone numbers, or personal message metadata.