Jawbreaker Submission Package

Jawbreaker is a private scam-defense app for someone you love.

It is built for the moment before someone clicks a suspicious link, replies to an impersonator, shares a code, calls an unknown number, or sends money. The motivating user is a friend's grandmother who had already been affected by scam messages; private identity and message details are intentionally omitted.

Main Links

• Live Space: https://huggingface.co/spaces/build-small-hackathon/jawbreaker
• Final model adapter: https://huggingface.co/build-small-hackathon/jawbreaker-minicpm5-1b-lora-v8
• Dataset/eval bundle: https://huggingface.co/datasets/build-small-hackathon/jawbreaker-scam-defense-data
• Collection: https://huggingface.co/collections/build-small-hackathon/jawbreaker-6a263632dcd0b6d41ca914ff
• GitHub repo: https://github.com/gowtham0992/jawbreaker
• Article draft: https://huggingface.co/spaces/build-small-hackathon/jawbreaker/blob/main/docs/huggingface-article-draft.md

Model + Runtime

• Base model: openbmb/MiniCPM5-1B
• Adapter: build-small-hackathon/jawbreaker-minicpm5-1b-lora-v8
• Runtime: Hugging Face ZeroGPU, direct Transformers inference
• Training/eval: Modal A100 PEFT/LoRA runs
• No external hosted LLM API is used in the scam-analysis path

Final Eval Evidence

Final guarded hard-suite report:

https://huggingface.co/spaces/build-small-hackathon/jawbreaker/blob/main/eval/reports/jawbreaker-minicpm5-1b-lora-v8-hard632-safetyguard-v4.json

Headline metrics:

• Cases: 632
• Risk accuracy: 579/632, 91.61%
• Scam type accuracy: 561/632, 88.77%
• Mean tactic recall: 90.69%
• Dangerous as safe: 0
• Dangerous as needs_check: 0
• Safe as dangerous or suspicious: 0
• Unsafe action violations: 0
• Invalid predictions: 0
• Model errors: 0

Evidence Docs

• Field notes: https://huggingface.co/spaces/build-small-hackathon/jawbreaker/blob/main/FIELD_NOTES.md
• Codex evidence: https://huggingface.co/spaces/build-small-hackathon/jawbreaker/blob/main/CODEX_JUDGE_EVIDENCE.md
• Build log: https://huggingface.co/spaces/build-small-hackathon/jawbreaker/blob/main/CODEX_BUILD_LOG.md
• Honest submission guardrails: https://huggingface.co/spaces/build-small-hackathon/jawbreaker/blob/main/HONEST_SUBMISSION.md

Claimed Fit

• Backyard AI: real family-safety workflow for scam defense
• Best MiniCPM Build: MiniCPM5-1B is the live model family
• Well-Tuned: published MiniCPM5-1B LoRA adapter and eval suite
• Tiny Titan: 1B model under the 4B threshold
• Off the Grid: no external hosted LLM API in the analysis path
• Off Brand: custom Gradio Server UI, not stock component layout
• Sharing is Caring: public model, dataset/evals, collection, build docs
• Field Notes: product/model/runtime decisions documented
• Modal: training and guarded eval runs on Modal A100
• Codex/OpenAI: public GitHub with Codex-attributed commits and file-level evidence

Remaining Public Story Work

• Demo video: pending
• Social post: pending
• Hugging Face Article: drafted, pending publication through the Articles UI

The current submission is intended to be judged from the live Space plus the public model/dataset/collection/GitHub evidence above.