The Gemma 4 Good Hackathon

The Gemma 4 Good Hackathon · Special Technology Track · Unsloth prize

MindBridge

Built by Huzayfah Patel

A Hindi mental-health screening prototype with a coded safety acknowledgment and on-device base Gemma after download. The fine-tune is separate from the app, and its claimed 25-point accuracy gain was not independently confirmed.

I also participated in this hackathon with Memory Moment. This coverage was prepared independently after results were announced.

The idea worth stealing

Decide the kill gate before you train.

A UK psychiatrist building for a million community health workers pre-specified three thresholds the fine-tune had to clear in order: JSON validity at 95%, then suicidality-item sensitivity at 90%, then a ten-point accuracy gain, and only then would it ship. A separate rule and acknowledgment flow handles one safety condition: a positive answer on item 9 blocks the app until the worker confirms the helpline was discussed. The 144 seed utterances were co-written with the author’s father, a native Hindi speaker, who caught a dictation error that had turned a self-harm euphemism into “make yourself happy”.

What we checked

Date checked: .

Public repository at the submission revision, Kaggle writeup, and linked artifacts; code was read, not executed. Deployment checks and limits are described below.

Labels apply to each finding and the evidence described, not to the project as a whole.

Not independently verified means the available evidence was insufficient to confirm a claim. It does not mean the claim is false.

  1. Item 9 blocks the app in code

    Verified

    selectNextStep routes to Item9Blocker when the PHQ-9 item 9 score is above zero or when a Hindi keyword rule fires on any item, and keeps routing there until the ASHA taps “I have discussed the helpline.” Source

  2. The app ships the base model

    Verified

    At first launch the app downloads Cactus-Compute/gemma-4-E2B-it. The LoRA and the 2,883-row dataset were on Hugging Face by May 17 but are not in the build, which the writeup states. Source

  3. Reported sensitivity and accuracy after tuning

    Not independently verified

    The builder reports item-9 sensitivity falling from 95.8 percent base to 91.7 after tuning, one adversarial case, still above the 90 percent gate; Likert accuracy rose 25 points. The evaluation notebook has no outputs. Source

    What we need: Executed base-versus-tuned evaluation outputs with per-item predictions, scoring code, and model revisions.

Under the hood

Negation-aware keywords
item9RuleEngine.ts holds about fifty Hindi phrases for active and passive ideation and deliberately excludes “मरने का” so that “didn’t feel like dying” does not trip the blocker.
The acknowledgment fires once
After the ASHA acknowledges, the trigger never re-fires in that session, even if item 9 is revised upward, so the audit trail records exactly one safety event per screening.

Nuance

The writeup discloses limits to its evaluation: one author with his father reviewing Hindi, a 200-row held-out drawn from the teacher-generated distribution, and two disclosed leaks. We did not install the TestFlight build, and the three pytest files only check that notebooks match their scripts.

Sources

Last updated: . Editorial updates do not imply a new technical check.

Builder credits reviewed: , using the official announcement, submission, and any linked credit sources.

The submission credits Huzayfah and links the Huzayfah-Patel model account.

Builder? Add context, request a correction, or ask for a re-check →