drax234/smartmemory-privacy-gemma

SmartMemory privacy classifier (Gemma 4 E2B, fine-tuned)

A small on-device classifier used by the SmartMemory Android app to decide, on the phone and without any network, whether a short note contains a secret (password, card number, API key, ...), sensitive personal information (health, legal, financial, intimate, children, location), or neither. It answers with one JSON object {"label": "SECRET|SENSITIVE|ORDINARY", "category": ..., "score": ...}. The app uses it so that secrets and sensitive notes are not sent to a cloud AI unasked.

This is a modified version of Google's Gemma 4 E2B instruction-tuned model: a LoRA adapter was merged into the weights and the result converted to the .litertlm format of LiteRT-LM. This file is a Model Derivative of Gemma and was modified from the original.

Files

file	what
gemma-privacy-v2.litertlm	the merged model for LiteRT-LM (about 3.7 GB, int8 decoder)
NOTICE	the Gemma notice and the modification notice

The app downloads one pinned revision of this file over Wi-Fi after the person agrees, verifies its SHA-256, and runs it locally. Nothing about a person's notes is ever sent to this repository or to Hugging Face.

Training data

Synthetic only. About 6,600 short notes in English, Hebrew and mixed text: templated notes with random fake values, and natural-looking notes written by a language model with instructions to use invented people, reserved example domains and never-issued numbers, each labelled by a second model. No real personal data, no user data and no real credentials were used. Card numbers pass the checksum but were never issued; keys are random strings with well-known prefixes.

Evaluation (held-out synthetic notes, the model's plain output)

Share of secrets flagged (answered SECRET or SENSITIVE) / ordinary notes wrongly flagged: 100 % / 0.7 % on 1,284 notes of the first test set, 100 % / 2.5 % on 389 notes written in a different style, 100 % / 5.2 % on 473 notes of an older templated set. Small counts, synthetic text, one training run per recipe: these are evidence, not guarantees. See the SmartMemory project documents for the method.

Limitations and intended use

• It is a screening aid that errs towards flagging: it wrongly flags a few ordinary notes (a few percent) so that very few secrets slip through, but it can miss a secret, especially one that is unlike its training notes. It is not a security guarantee.
• Trained on short notes (up to a few hundred characters); longer text is checked in windows.
• Languages: English and Hebrew (and mixed). Other languages are untested.
• Intended for on-device use inside SmartMemory. Using it to build surveillance or to extract private data from others is not an intended use and is prohibited by the Gemma Prohibited Use Policy.

License

Gemma is provided under and subject to the Gemma Terms of Use found at https://ai.google.dev/gemma/terms. Use of this model is also subject to the Gemma Prohibited Use Policy at https://ai.google.dev/gemma/prohibited_use_policy. Recipients of this model must be given a copy of those terms.