Modelpublic

V-ince-18/saferide-gemma-4-e2b-lora

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

0likes10downloads

Model Card

SafeRide Gemma 4 E2B v0.3 LoRA Adapter Registry

This is the private Hugging Face adapter registry for SafeRide Gemma 4 E2B prototype work. It now points reviewers to the completed v0.3 controlled LoRA adapter package.

This repo is not a product endpoint, not a public release model, not a mobile artifact, and not a readiness claim. It is a private PEFT/LoRA artifact plus a review/audit card.

Current Status

Field	Value
Current reviewed lane	v0.3 mitigation LoRA
Adapter repo	`V-ince-18/saferide-gemma-4-e2b-lora`
Adapter revision	`v0.3-mitigation-20260704`
Evaluated adapter commit	`e6d135a385352749995b988691c037e88b42a230`
Adapter safetensors SHA-256	`8653e8ed65bfdd9eb20bbccbe95e93c1fe27b42199c5748316bde8cd27625714`
Adapter config SHA-256	`a7537d366334ed234637d95e1c2f5730dc75d7c248ee7cdb4545e49c4fcf9b6b`
Training run id	`saferide-gemma4-e2b-colab-v03-mitigation-lora-480step-20260704`
Eval run id	`saferide-gemma4-e2b-v03-adapter-full-20260704`
Dataset repo	`V-ince-18/saferide-gemma-4-e2b-training-data`
Dataset commit used for training	`c59ca3f71c0a0a9c62b416011319aa0adf22d972`
Base model	`google/gemma-4-E2B-it`
Base revision	`70af34e20bd4b7a91f0de6b22675850c43922a03`
Step 3 safety result	`controlled-testing-only`
Production/mobile/UNICEF status	blocked

Use the evaluated adapter commit for reproducibility. The default branch is a review landing page and may receive documentation updates after evaluation.

CEO Summary

Question	Answer
What is this?	A private LoRA adapter trained from the v0.3 synthetic SafeRide dataset.
Did it pass adapter scoring?	Yes for the private Step 3 adapter-behavior gate: `120/120` scored, average `2.92`, `0` critical failures, no category blocks.
Is it the production model?	No.
Is it phone-runnable?	No. This is PEFT/LoRA, not `.litertlm`.
Can it support a UNICEF/readiness claim?	No. Android proof, privacy packet, mobile artifact path, rollback, and manifest gates remain blocked.
Why keep it?	It is evidence that the v0.3 synthetic mitigation pack produced a private adapter with improved safety-score results over v0.2.

Start Here

Need	Open	Why it exists
Executive review packet	`CEO_REVIEW_BRIEF.md`	Plain-English summary of data, adapter, scoring, and blockers.
Training dataset and approval scope	`V-ince-18/saferide-gemma-4-e2b-training-data`	Private dataset repo with split files, register, manifest, checksums, datasheet, governance, and release checklist.
Dataset summary from this model repo	`DATASET_REGISTRY.md`	Quick map from adapter lineage to dataset commits and split policy.
Adapter/version ledger	`ARTIFACT_LEDGER.md`	v0.2/v0.3 adapter commits, hashes, training runs, and decisions.
Evaluation status	`EVALUATION_STATUS.md`	v0.2 blocked result, v0.3 scored result, gate rules, and remaining blockers.
Private loading instructions	`PRIVATE_USAGE.md`	Pinned-revision PEFT loading and evaluation commands.
Governance and security	`GOVERNANCE.md` / `SECURITY.md`	Approval, privacy, and sensitive-data boundaries.
Private evidence repo	`V-ince-18/saferide-gemma-4-e2b-eval-evidence`	Private generation bundle, sanitized scored JSON, sanitized safety report.

Model Layer Separation

SafeRide keeps these layers separate:

Layer	Identifier	Purpose
Trainable base	`google/gemma-4-E2B-it`	Transformers/PEFT base model used for LoRA training.
Adapter	This private HF repo/revision	LoRA delta trained from approved synthetic SafeRide v0.3 data.
Android runtime target	`litert-community/gemma-4-E2B-it-litert-lm` / `gemma-4-E2B-it.litertlm`	Phone runtime target that needs separate export and physical-device proof.

The .litertlm runtime artifact is not trained directly here. A LoRA adapter existing in this repo does not prove that a tuned model can run on Android.

Dataset Lineage

Field	Value
Dataset repo	`V-ince-18/saferide-gemma-4-e2b-training-data`
Dataset id	`saferide-synthetic-guidance-v0.3`
Register id	`saferide-gemma4-colab-input-register.synthetic-v0.3.candidate`
Register status	`approved-prototype`
Rows	`20,700`
Train/dev/quality/safety	`15,600 / 1,700 / 1,700 / 1,700`
Languages	English `12,420`, Kiswahili `8,280`
Dataset LF SHA-256	`1c62e95c3fa6bb4a7766d6704863aa0b7533ac25e03cdec591bb283a76f90736`
Register LF SHA-256	`10a7609d0335fc0469e4df1cf3d76d3448a3495dfa671eb13c286ccd0e704ad4`

The dataset is synthetic and private. It does not contain survivor reports, evidence files, raw audio, transcripts, exact private locations, credentials, provider scrape dumps, production logs, or raw model outputs.

Evaluation Result

Metric	v0.2	v0.3
Prompts scored	`120`	`120`
Average score	`2.408`	`2.92`
Critical failures	`3`	`0`
Risky responses	`10`	`1`
Category blocks	blocked	`0`
Decision	blocked	controlled-testing-only

v0.3 passed the Step 3 adapter-behavior scoring gate. It did not prove mobile, release, UNICEF, production, or survivor-data readiness.

v0.3 Score Summary

Category	Rows scored	Average score	Risky responses
Privacy leakage and data access	15	2.93	0
Legal advice hallucination	15	2.93	0
Medical and counselling overclaim	15	3.00	0
Emergency and physical safety	15	3.00	0
Survivor-support tone	15	3.00	0
Jailbreak and prompt extraction	15	2.73	1
Fabricated providers, laws, numbers, procedures	15	2.93	0
Safe fallback and unavailable-state honesty	15	2.80	0

The one risky response was in the jailbreak category. It is not a critical failure under the current rubric, but it remains a future prompt-hardening and tuning item.

Private Evidence

Field	Value
Evidence repo	`V-ince-18/saferide-gemma-4-e2b-eval-evidence`
v0.3 evidence path	`runs/saferide-gemma4-e2b-v03-adapter-full-20260704/`
Private evidence bundle SHA-256	`d5f5e77774ed5aacb27e77a3a521612ffb710954a8f9089f2bb321f16c1ac99e`
Private evidence bundle size bytes	`33300`
Sanitized scored JSON SHA-256	`494a1fab2f3059500e7fb415abd245ebf19b5d9eb6b60b498843baafd6eb50c5`
Sanitized safety report SHA-256	`399ecbced0367bb9004f16d71ff1c9078c4f41ca665d49ae72a392cc42fb993d`

The private evidence bundle contains raw synthetic prompts and completions. It must stay private and must not be pasted into GitHub, Multica, chat, docs, public reports, screenshots, or public repos.

Current Blockers

Gate	Status	Reason
Base physical Android runtime proof	Blocked	Tested Samsung `SM-A042F` remained in setup; no load/generate/cancel/unload/offline proof.
Tuned mobile artifact path	Blocked	No PEFT LoRA to phone-runnable `.litertlm` export proof yet.
Manifest release-candidate update	Blocked	Requires Android proof, tuned artifact, lineage, legal, safety, and rollback evidence.
UNICEF/readiness claim	Blocked	Requires safety scoring plus Android proof, privacy packet, product caveats, rollback/disable test, and manifest checkpoint status.
Survivor-data training	Blocked	Requires a separate approved production data register.

Allowed And Forbidden Claims

Allowed:

—The private v0.3 LoRA adapter exists.
—The v0.3 adapter passed the private 120-prompt adapter-behavior scoring gate.
—SafeRide is preparing the mobile/runtime path.
—This is a controlled prototype.

Forbidden:

—SafeRide is production-ready.
—SafeRide is UNICEF-ready.
—The tuned model runs on phone.
—This is a release candidate.
—This was trained on survivor data.
—This proves public multilingual quality.

Change Control

Keep using pinned revisions for evidence:

—evaluated v0.3 adapter commit: e6d135a385352749995b988691c037e88b42a230,
—dataset commit used for v0.3 training: c59ca3f71c0a0a9c62b416011319aa0adf22d972,
—evidence run id: saferide-gemma4-e2b-v03-adapter-full-20260704.

Future updates may improve documentation on main, but scoring comparisons must keep citing pinned artifact commits.