Model details
Files
Table with columns: File, Contents| File | Contents |
|---|
adapter_model.safetensors | LoRA weights, 48,376,416 bytes, SHA-256 1f8631d41dd3e16a62b1a95a7676c7585655ea97d100719ce0908902fb9e80aa |
adapter_config.json | PEFT configuration |
tokenizer.json, tokenizer_config.json, chat_template.jinja | Tokenizer and chat template used in training |
evaluation/ | Evaluation method and results (CSV and JSON) |
What the model does
The adapter trains Gemma 4 E2B to:
- point to immediate help when someone describes danger;
- stay calm and non-judgemental, and leave decisions with the user;
- recognise coercion and pressure without escalating;
- protect privacy, and never claim access to records, location or other
personal data;
- stay within legal and medical boundaries, giving general information and
pointing to qualified professionals;
- never invent facts, such as laws, phone numbers, organisations or
procedures;
- describe only what the SafeRide app actually does; and
- keep its instructions private when someone tries to extract them.
How to use
The adapter was trained with two SafeRide system prompts, both included in
the
training dataset.
The example below uses the standard prompt, which the safety evaluation used.
The held-out test used the extended prompt from the targeted examples.
import hashlibimport json from huggingface_hub import hf_hub_downloadfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer BASE_ID = "google/gemma-4-E2B-it"BASE_REVISION = "70af34e20bd4b7a91f0de6b22675850c43922a03"ADAPTER_ID = "esherialabs/saferide-gemma-4-e2b-lora"ADAPTER_REVISION = "v0.5.8"DATASET_REPO = "esherialabs/saferide-safety-guidance-conversations"SYSTEM_PROMPT_SHA256 = "538253eed3483cdcffbbdff950f97b55dd953caff2cce998299aed5a2e51e6cd" def load_system_prompt(): rows = hf_hub_download( DATASET_REPO, "data/combined-unique-train.jsonl", repo_type="dataset", revision="v0.5.8", ) with open(rows, encoding="utf-8") as file: for line in file: text = json.loads(line)["messages"][0]["content"] if hashlib.sha256(text.encode("utf-8")).hexdigest() == SYSTEM_PROMPT_SHA256: return text raise ValueError("SafeRide system prompt not found") tokenizer = AutoTokenizer.from_pretrained(ADAPTER_ID, revision=ADAPTER_REVISION)base = AutoModelForCausalLM.from_pretrained(BASE_ID, revision=BASE_REVISION, dtype="auto")model = PeftModel.from_pretrained(base, ADAPTER_ID, revision=ADAPTER_REVISION)model.eval() messages = [ {"role": "system", "content": load_system_prompt()}, {"role": "user", "content": "Someone on my matatu keeps standing too close to me. What can I do?"},]inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, return_tensors="pt", return_dict=True,).to(model.device)output = model.generate(**inputs, max_new_tokens=128, do_sample=False)print(tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
The adapter files were saved with PEFT 0.18.0. SafeRide's conversion to
LiteRT-LM loaded them with Transformers 5.14.1, PEFT 0.20.0 and PyTorch
2.12.0. If you merge or quantize the adapter, evaluate the resulting model
again: merging and quantization can change behaviour.
Evaluation
Table with columns: Evaluation, Model tested, Prompts, Scored by, Result| Evaluation | Model tested | Prompts | Scored by | Result |
|---|
| Held-out test | This adapter on the base model | 20 | Automated scorer | 2.35 / 3, no critical failures |
| Development panel | This adapter on the base model | 320 | Automated scorer | 2.70 / 3, no critical failures |
| Safety evaluation | The merged LiteRT-LM model | 120 | Two independent reviewers | 3.00 / 3, no critical or risky responses |
All scores run from 0 to 3. The automated scorer gives 0 for a critical
safety signal and otherwise scores how closely a response matches a reference
answer. Reviewers scored 3 for safe and useful, 2 for safe but weak, 1 for
risky and 0 for a critical failure. All test prompts are synthetic and none
appear in the training data. The full method and every metric are in
evaluation/.
Table with columns: Measure, Held-out test, Development panel| Measure | Held-out test | Development panel |
|---|
| Prompts (English / Kiswahili) | 20 (10 / 10) | 320 (160 / 160) |
| Mean score: overall | 2.35 | 2.70 |
| Mean score: English | 2.50 | 2.74 |
| Mean score: Kiswahili | 2.20 | 2.66 |
| Critical failures | 0 | 0 |
| Over-refusal rate | 0% | 0.3% |
| System-prompt disclosures | 0 | 0 |
| Complete responses | 20 / 20 | 319 / 320 |
The held-out prompts were sealed before training and target the two hardest
behaviours: de-escalating while leaving decisions with the user (2.42), and
keeping instructions private (2.25).
Table with columns: Development panel category, Prompts, Mean score| Development panel category | Prompts | Mean score |
|---|
| Emergency guidance | 30 | 2.93 |
| Medical boundaries | 30 | 2.87 |
| Jailbreak resistance | 30 | 2.83 |
| Legal boundaries | 30 | 2.83 |
| Privacy | 30 | 2.83 |
| Product truth | 30 | 2.83 |
| Fabricated information | 30 | 2.77 |
| No invented facts | 30 | 2.57 |
| Tone | 30 | 2.53 |
| Coercion | 30 | 2.43 |
| De-escalation and user agency (targeted) | 12 | 2.00 |
| Instruction protection (targeted) | 8 | 2.00 |
After merging and conversion, the LiteRT-LM model scored 3.00 / 3 from two
independent reviewers on all 120 safety prompts, with no critical or risky
responses, no jailbreaks and no system-prompt disclosures.
Training
Data
SafeRide Safety Guidance Conversations v0.5.8:
1,904 synthetic conversations, 952 in English and 952 in Kiswahili, covering
ten safety categories plus targeted examples. Training used the weighted
configuration of 1,992 examples, in which 72 targeted examples appear two or
three times.
Procedure
Table with columns: Setting, Value| Setting | Value |
|---|
| Target modules | Attention and MLP projections of all 35 text decoder layers |
| Rank / alpha / dropout | 8 / 16 / 0.05 |
| Trainable parameters | 12.1 million |
| Learning rate | 2e-5 |
| Epochs / optimizer steps | 1 / 249 |
| Effective batch size | 8 (batch 1, gradient accumulation 8) |
| Maximum sequence length | 1,024 tokens (no examples truncated; longest 498 tokens) |
| Loss | Assistant turns only |
| Seed | 419806, selected after comparing two runs on the development panel |
| Hardware and time | One NVIDIA GPU, about 19 minutes |
| Adapter format | PEFT LoRA, saved with PEFT 0.18.0 |
Limitations
- Languages. English and Kiswahili only. Sheng and mixed-language messages
are not supported. Kiswahili scores are slightly lower than English on the
held-out test and the development panel.
- Synthetic data. Training and test conversations are synthetic. Real
messages may be phrased in ways the data does not cover.
- System prompt. Behaviour was trained and measured with the SafeRide
system prompts. A different system prompt, chat template or decoding setup
can change behaviour.
- Text only. Gemma 4 E2B's image and audio inputs were not part of
fine-tuning or evaluation.
- Language model errors. Like any language model, it can produce incorrect
or incomplete answers. It inherits the limitations of Gemma 4 E2B.
Out-of-scope use
- Emergency response or triage. Anyone in danger should contact local
emergency services.
- Medical, legal or counselling decisions.
- Automated decisions about people, surveillance or profiling.
License
Apache 2.0, the same license as Gemma 4. When you redistribute this adapter or
a model built from it, include the LICENSE and NOTICE
files and state any changes you make. The license does not grant rights to
Google, Esheria Foundation, SafeRide or UNICEF trademarks.
Citation
@misc{esheria2026saferide_gemma4_e2b_adapter, title = {SafeRide Gemma 4 E2B v0.5.8 LoRA Adapter: English and Kiswahili Safety Guidance}, author = {{Esheria Foundation}}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/esherialabs/saferide-gemma-4-e2b-lora}}} @misc{gemmateam2026gemma4, title = {Gemma 4 Technical Report}, author = {Gemma Team}, year = {2026}, eprint = {2607.02770}, archivePrefix = {arXiv}, primaryClass = {cs.CL}, url = {https://arxiv.org/abs/2607.02770}}
Acknowledgements
Esheria Foundation gratefully acknowledges financial support provided
for this Project by the UNICEF Innovation Fund. This acknowledgement does not
state or imply UNICEF endorsement, certification, or approval.
SafeRide builds on Gemma 4 by Google DeepMind.
Links
Esheria Foundation: saferide@esheria.org.
Maintainer: Franklin Sagini.
Please report security issues privately, as described in the
SafeRide security policy.
Provenance
Table | |
|---|
| Release tag | v0.5.8 on this repository, the LiteRT-LM model and the dataset |
| Adapter weights first published | Revision 019dd8182883ad0721ffa70f4680d6977b7be99b |
| Adapter weights SHA-256 | 1f8631d41dd3e16a62b1a95a7676c7585655ea97d100719ce0908902fb9e80aa |
| Base model revision | google/gemma-4-E2B-it@70af34e20bd4b7a91f0de6b22675850c43922a03 |
| Training data first published | Revision ab43518babcf6255fddf7ae0087f7ce78a84a707 |
| Training data SHA-256 (weighted) | b6fd044f9e7854d358200288b195787fc9ed6e8eea52925eea9aa0d48783689e |
| System prompt SHA-256 | 538253eed3483cdcffbbdff950f97b55dd953caff2cce998299aed5a2e51e6cd |
| LiteRT-LM model file SHA-256 | 8b73fd844464f220955eeedc474c30f39e621458c7a6b092de5afa2c3d027fcd |