Open models · Open data
Deep tech,
built in the open.
We train small, open AI models on Kenyan and East African data to solve problems our partners bring us. Then we publish the weights, the data, and what we learned, including the limits.
- Models and adapters published
- 4
- Open datasets
- 2
- Kenyan court submission pairs
- 14,384
- English and Kiswahili conversations
- 1,904
SafeRide · Public safety
Safety guidance that works offline, on the phone.
SafeRide helps women on public transport record what happened and understand their options. We fine-tuned Google’s Gemma 4 E2B on English and Kiswahili safety conversations and packaged it to run on the phone itself, so the app can give guidance with no connection.
- Base model
- google/gemma-4-E2B-it, pinned revision
- Method
- LoRA fine-tuning: rank 8, one epoch, learning rate 2e-5, 249 optimizer steps
- Training data
- 1,904 synthetic conversations (952 English, 952 Kiswahili), 1,992 weighted examples
- On the phone
- Merged and quantized to 8-bit weights, exported to LiteRT-LM: a 5.07 GB package with a 2,048-token context
- Tested
- Offline load, generation, restart and replay on a server; download, checksum, restart and chat on one Android handset with more than 8 GB of memory
- Internal check
- 20-example holdout: no critical failures; average usefulness 2.35 of 3 (English 2.50, Kiswahili 2.20)
- Model SafeRide Gemma 4 E2B · offline mobile model v0.5.8
esherialabs/saferide-gemma-4-e2b-v058-original-419806-litertlmThe adapter merged into Gemma 4 E2B and exported for on-device inference with LiteRT-LM. (opens on Hugging Face in a new tab) - Adapter SafeRide bilingual safety guidance adapter v0.5.8
esherialabs/saferide-gemma-4-e2b-v058-original-419806-adapter48 MB of LoRA weights for Gemma 4 E2B: cautious, agency-preserving guidance that avoids inventing facts. (opens on Hugging Face in a new tab) - Dataset SafeRide synthetic bilingual safety guidance dataset v0.5.8
esherialabs/saferide-gemma-4-e2b-v058-original-419806-training-dataThe exact conversations the adapter was trained on, with unique and weighted configurations. (opens on Hugging Face in a new tab) - ModelSuperseded SafeRide Gemma 3n E4B · GBV incident assistant
esherialabs/saferide-gemma-3nThe first SafeRide model, for text and audio. Kept for the record; the app now uses Gemma 4 E2B. (opens on Hugging Face in a new tab)
A research preview, not a support service. It is not approved for survivor-facing, emergency, medical, or legal use, and lower-memory phones are still being tested. The training data is synthetic. UNICEF’s Innovation Fund supported this work financially; that is not an endorsement.
LexChat · Llama Impact Accelerator
Court-ready drafting for legal aid, under supervision.
Legal aid clinics, paralegals, and community justice centres need first drafts they can trust to follow Kenyan practice. We fine-tuned Meta’s Llama 3.1 8B on Kenyan court submissions, then trained it further with reinforcement learning so drafts keep the structure, side, and jurisdiction an advocate expects. This is the work recognized at the Llama Impact Accelerator Kenya Demo Day.
- Base model
- meta-llama/Llama-3.1-8B-Instruct
- Step 1
- Supervised fine-tuning on Kenyan court submissions, from the dataset below
- Step 2
- GRPO reinforcement learning: six drafts per prompt, each scored against the others
- Rewards
- Section order, the side argued, Kenyan authorities only, length, citation format, and doctrine coverage, all checked by code with no judge model
- Efficiency
- QLoRA (4-bit base, LoRA rank 32) on a single 24 GB NVIDIA A10
- Output
- BLUF → Governing rules → Controlling holdings → Application → Relief sought
- Model LexChat 8B · Kenya court submissions generator
esherialabs/lexchat-Llama-3.1-8B-GRPODrafts structured submissions from facts, issues, side, and posture for the High Court, Court of Appeal, and Supreme Court. (opens on Hugging Face in a new tab) - Dataset Kenya court submissions Q/A
esherialabs/kenya-court-submissions-qa14,384 Kenya-only question and answer pairs across 14 areas of law, split 12,945 / 719 / 720 for training, validation, and test. (opens on Hugging Face in a new tab)
Not legal advice. Drafts are for advocates and supervised paralegals to review before anything is filed.
How we build
Small models.
Real problems.
Public-interest AI has to run where our partners are: on modest hardware, in local languages, often offline.
- 01
Start from the need
A real gap a partner described: a survivor with no data bundle, a clinic with no time to draft.
- 02
Build the data
Kenya-first and bilingual, with privacy kept out of the data by design.
- 03
Train small
Open base models, parameter-efficient methods, and hardware a public-interest team can afford.
- 04
Check it
Automated reward checks and held-out tests, with the results published, including the weak ones.
- 05
Publish openly
Weights, data, and model cards that state the limits, so others can review and build on the work.
Use the work
Open for review.
Open to build on.
Everything here is public. Download a model, test it in your context, and tell us where it falls short. Read each model card first: it states what the model is for and what it must not be used for.
The Finance Bill explainer helped the public explore Kenya’s Finance Bill 2024. This is a historical release, not guidance on current legislation.
View the announcement (opens in a new tab) Original bill (opens in a new tab)