Accountability
Model manifest v1 · recorded 11 October 2026
Sikivu has trained no models. Every model we run was built by someone else and is used unmodified, pinned to a specific published revision. We have not fine-tuned, distilled, or otherwise derived a model from anything you have written.
We say this here because it is the claim most worth checking, and because it will not always be true. Our stated intention is to build a mental-health model for East African languages, trained only on data from people who explicitly opt in. If we do, we intend to publish what we trained on, when, and which model on this page before the model ships.
These run locally, inside your browser on the Sikivu web app. Text classified by them does not leave your device. The mobile app does not run any model on the device today: what it sends for classification is handled by the cloud models in section 3. Each is pinned to a published revision, and the file hash lets you verify that the file we serve is the file we describe.
13-class emotion classifier
Published by boltuix/bert-emotion. Quantised ONNX, 10.86 MB.
Mental health risk classifier (Normal/Anxiety/Depression/Suicidal)
Published by ourafla/mental-health-bert-finetuned. Quantised ONNX, 105.23 MB.
Sentence embeddings for pgvector semantic memory
Published by sentence-transformers/all-MiniLM-L6-v2. Quantised ONNX, 21.77 MB.
Some replies need a larger model than a phone can run. For those, your message is sent to Anthropic's Claude API. We call these models and do not train them. How Anthropic itself treats data sent through its API is set out on Anthropic's privacy page. See our Privacy Policy for what is sent and Subprocessors for the processing terms.
The on-device models listed above are English-only. There is no Kiswahili classifier running in Sikivu today, and no Sheng or code-switched classifier either. We state this plainly because we once claimed otherwise in error, and corrected it.
Input the on-device models cannot classify, including anything not detected as English, is sent to the cloud models in section 3. The practical consequence is that on-device features behave less well in Kiswahili than in English. Closing that gap is the reason we want to train a model at all.
This page is about models and training data. It does not report how many people use Sikivu or what they talk about. Aggregate usage statistics, if we ever publish them, will live on their own page, will cover only people who have specifically opted in, and will suppress any figure small enough to identify someone.
Something here look wrong? Tell us at privacy@sikivu.app.