01
It loses the thread
Answer quality degrades reliably after roughly seven to ten turns. Past that it invents facts and statistics and starts repeating its own phrasing.
R.01The models, released when they are real
KW5-Lite.
Our first step in Kiswahili. Published so it can be checked, not because it is finished.
KW5-Lite is a small conversational fine-tune: a 74 MB LoRA adapter trained with 4-bit QLoRA over Qwen2.5-1.5B-Instruct. It holds a short conversation in standard Kiswahili on modest hardware. It is not a foundation model, it was not trained from scratch, and it is not ready to be put in front of anyone making a decision that matters.
Why publish it at all
Because the alternative is a language-equity programme that nobody outside the building can inspect. The weights are open, the evaluation notes say where it breaks, and the next model gets measured against this one.
Read the model card on Hugging Face↗
LoRA adapter · Apache 2.0 · loads on top of the base model
Model card, as it stands
Class
Conversational fine-tune, Kiswahili
Base
Qwen2.5-1.5B-Instruct (Unsloth build)
Form
LoRA adapter, ~74 MB. Not a merged model
Training
4-bit QLoRA, supervised fine-tuning (TRL SFTTrainer)
License
Apache 2.0, open weights
Status
Research preview. Not for production
Programme
R.01 LUGHA, sovereign language models
No benchmark table, because we do not have one worth printing. Numbers appear here when they are measured and reproducible.
01Where it holds up
A short conversation, in good Kiswahili.
Treat it as a lightweight Kiswahili-speaking assistant: casual question and answer, language practice, general knowledge that is well attested. That is the envelope it was evaluated inside.
- 01Standard-register Kiswahili grammar, handled correctly across the evaluation set
- 02Consistent safety and refusal behaviour, including redirecting acute medical questions to a clinician rather than answering them
- 03Familiar cultural and factual ground, where the answer is well attested
- 04Short exchanges: a question, an answer, a follow-up
02Where it fails
Four ways
it breaks.
These are on the published card. A 1.5B model with an adapter on top has limits, and printing them is cheaper for everyone than letting someone discover them in front of a class.
02
It expects the formal register
Casual phrasing and Sheng-influenced input make it noticeably less reliable. It was trained on standard Kiswahili and it shows.
03
It fabricates when the question is open
On creative or open-ended prompts it produces plausible-sounding invented content, including invented linguistic detail. Confident and wrong is its failure mode.
04
It sometimes refuses things it shouldn't
The safety behaviour that makes it well-mannered also produces false refusals on benign requests.
Do not use it for
Medical, legal, or financial advice. Any decision a person or an institution has to stand behind. Any workflow where a confident invented answer would be acted on.
Nothing in the Regnant portfolio depends on this model for consequential work. Where a system needs a decision it can defend, it is built to show its evidence, not to be trusted on fluency.
03What it is a step toward
Language equity
is a model problem,
not a translation one.
A translation layer over an English model inherits every place that model has nothing to say. The programme is aimed at models that hold Kiswahili at the weights. KW5-Lite is the first thing we have published on that road, and it is small, adapted, and limited. The point of publishing it is that the next one has something to beat.
Get it
Weights and model card on Hugging Face ↗Loads as a PEFT adapter over the Unsloth build of the base model. When you apply the chat template, pass return_dict=True so the attention mask comes with it.
Related
Milkshake, the education ecosystem →The LUGHA programme →