Nadhari AI Lab

AI research lab advancing frontier AI research and applications in Sub-Saharan Africa (and beyond).

[Alfaxad @alfxad]

Excited to finally launch a demo of Sara, a clinical workflow agent that can autonomously orchestrate end-to-end digital clinical tasks.

Think of it like Devin, for healthcare.

We built Sara by fine-tuning Google's MedGemma 1.5 (4B) to adapt medical tool-use capabilities.

On MedAgentBench, Sara outperforms models 2x - 200x its size and is SOTA on several tasks.

learn more: https://t.co/ugBuelp0Yu
demo: https://t.co/dGVkLcYZlT

[Nadhari AI Lab @NadhariAI]

https://t.co/ad792K0nZR

[Nadhari AI Lab @NadhariAI]

Today, we’re excited to introduce Soga (Preview), a Swahili voice AI app from Nadhari AI Lab .

Soga is powered by swa-csm-1b, the best open-source Swahili text-to-speech model. We fine-tuned Sesame's CSM-1B and achieved state-of-the-art performance in Swahili TTS, nailing Swahili prosody across various accents and dialects. swa-csm-1b is now available on Hugging Face.

Soga is now rolling out, with 6 minutes daily conversational access. You can now talk with Asha and Mosi, our Swahili AI personas.

Blogpost: https://t.co/yec36lgQYY

Soga: https://t.co/GZ3l1PJb6E

Model: https://t.co/XPcXjpO8Uc
Youtube: https://t.co/lJvrCLulf5

Soga will have a full release soon. In the coming months, we'll be making various improvements on both the model and user experience of the app.

We're excited for what's next.

[Nadhari AI Lab @NadhariAI]

@masterplan404 It’s a good start, working on expanding the dataset.

[Nadhari AI Lab @NadhariAI]

Introducing the Swahili Thinking Dataset.

Excited to release the first open-source chain-of-thought reasoning dataset for Swahili. Following OpenAI's Harmony response format, the dataset comprises of high-quality Swahili conversational AI responses along with their chain-of-thought.

While such datasets exist for English, French, Spanish, e.t.c, there were no publicly accessible high-quality reasoning datasets for African languages.

Until now!!

This dataset enables researchers and developers to build Swahili language models with native reasoning capabilities, advancing AI for 200+ million Swahili speakers.

Release announcement: https://t.co/sTeG2MrUTh

Dataset: https://t.co/pPPbdrxEGh

The dataset built upon the excellent work by @huggingface H4's Multilingual-Thinking dataset. We intend to extend the dataset in the future and we welcome further contributions to the dataset.

[Nadhari AI Lab @NadhariAI]

https://t.co/fG4ABvRTxc

[Alfaxad @alfxad]

I’m incredibly honored to build @NadhariAI with the support of @osventuresllc. Quite challenging and exciting work ahead, it’ll be fun. It’s time to build!! Much thanks to @jposhaughnessy and the whole team at OSV!!

[Nadhari AI Lab @NadhariAI]

Our page is finally live on @huggingface !! Kindly follow for our latest updates.
Link: https://t.co/ygFrrl1gkR

[Alfaxad @alfxad]

Website: https://t.co/6rcJRiaBYf

GitHub: https://t.co/xtdSSpTITG

[Nadhari AI Lab @NadhariAI]

Introducing Gemma-3n-Swahili :)

[Alfaxad @alfxad]

Introducing Gemma-3n-Swahili preview:
In the past two weeks I have been working on Swahili variants of Gemma-3n. The Gemma-3n models are multimodal and have a very efficient architecture enabling them to run locally on most devices, which is amazing.
However, we found out the model, while they have a fundamental Swahili language understanding and text generation, at times they make-up non-existent words and fail on basic Swahili prompts.
🧵

[Alfaxad @alfxad]

Introducing Gemma-3n-Swahili preview:
In the past two weeks I have been working on Swahili variants of Gemma-3n. The Gemma-3n models are multimodal and have a very efficient architecture enabling them to run locally on most devices, which is amazing.
However, we found out the model, while they have a fundamental Swahili language understanding and text generation, at times they make-up non-existent words and fail on basic Swahili prompts.
🧵

[Nadhari AI Lab @NadhariAI]

o4-mini and o3 for archeological discovery.

[Alfaxad @alfxad]

I just woke up to the news that my friend and I won Google’s @kaggle competition to develop language-specific variants of Gemma-2 models.

This was a really fun and exciting challenge, we worked on a suite of models we named Gemma 2 Swahili which excelled in Swahili understanding, technical and creative writing and we significantly improved cultural context understanding of the Gemma 2 models.

Looking forward to taking on more challenges like this one :)

[Alfaxad @alfxad]

the gemma 2 swahili website is finally live!! visit gemma-2-swahili(.)github(.)io🙃

[Alfaxad @alfxad]

I’m having a lot of fun pushing gemma 2 swahili to the limit. A lot of people have requested a chat playground, we will make a limited hugging face space soon. But I really wish we had gpus/tpus to provide wide access chat playground for anyone to try it.

[Alfaxad @alfxad]

over the holidays, inspired the gemma 2 JPN release, i teamed up with my friend to build swahili variants of gemma 2 models, we named them “gemma 2 swahili”.

we had a lot of fun, learnt a lot, and improved the models swahili understanding by huge margin, we hope to scale this work!!

we have submitted the notebooks @kaggle gemma competition. the links to the code, and models can be found below ;)