Atlas Vision
Ask questions about a picture in English, Igbo, Yoruba or Hausa.
- Role
- Team member, SDK
- Type
- Team
- Year
- 2026
- Status
- Open source
- Category
- Hackathons
- Main stack
- Python, Hugging Face +3
Overview
Atlas Vision lets you ask questions about a picture in English, Igbo, Yoruba or Hausa and get the answer back in the same language. Show it a photo of two cats, ask in Igbo what is in it, and it answers in Igbo.
It joins three pieces. SigLIP2 is the vision model that looks at the image. N-ATLaS is the language model that understands the African languages, and it is gated on Hugging Face, so it needs approved access and a token. Atlas Vision's own trained weights, a projector and a LoRA adapter, connect the two: the projector turns the image into something the language model can read, and the adapter teaches it to answer questions about images.
We built Atlas Vision as Team HackFirst for the National AI Innovation Challenge 2026, a remote hackathon themed 'Build with N-ATLAS'. The challenge is run by Nigeria's Federal Ministry of Communications, Innovation and Digital Economy through NCAIR, with Awarri Technologies, and asks teams to build working solutions on N-ATLAS, Nigeria's open source multilingual language model. The team built an SDK, a Python package called testatlasvision with a client class named TestAtlasVision. We started from a working Colab notebook, kept its model logic and restructured it into proper modules, then made it safe to run behind a server, tested it, fixed the bugs that turned up and published it to PyPI.
Highlights
- •Functions for asking, chatting in any of the four languages, describing an image, plus plain text and translate without an image.
- •Sessions that remember earlier questions, owned by the caller so different users never get mixed together.
- •A lock so requests take turns on the one GPU, and image loading that can block remote URLs.
- •Named errors such as AuthenticationError, GatedModelAccessError and ImageLoadError, so an API team can map them to HTTP codes.
- •Hugging Face login handling that checks access before the 16 GB download.
- •34 offline tests that run without a GPU, plus a real GPU run in Colab, with CI on GitHub Actions.
A hard part
The problem
Yoruba and Hausa answers were being cut off mid sentence, and the bad token check passed with a fake token because the Hugging Face library quietly returned a cached file.
How it was solved
The translate back step now gets extra room so answers finish, and the token check no longer trusts the cache. Both fixes have tests so they stay fixed.
Built with
- Python
- Hugging Face
- PyPI
- Google Colab
- pytest
