NorthernTribe-Research/maasai-lang

A production-oriented language AI platform for English–Maasai translation, Maasai speech transcription, and low-resource NLP research. # Maasai Language Platform Production-oriented repository for English Maasai translation, Maasai speech transcription, culturally grounded language tooling, and the MLOps workflows required to train, publish, monitor, and operate those assets across GitHub, Hugging Face, Kaggle, and self-hosted GPU infrastructure. > Status > Active development. This repository is proprietary and all rights reserved. See `LICENSE` for the governing terms. ## Executive Summary This project is an end-to-end low-resource language AI platform focused on Maasai language preservation, accessibility, and operational delivery. It combines: - A user-facing Gradio application for translation, speech transcription, glossary-backed exploration, and curated cultural content. - A QLoRA fine-tuning stack for English Maasai machine translation and optional Maa generation-oriented tasks. - A Hub-centric publication model where Hugging Face stores the canonical dataset, model artifacts, Space bundle, and resumable training checkpoints. - A control-plane architecture where GitHub Actions orchestrates scheduled training while Kaggle or self-hosted runners provide GPU execution. ## Platform Scope ### Core Capabilities - Bidirectional English -> Maasai and Maasai -> English translation. - Maasai speech transcription using a dedicated ASR model. - Transcribe-then-translate workflows inside the same application surface. - Glossary-aware and culturally grounded user experiences for language preservation workflows. - Repeatable publication of model, dataset, and Space assets to Hugging Face. - Resumable daily training that restores checkpoints from Hugging Face and pushes updated checkpoints back during execution. ### Operating Model This repository follows a clear control-plane and data-plane split: - GitHub is the automation and source-control control plane. - Hugging Face is the durable system of record for dataset, model, Space, and optional run artifacts. - Kaggle and self-hosted GPU …