aliyudikko/dikko-ai-noma

A Hausa-only decoder-only Transformer language model built from scratch with PyTorch to support agricultural knowledge for Hausa-speaking communities. The project includes custom Hausa tokenization, causal language model pretraining, inference, RAG integration, and FastAPI deployment. # Dikko AI Noma **Dikko AI Noma** is a Hausa-language agricultural AI assistant designed to provide practical knowledge and guidance on **Noma (crop farming and agriculture)**. The system combines a custom lightweight **decoder-only Transformer language model** with **Retrieval-Augmented Generation (RAG)** to improve the accuracy and relevance of Hausa agricultural responses. ## Features * Hausa-focused agricultural AI assistant * Custom decoder-only Transformer language model * Custom SentencePiece tokenizer * Fine-tuned on Hausa agricultural instruction-response data * Retrieval-Augmented Generation using agricultural knowledge * ChromaDB vector database * LangChain-based RAG pipeline * FastAPI backend * Next.js frontend * Persistent local vector database * GPU/CPU-compatible inference * Support for Hausa characters such as `Ƙ`, `ƙ`, `Ɗ`, and `ɗ` ## System Architecture ```text Dikko AI Noma │ ▼ Next.js UI │ ▼ FastAPI Backend │ ▼ User Query │ ▼ LangChain RAG System │ ▼ ChromaDB Search │ ▼ data/rag/noma.txt │ ▼ Relevant Agricultural Context │ ▼ Custom Dikko AI Noma Model │ ▼ Hausa AI Response │ ▼ Next.js UI ``` ## Technology Stack ### AI / Machine Learning * Python * PyTorch * Custom Decoder-only Transformer * SentencePiece * Hugging Face * LangChain * ChromaDB * Hugging Face Sentence Transformers ### Backend * FastAPI * Uvicorn ### Frontend * Next.js * React * TypeScript ### Data * Hausa agricultural text * Hausa instruction-response fine-tuning data * `noma.txt` knowledge base * ChromaDB vector index ## Model Dikko AI Noma uses a lightweight custom decoder-only Transformer designed for experimentation and deployment with limited computational resources. Current model configuration: ```text Vocabulary Size: 8,000 Hidden Size: 256 Transformer Layers: 2 Attention Heads: 4 Context Length: Configurable Tokenizer: SentencePiece Language: Hausa Domain: Noma / Agriculture ``` The model is trained using causal language modeling and fine-tuned using Haus …