Ygtopher/NelsonAI

Nelson AI is a self-evolving, trilingual (Kinyarwanda, English, French) language model trained entirely from scratch. It features a custom transformer architecture, offline tool-calling, live internet access, and a background evolution engine that continuously learns from conversations and web data. # Nelson AI 🇷🇼 > ⚠️ **WORK IN PROGRESS:** This artificial intelligence model is currently in active development. It is still learning and undergoing alignment fine-tuning. Responses may be inaccurate, experimental, or incomplete. **A Self-Evolving, Internet-Aware Trilingual AI — Trained From Scratch** Nelson is a custom-built GPT-style transformer trained natively on **Kinyarwanda, English, and French**. It features live internet access via tool-calling and a self-evolution engine that continuously improves from new conversations and web data. > **Language behavior:** Nelson detects which language you're using and responds in the same language. > It defaults to Kinyarwanda and has Kinyarwanda as its primary identity. --- ## 🚀 Full Training Pipeline ### Step 1 — Environment Setup (Python 3.11) *Note: PyTorch CUDA requires Python 3.9 - 3.12. Ensure you are using Python 3.11!* ```powershell # Create and activate a Python 3.11 virtual environment py -3.11 -m venv venv venv\Scripts\activate # Install PyTorch with CUDA 12.1 (for Quadro T2000) pip install torch torchvision torchaudio --index-url download.pytorch.org # Install other dependencies pip install -r requirements.txt ``` ### Step 2 — Download Multilingual Data Uses direct Parquet downloads and streaming to pull high-quality data from Wikipedia, OPUS-100, Wikitext, and TinyStories. ```powershell python data_collection/download_data.py ``` ### Step 3 — Clean and Balance Corpus Cleans text and strictly balances the final corpus to **60% Kinyarwanda, 25% English, 15% French**. ```powershell python data_collection/clean_text.py ``` ### Step 4 — Train the Tokenizer Learns a custom 32,000-token BPE vocabulary covering all three languages. ```powershell python tokenizer/train_tokenizer.py ``` ### Step 5 — Tokenize the Dataset Converts text into fast, binary `.npy` format for rapid training. ```powershell python training/tokenize_dataset.py ``` ### Step 6 — Train Nelson 🧠 Fully custom traini …