sagenanochat

A lightweight language model training and inference framework.

Features

• Language model training with multiple optimizers (AdamW, Muon, Lion, Sophia)
• Custom tokenizer with Rust-based BPE implementation
• Inference engine with KV caching
• Chat interface (CLI and web)
• Evaluation tasks (ARC, GSM8K, HumanEval, MMLU, SmolTalk)

Quick Start

# Install dependencies
pip install -r requirements_microlm.txt

# Train a model
python scripts/base_train.py

# Run chat interface
python scripts/chat_cli.py

Project Structure

• microlm/ - Core language model implementation
• nanochat/ - Training and inference utilities
• scripts/ - Training and evaluation scripts
• tasks/ - Benchmark evaluation tasks
• rustbpe/ - Rust-based tokenizer