Vibe Coding Discoverguppylm
View on GitHubA ~9M parameter LLM that talks like a small fish.
GuppyLM is a ~9M parameter vanilla transformer trained from scratch to chat as a fish persona. It includes data generation, BPE tokenizer training, training loop, inference, ONNX/WASM browser demo, and Colab notebooks, making it a compact reference for building your own tiny LLM.
Use Cases
Train a small LLM from scratch in a free Colab T4 in ~5 minutesLearn transformer architecture end to end (data, tokenizer, model, training, inference)Generate synthetic conversation datasets from templatesRun quantized LLM inference in the browser via ONNX + WASMExport a PyTorch model to quantized ONNXChat with a local tiny model from the CLIUse as a teaching template for building your own character LLM
Built With
- Language
- Python
- Frameworks
- PyTorch · Hugging Face Datasets · tokenizers · ONNX · WebAssembly · NumPy · tqdm
Tags
llm · tiny-llm · transformer · from-scratch · training · tokenizer · bpe · onnx · wasm · education · pytorch · synthetic-data · inference · colab · quantization