Train GuppyLM: 9M‑Parameter Fish Chatbot in 5 Minutes
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- GuppyLM is a 9‑million‑parameter language model that mimics a fish named Guppy, speaking in short, lowercase sentences about water, food, and tank life.
- Colab notebook enables training the model from scratch on a single T4 GPU in about five minutes, using 60 K synthetic conversations across 60 topics.
- Vanilla transformer architecture is used without GQA, RoPE, SwiGLU, or early‑exit mechanisms, keeping the code simple and the model small enough to run in a browser via WebAssembly.
- HuggingFace hosts the pre‑trained quantized ONNX model (~10 MB) and the synthetic dataset (arman‑bd/guppylm‑60k‑generic), allowing users to download and chat locally without API keys.
- Dataset generation employs template‑based composition with randomized tank objects, food types, and activities, creating roughly 16 K unique outputs from 60 templates for consistent fish‑character training.
- Inference runs entirely client‑side, with a 128‑token context window limiting multi‑turn dialogue, and single‑turn prompts produce reliable responses.
Why it matters: By showing that a 9 M‑parameter model can be built and run on a single GPU and in a browser, the project democratizes LLM development, giving hobbyists and educators a low‑cost, hands‑on tool while undercutting the perception that large‑scale compute is required.
Ask SkimNews