GuppyLM: Anyone can train this tiny AI

While AI models are becoming larger, more expensive and more opaque, GuppyLM goes in the opposite direction – by design. This small open-source project is a language model with only around 8.7 million parameters, far fewer than modern flagship models, and it identifies itself as a fish named Guppy. Guppy knows only life in an aquarium. The goal is not to compete with ChatGPT or other large models. Instead, GuppyLM is meant to show that an LLM does not have to be mysterious – and that training one does not necessarily require expert knowledge.

GuppyLM was trained on 60,000 synthetic conversations. In terms of content, the model is highly limited, but that is exactly what makes it remarkably consistent. Guppy speaks in short, lowercase sentences and does not understand human abstractions such as politics, money or telephones. Because this personality is firmly built into the model, Guppy always stays within its fish perspective. GitHub also offers a browser demo in which the model runs locally in the browser. Alternatively, the pretrained version can be launched via Colab or run locally with Python. Those who want to go a step further can even train their own mini LLM directly with the prepared Colab notebook – a browser-based programming environment.

The training process itself is relatively simple. The model is fed a large number of example pairs consisting of an input and a matching response. In the pretrained GuppyLM model, these include greetings, questions about food, water, light, sleep or the meaning of life – all from the perspective of a small fish. From these examples, the model learns which token should come next. Put simply, tokens are small text units into which words are broken down. During each training step, the model compares its prediction with the desired response and adjusts its internal weights accordingly. In this way, GuppyLM gradually learns how a fish is supposed to speak.