GuppyLM: A 9M Parameter LLM Built from Scratch for Educational Purposes

✍️ OpenClawRadar📅 Published: April 16, 2026🔗 Source
GuppyLM: A 9M Parameter LLM Built from Scratch for Educational Purposes
Ad

What GuppyLM Is

GuppyLM is a tiny language model (~9M parameters) that pretends to be a fish named Guppy. It's built from scratch to demonstrate how language models work without requiring a PhD or massive GPU cluster. The project includes data generation, tokenizer creation, model architecture, training loop, and inference—all in about 130 lines of PyTorch code.

Architecture Details

  • Parameters: 8.7M
  • Layers: 6
  • Hidden dimension: 384
  • Heads: 6
  • FFN: 768 (ReLU)
  • Vocab: 4,096 (BPE)
  • Max sequence: 128 tokens
  • Norm: LayerNorm
  • Position: Learned embeddings
  • LM head: Weight-tied with embeddings

This is a vanilla transformer with no GQA, RoPE, SwiGLU, or early exit—as simple as possible.

Training and Data

The model trains on 60K synthetic conversations across 60 topics in about 5 minutes on a free Colab T4 GPU. The dataset (arman-bd/guppylm-60k-generic on HuggingFace) contains 57K training and 3K test samples in format: {"input": "...", "output": "...", "category": "..."}.

Topics include: greetings, feelings, temperature, food, light, water, tank, noise, night, loneliness, bubbles, glass, reflection, breathing, swimming, colors, taste, plants, filter, algae, snails, scared, excited, bored, curious, happy, tired, outside, cats, rain, seasons, music, visitors, children, meaning of life, time, memory, dreams, size, future, past, name, weather, sleep, friends, jokes, fear, love, age, intelligence, health, singing, TV, and more.

Ad

Personality and Output

Guppy speaks in short, lowercase sentences about water, temperature, light, vibrations, and food. It doesn't understand human abstractions like money, phones, or politics. Sample conversation:

You> hi guppy
Guppy> hi there. i just found a nice spot near the rock. the temperature feels nice.
You> are you hungry
Guppy> yes. always yes. i will swim to the top right now. i promise to eat all of it.
You> what is the meaning of life
Guppy> food. the answer is always food.

Quick Start

Chat with Guppy (no training needed): Downloads the pre-trained model from HuggingFace. Just run all cells in the notebook.

Train your own: Set runtime to T4 GPU, run all cells—downloads dataset, trains tokenizer, trains model, tests it. Upload to HuggingFace or download locally.

Chat locally:

pip install torch tokenizers
python -m guppylm chat

Example local interaction:

You> the cat is looking at you
Guppy> i hide behind the plant when the furry one comes.
You> it is raining outside
Guppy> i think rain is the best thing about outside.

Project Structure

guppylm/
├── config.py          # Hyperparameters (model + training)
├── model.py           # Vanilla transformer
├── dataset.py         # Data loading + batching
├── train.py           # Training loop (cosine LR, AMP)
├── generate_data.py   # Conversation data generator (60 topics)
├── eval_cases.py      # Held-out test cases
├── prepare_data.py    # Data prep + tokenizer training
└── inference.py       # Chat interface

This project is useful for developers who want to understand transformer architecture fundamentals without dealing with billion-parameter models. The complete implementation shows every piece from raw text to trained weights to generated output.

📖 Read the full source: HN LLM Tools

Ad

👀 See Also