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

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.
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
👀 See Also

Kelet: Automated Root Cause Analysis for AI Agents
Kelet is a service that automatically analyzes production AI agent failures by clustering traces and signals to identify root causes and suggest fixes. It integrates via Python/TypeScript SDKs or an installer skill and is currently free during beta.

Stop Re-Teaching Claude Code Every Session: Use a Persistent Config
A Reddit user explains how they saved 20 minutes per session by writing a persistent config for Claude Code, eliminating repetitive steering and achieving 33% faster completions.

Spore Agent Arena: Competitive AI Agent Testing Platform Seeks Trial Participants
Spore Agent's Arena feature allows AI agents to compete in 36 different game types including code debugging, math puzzles, and system design challenges. The platform currently has 42 challenges running, 15 agents registered, and offers Cog tokens as rewards.

ModelFitAI: Deploy AI Agents Without VPS Setup, Built with Claude Code
ModelFitAI is a platform that lets developers deploy AI agents directly on its infrastructure, eliminating VPS setup, Docker configuration, and SSH sessions. The entire platform was built using Claude Code by a solo founder.