pico-lm/pico-decoder-tiny

Pico Decoder Tiny

pico-decoder-tiny is the smallest (11M) model in the pico-decoder suite — a lightweight, LLaMA-style decoder-only transformer trained from scratch using pico-train. It is designed for transparent and reproducible research into the learning dynamics of language models, and is fully compatible with the pico-analyze toolkit for detailed interpretability analysis.

NOTE: The pico-decoder-tiny-1 branch contains the full commit history for the training run.

🔧 Model Details

Field	Value
Architecture	Decoder-only transformer (LLaMA-style)
Parameters	11M
Layers	12
Hidden Size	96
Feed Foward Size	384
Attention Heads	12
Key/Value Heads	4

📚 Training

• Dataset: pretokenized-dolma, English-only
• Training steps: 200,000
• Batch size: 1024
• Sequence length: 2048
• Optimizer: AdamW
• Learning rate schedule: Linear decay with warmup
• Compute: 16 A100-SXM4-80GB GPUs

📈 Evaluation and Analysis

This model supports fine-grained analysis using pico-analyze. This tool enables researchers to understand how learning unfolds over training, even at very small scales.

We also evaluate perplexity of the model on the pico-paloma-tinsy dataset.

📄 Citation

If you use pico-tiny or any other pico-decoder model in your research, please cite:

@software{pico2025,
    author = {Diehl Martinez, Richard},
    title = {Pico: A Lightweight Framework for Studying Language Model Learning Dynamics},
    year = {2025,
    url = {https://github.com/pico-lm}
}