addGPT: Teaching Transformers to Add

A transformer encoder-decoder that learns integer addition, demonstrating algorithmic learning capabilities.

Model Description

• Architecture: Transformer encoder-decoder
• Parameters: ~2.7M (tiny config)
• Task: Integer addition (up to 5-digit numbers)
• Accuracy: >99% on held-out test set

Usage

# Download checkpoint
from huggingface_hub import hf_hub_download
checkpoint_path = hf_hub_download(repo_id="m-bano/addGPT", filename="ckpt_step12500.pt")

# Or use the provided script
# python download_checkpoint.py

See the GitHub repository for complete code and usage instructions.

Training Details

• Framework: PyTorch 2.0+
• Training Time: ~10 minutes on GPU
• Dataset: 6.4M randomly generated addition problems
• Optimizer: AdamW with cosine annealing

Configuration

n_emb: 512
n_heads: 8
n_blocks: 4
max_src_len: 5  # 5-digit operands

License

Apache 2.0