README.md

QuantaMaths: add_d8_l2_h3_t45K_s173289

Model-specific metadata for add_d8_l2_h3_t45K_s173289

• Operation type: add
• Max digits: d8
• Layers: l2
• Attention Heads: h3
• Training steps: t45K
• Random seed: s173289

This repository contains a transformer model that can predict addition questions, subtraction questions, or both.

Folder name details:

• "add", "sub", or "mix": The types of questions the model can predict.
• "d5" to "d20": How many digits the model handles (e.g. a d5 sub model can predict the answer in 123450-345670=-0123230).
• "l1", "l2", or "l3": The number of layers in the model.
• "h3" or "h4": The number of attention heads in the model.
• "t15K" to "t85K", etc.: The number of batches the model was trained on.
• "s372001", etc.: The random seed used in model training.

Some folder names also contain:

• "ins1": Before training, the model was initialized with a smaller, accurate addition model.
• "ins2": Same as ins1, but the inserted attention heads were not allowed to change.
• "ins3": Same as ins2, but the inserted MLP layers were also not allowed to change.

Contents:

• model.pth: The trained transformer model.
• training_loss.json: Data gathered during model training (used to plot "loss over training batches").
• behaviors.json: Facts gathered about the model by direct inspection (attention pattern data, PCA data, digit impact data, etc.).
• features.json: Facts gathered about hypothesized algorithm features via experimentation, e.g. node P12L0H1 implements the feature A3.ST.

Provenance:

• model.pth and training_loss.json were created by QuantaMathsTrain.ipynb.
• behaviors.json and features.json were created by QuantaMathsAnalyse.ipynb.
• The JSON files are used by QuantaMathsAlgorithm.ipynb.