homericintelligence/scylla/extract-hyperparameters

Extract Hyperparameters

Locate and document all hyperparameters mentioned in research papers including learning rates, batch sizes, and model configurations.

When to Use

• Reproducing paper results
• Setting up model training configurations
• Comparing hyperparameter choices across papers
• Planning hyperparameter tuning experiments

Quick Reference

# Extract numeric values and parameters from papers
pdftotext paper.pdf - | grep -i "learning rate\|batch\|epochs\|weight decay\|dropout" | head -20
 
# Common pattern search
grep -E "\\b(lr|batch_size|epochs|momentum|dropout|layers)\\s*[=:]" config.py

Workflow

1. Find hyperparameter table: Look for "Table 1" or "Hyperparameters" section
2. Document architecture parameters: Layer sizes, activation functions, normalization
3. Extract training parameters: Learning rate, batch size, epochs, optimizers
4. Note regularization: Dropout, weight decay, batch normalization
5. Create configuration file: Translate to implementation format (YAML/JSON/Mojo)

Output Format

Hyperparameter documentation:

• Model architecture (layers, sizes, activations)
• Training parameters (LR, batch size, epochs)
• Optimizer configuration (type, momentum, decay)
• Regularization settings (dropout, L1/L2)
• Data preprocessing (normalization, augmentation)
• Hardware and precision (float32, float64)

References

• See prepare-dataset skill for data configuration
• See train-model skill for training implementation
• See /notes/review/mojo-ml-patterns.md for Mojo configuration patterns