Sweep + Analysis Framework

This release marks the next completed phase of the project, which includes:

• Sweep scripts (random + focused)
• YAML config refactor
• TrainingLossMonitor early stopping logic
• Sweep analysis via early-slope scoring
• WandB training log downloader
• Config + artifact uploader
• Full training rerun capability with logged configs

Example of a sweep. In practice, this release gives a framework for hyperparameter tuning which enables hill-climbing quality metrics.

A few artifacts include best_configs.txt generated from running

uv run experiments/analysis/sweep_analysis.py \
    --wandb-csv=/tmp/training_history_long.csv \
    --output-yaml=/tmp/best_configs.txt

Where training_history_long.csv given for reference (obtained via fetch_wandb_runs.py).