Characterize model scaling with additional training data

[forklady42]

Overview

Understand how model performance scales as the amount of training data increases. This will inform data collection priorities and set expectations for future training runs.

Tasks

• Define evaluation metric(s) to track (e.g., validation loss, MAE on charge density)
• Train models on increasing subsets of available data (e.g., 10%, 25%, 50%, 75%, 100%)
• Plot learning curves as a function of dataset size
• Identify whether the model is data-limited or compute-limited at current scale
• Summarize findings and recommend next steps for data acquisition if needed

Acceptance Criteria

• Scaling curves produced and documented
• Clear conclusion on whether more data is expected to yield meaningful gains