Model Validation

Run validation on a trained checkpoint.

Step 1: Identify Checkpoint

Path to checkpoint, run name + step, or "best"/"latest"

Step 2: Identify Validation Data

Validation set, test set, custom set, or multiple datasets

Step 3: Define Metrics

• Standard: loss, perplexity (for LM), accuracy
• Task-specific: BLEU, ROUGE, F1, precision, recall

Step 4: Run Validation

Load model, set to eval mode, iterate through data, compute metrics

Step 5: Report Results

Single checkpoint table, multi-checkpoint comparison, cross-dataset evaluation Compare to baselines, provide recommendations based on results