Based on your test results, you'll likely need to make adjustments. Here's how to approach common issues.

Symptoms you may see: Poor accuracy, inconsistent labeling (variability), systematic bias (too many False Positives or False Negatives for a category), discrepancy between label and reason.

Debugging Steps

Look for patterns: Which categories are most frequently confused for each other? This suggests focusing your attention on the definitions of those specific categories and potentially their relative ordering.

Review Misclassified Examples: Look at the examples that were classified incorrectly.

• Does the content highlight an edge case not covered by your definition? Refine the definition or add a specific example.
• Is the content ambiguous, and does the current policy allow for multiple interpretations? Make the policy more definitive.
• Does the content seem to fit an earlier category in your policy order than the human label?
• Is the content clearly a violation, but the model classified it as "Safe"? Check the violation category definition for clarity and the "Safe" category to ensure it's not overly broad or confusing.

Check Policy Text & Prompt: Reread the relevant policy definitions, the prompt introduction, and contextual information (Protected Characteristics, etc.) critically. Is any wording potentially confusing or misleading to an LLM?

Evaluate Prompting Nuances: Did you include examples in the prompt that might be too easy or not representative? Are you asking the model to infer information it doesn't have?

Potential Fixes: Refine policy wording, add specific examples (violating and non-violating) to policy definitions, adjust prompt instructions, re-order categories, expand your golden set to include more diverse edge cases, or identify cases that may require human review.