Model based grading

Overview of Graders

• A grader takes model output and returns an objective signal (commonly a number between 1-10, where 10 = high quality, 1 = low quality)
• Output can be any format: numbers, true/false, or other signals

Three Types of Graders

Code-Based Graders

• Author custom code snippets to evaluate model output
• Can perform programmatic checks: length validation, keyword verification, syntax validation for JSON/code
• Can implement complex checks like readability scores
• Must return an objective signal

Model-Based Graders

• Feed model output to an additional model via API call
• Provides flexibility to evaluate on any criteria: general quality, instruction-following, completeness, etc.
• Model must return a hard objective signal (usually 1-10 scale)
• Useful when evaluation criteria are complex or subjective

Human-Based Graders

• Have actual people evaluate responses
• Extremely flexible for any evaluation metric
• Major downside: time-consuming and tedious

Setting Evaluation Criteria

• Define upfront exactly what aspects of responses you will focus on
• Example criteria: correct format (Python/JSON/regex only), valid syntax, task completion and accuracy

Implementation Tips for Model Graders

• Write detailed prompts that set a role and clearly explain the evaluation task
• Ask model to provide strengths, weaknesses, and reasoning alongside scores
• Requesting reasoning prevents middling scores (like always scoring 6) by forcing the model to be more concrete
• Use JSON output format with pre-filled assistant messages and stop sequences for clean extraction

Measuring Overall Performance

• Calculate average score across all test cases to get an objective metric for prompt quality
• Use this metric to track improvements when iterating on prompts

Takeaways

• A grader evaluates model output and returns an objective signal (typically 1-10) to measure quality
• Three grader types exist: code-based (programmatic checks), model-based (API calls to evaluate), and human-based (manual evaluation)
• Define evaluation criteria upfront (format, syntax, accuracy, etc. ) before building graders
• Model-based graders work best when you request reasoning alongside scores, which prevents vague middle-range scoring
• Calculate average scores across test cases to objectively track prompt quality improvements