Model based grading

Prompt Evaluation Workflow: Grading Systems

• Purpose of a Grader: To take model output and provide an objective signal (e. g. , a number or true/false) regarding the quality of the response.
• Common Scoring: Scores are frequently given on a scale of 1 to 10 (1 = low quality, 10 = high quality).

Three Types of Graders

• Code-Based Graders:
• Mechanism: Feed model output into a custom snippet of code.
• Checks: Programmatic verification of length (too long/too short), presence/absence of specific words, syntax validation (for JSON or code), and readability scores.
• Requirement: Must return a usable signal.

• Model-Based Graders:
• Mechanism: Feed model output into an additional model (a separate API request).
• Flexibility: Can evaluate general quality, instruction following, completeness, or adherence to specific criteria.
• Requirement: Must return an objective signal (usually 1 to 10).

• Human-Based Grading:
• Mechanism: A person manually evaluates the model output.
• Flexibility: Can evaluate for any imaginable metric.
• Drawback: Highly time-consuming and tedious.

Defining Evaluation Criteria

• Upfront Decision: Before grading, you must define exactly what aspects of the response you are focusing on.
• Example Criteria:
• Format: Is the output only Python, JSON, or regular expression, with no extra explanation?
• Syntax: Does the generated code/JSON have valid syntax (no typos)?
• Task Following: Does the model clearly address the user's task with generally accurate code?

Implementation Strategy

• Code Grader Use: Best for format and syntax checks (e. g. , ensuring valid JSON or Python).
• Model Grader Use: Best for subjective or complex checks (e. g. , general task following and accuracy).

Model Grader Implementation Details

• Prompt Engineering: The grading prompt must clearly set a role, state the task, list the model's solution, and provide specific directions on how to respond.
• Improving Score Quality: To avoid generic scores (like 6), ask the model to provide specific strengths, weaknesses, and reasoning. This forces a more concrete evaluation.
• Process:
• Call the additional grading model.
• Ensure the model returns a structured JSON object.
• Parse the returned JSON to extract the score and reasoning.

Final Workflow Step

• Objective Metric: After running the evaluation, calculate the average score across all test cases to establish a final, objective metric for the prompt's performance.

Takeaways

• The primary purpose of a grader is to provide an objective signal (e. g. , a score on a scale of 1 to 10) regarding the quality of model output.
• There are three main types of graders: Code-Based (verifies syntax and format), Model-Based (evaluates general quality and instruction following), and Human-Based (most flexible but time-consuming).
• Evaluation criteria (e. g. , format, syntax, task following) must be defined upfront before any grading begins.
• Code Graders are best suited for objective checks (like valid JSON or Python syntax), while Model Graders are best for subjective or complex checks (like overall accuracy).
• To improve the quality of scores from a Model Grader, the grading prompt must be highly structured and require the model to provide specific strengths, weaknesses, and reasoning.