Model based grading

Prompt Evaluation Workflow: Grading Systems

**I. Purpose of a Grader**

• A grader takes model output and provides an objective signal (e. g. , a number or true/false value).
• The most common signal is a score between 1 and 10 (10 = high quality, 1 = low quality).

**II. Types of Graders**

• Code-based Grader:
• Feeds model output into custom, authored code.
• Allows for programmatic checks (e. g. , output length, presence of specific words, JSON/code syntax validation, readability score).
• Requirement: Must return a usable signal.
• Model-based Grader:
• Feeds the original model output into an additional model API request.
• Offers high flexibility for evaluation (e. g. , general quality, instruction following, completeness).
• Requirement: Must return an objective signal (usually a score between 1 and 10).
• Human-based Grading:
• A person manually evaluates the model outputs.
• Highly flexible for any metric, but very time-consuming and tedious.

**III. Defining Evaluation Criteria**

• Evaluation criteria must be defined upfront.
• Example Criteria:
• Format: Output must be Python, JSON, or regular expression, with no additional explanation.
• Syntax: The generated code must have valid syntax (no typos).
• Task Following: The model must clearly address the user's task with generally accurate code and no major logic mistakes.

**IV. Mapping Criteria to Graders**

• Code Grader: Best for format and syntax checks (e. g. , ensuring the output is valid JSON or Python).
• Model Grader: Best for general response quality and task fulfillment due to its flexibility.

**V. Implementing a Model Grader**

• Prompt Design: The grading prompt must be detailed and set a clear role for the grader.
• Improving Score Quality: To get more concrete scores, the prompt should ask the model for:
• A score.
• A list of strengths.
• A list of weaknesses.
• Reasoning behind the score.
• Structured Output: Use techniques like pre-filled assistant messages and stop sequences to ensure the model returns a clean, parsable JSON object.

**VI. Final Evaluation Workflow**

• Process: Run Model → Get Output → Run Grader → Get Score/Reasoning.
• Objective Metric: Aggregate all individual scores to calculate an average score, providing a final, objective metric for prompt performance.

Takeaways

• A grader provides an objective signal (e. g. , a score between 1 and 10) to evaluate model output.
• Grader types include Code-based (programmatic checks like syntax), Model-based (using an additional LLM for quality), and Human-based (manual evaluation).
• Code Graders are ideal for format and syntax validation, while Model Graders are best suited for general response quality and task fulfillment.
• Implementing a Model Grader requires detailed prompt design, asking the model for structured output such as a score, strengths, weaknesses, and reasoning.
• The final evaluation workflow involves running the model, applying the grader, and aggregating individual scores to establish a final objective metric.