Chapter 4: Evaluate AI Systems (Practical Tooling)

+-----------------------------------------------------------------------------+
|                          LLM-AS-A-JUDGE EVALUATION                          |
+-----------------------------------------------------------------------------+
|  Question + Context + Candidate Output ---> Frontier Model (GPT-4o)         |
|                                                       |                     |
|                                                       v                     |
|  Output: Structured JSON Score (1-5) + Rigorous Step-by-Step Justification  |
+-----------------------------------------------------------------------------+

The LLM-as-a-Judge Paradigm

Instead of relying solely on expensive human annotators, production teams use a larger, more capable model (the "Judge") to evaluate smaller operational models:

1. Single-Answer Grading

The judge evaluates one output against a strict rubric.

2. Pairwise Comparison (Side-by-Side Arena)

The judge is given two outputs (Model A vs. Model B) blindly and selects the winner to compute Elo ratings.

Implementation: Production Faithfulness Evaluator

from pydantic import BaseModel, Field
from openai import OpenAI

class FaithfulnessVerdict(BaseModel):
    is_faithful: bool = Field(description="True if all claims are backed by context")
    hallucinated_claims: list[str] = Field(description="List of statements unsupported by context")
    confidence_score: float = Field(ge=0.0, le=1.0)
    reasoning: str

def evaluate_faithfulness(question: str, context: str, answer: str) -> FaithfulnessVerdict:
    client = OpenAI()
    prompt = f"""
You are an impartial evaluator auditing an AI assistant for hallucinations.
Verify if the ANSWER contains any factual claims not directly supported by the CONTEXT.

QUESTION: {question}
CONTEXT: {context}
ANSWER: {answer}
"""
    res = client.beta.chat.completions.parse(
        model="gpt-4o",
        response_format=FaithfulnessVerdict,
        messages=[{"role": "user", "content": prompt}],
        temperature=0.0
    )
    return res.choices[0].message.parsed

Avoiding Judge Biases

Chip warns of four prominent biases when using LLM judges:

1. Position Bias: Models tend to favor the first answer in pairwise evaluations (Mitigation: swap positions and average).
2. Verbosity Bias: Models favor longer, wordier answers even if they contain fluff (Mitigation: penalize unnecessary word length).
3. Self-Enhancement Bias: A model (e.g. GPT-4) tends to score its own outputs higher than Claude or Gemini outputs.