Evaluation Methodology

Methods for evaluating Foundation Model outputs.

Evaluation Approaches

1. Exact Evaluation

| Method | Use Case | Example | |--------|----------|---------| | Exact Match | QA, Math | "5" == "5" | | Functional Correctness | Code | Pass test cases | | BLEU/ROUGE | Translation | N-gram overlap | | Semantic Similarity | Open-ended | Embedding cosine |

# Semantic Similarity
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity

model = SentenceTransformer('all-MiniLM-L6-v2')
emb1 = model.encode([generated])
emb2 = model.encode([reference])
similarity = cosine_similarity(emb1, emb2)[0][0]

2. AI as Judge

JUDGE_PROMPT = """Rate the response on a scale of 1-5.

Criteria:
- Accuracy: Is information correct?
- Helpfulness: Does it address the need?
- Clarity: Is it easy to understand?

Query: {query}
Response: {response}

Return JSON: {"score": N, "reasoning": "..."}"""

# Multi-judge for reliability
judges = ["gpt-4", "claude-3"]
scores = [get_score(judge, response) for judge in judges]
final_score = sum(scores) / len(scores)

3. Comparative Evaluation (ELO)

COMPARE_PROMPT = """Compare these responses.

Query: {query}
A: {response_a}
B: {response_b}

Which is better? Return: A, B, or tie"""

def update_elo(rating_a, rating_b, winner, k=32):
    expected_a = 1 / (1 + 10**((rating_b - rating_a) / 400))
    score_a = 1 if winner == "A" else 0 if winner == "B" else 0.5
    return rating_a + k * (score_a - expected_a)

Evaluation Pipeline

1. Define Criteria (accuracy, helpfulness, safety)
   ↓
2. Create Scoring Rubric with Examples
   ↓
3. Select Methods (exact + AI judge + human)
   ↓
4. Create Evaluation Dataset
   ↓
5. Run Evaluation
   ↓
6. Analyze & Iterate

Best Practices

1. Use multiple evaluation methods
2. Calibrate AI judges with human data
3. Include both automatic and human evaluation
4. Version your evaluation datasets
5. Track metrics over time
6. Test for position bias in comparisons