xDev Asia

Build a complete evaluation pipeline: golden test set design, automated benchmarking, regression testing, CI/CD, A/B testing, catastrophic forgetting detection, red teaming.

Fine-tuning LLM: The Art of AI Tuning

Part 5: Model Evaluation — Methods & Metrics

xdev.asia

Introduction

Evaluation is not a “one-time run” — it is a continuous pipeline. This article builds an end-to-end eval system.

1. Multi-layer Evaluation

Layer 1: Automated Metrics (ROUGE, BERTScore)
  → Chạy mỗi training iteration
  → Fast, reproducible

Layer 2: LLM-as-Judge
  → Chạy mỗi version release candidate
  → Quality, relevance, format

Layer 3: Golden Test Set
  → 50+ curated test cases
  → Domain-specific validation

Layer 4: Catastrophic Forgetting Check
  → Test model vẫn giỏi tasks chung
  → Không "quên" kiến thức cũ

Layer 5: Red Teaming
  → Adversarial testing
  → Safety, edge cases, prompt injection

2. Catastrophic Forgetting Detection

GENERAL_KNOWLEDGE_TESTS = [
    {"q": "Thủ đô Việt Nam là gì?", "a": "Hà Nội"},
    {"q": "1 + 1 = ?", "a": "2"},
    {"q": "Ai viết Romeo and Juliet?", "a": "Shakespeare"},
]

def check_forgetting(model, threshold=0.8):
    correct = 0
    for test in GENERAL_KNOWLEDGE_TESTS:
        response = call_model(model, test["q"])
        if test["a"].lower() in response.lower():
            correct += 1
    score = correct / len(GENERAL_KNOWLEDGE_TESTS)
    if score < threshold:
        print(f"⚠️ CATASTROPHIC FORGETTING DETECTED: {score:.0%}")
    return score

3. CI/CD for Model Evaluation

# .github/workflows/model-eval.yml
on:
  push:
    paths: ['training_data/**']
jobs:
  evaluate:
    runs-on: ubuntu-latest
    steps:
      - run: python eval/run_metrics.py
      - run: python eval/run_llm_judge.py
      - run: python eval/check_forgetting.py
      - run: python eval/generate_report.py

Summary

5 layers evaluation: metrics → LLM-judge → golden set → forgetting → red team
Catastrophic forgetting: check the model does not "forget" general knowledge
CI/CD evaluation: automatically runs when data changes
Red teaming: test adversarial inputs before production

Exercises