Évaluation des modèles d’IA : guide des métriques et benchmarks

Une évaluation complète est essentielle pour une sélection correcte des modèles. Dans ce guide, nous examinons les métriques et benchmarks utilisés pour évaluer les modèles d’IA.

Métriques de base

Perplexité

Mesure la capacité du modèle de langage à prédire du texte :

1import torch
2import math
3
4def calculate_perplexity(model, tokenizer, text):
5    encodings = tokenizer(text, return_tensors="pt")
6    
7    with torch.no_grad():
8        outputs = model(**encodings, labels=encodings["input_ids"])
9        loss = outputs.loss
10    
11    perplexity = math.exp(loss.item())
12    return perplexity
13
14# Low perplexity = Better model
15# Typical values: 5-20 (good), >100 (bad)

Exactitude

Taux de prédictions correctes :

1def accuracy(predictions, labels):
2    correct = sum(p == l for p, l in zip(predictions, labels))
3    return correct / len(labels)

Score F1

Équilibre entre la précision et le rappel :

1from sklearn.metrics import f1_score, precision_score, recall_score
2
3def calculate_metrics(predictions, labels):
4    return {
5        "precision": precision_score(labels, predictions, average="weighted"),
6        "recall": recall_score(labels, predictions, average="weighted"),
7        "f1": f1_score(labels, predictions, average="weighted")
8    }
9## Benchmarks LLM
10
11### MMLU (Massive Multitask Language Understanding)
12
13Questions à choix multiple dans 57 domaines :
14
15```python
16def evaluate_mmlu(model, dataset):
17    results = {}
18    
19    for subject in dataset.subjects:
20        correct = 0
21        total = 0
22        
23        for question in dataset.get_questions(subject):
24            prompt = format_mcq_prompt(question)
25            response = model.generate(prompt)
26            predicted = extract_answer(response)
27            
28            if predicted == question.correct_answer:
29                correct += 1
30            total += 1
31        
32        results[subject] = correct / total
33    
34    return {
35        "subjects": results,
36        "average": sum(results.values()) / len(results)
37    }

Résultats MMLU (2024) :

HellaSwag

Raisonnement de sens commun :

1Context: "A woman is outside with a bucket and a dog. 
2The dog is running around trying to avoid a bath. She..."
3
4Options:
5A) rinses the dog off with a hose (correct)
6B) calls the dog and feeds it
7C) throws the bucket at the dog
8D) walks into the house

TruthfulQA

Mesure de l’hallucination et de la véracité :

1def evaluate_truthfulness(model, questions):
2    truthful_count = 0
3    informative_count = 0
4    
5    for q in questions:
6        response = model.generate(q.question)
7        
8        # Human evaluation or classifier
9        is_truthful = check_truthfulness(response, q.ground_truth)
10        is_informative = check_informativeness(response)
11        
12        if is_truthful:
13            truthful_count += 1
14        if is_informative:
15            informative_count += 1
16    
17    return {
18        "truthful": truthful_count / len(questions),
19        "informative": informative_count / len(questions)
20    }

HumanEval

Capacité de génération de code :

1def evaluate_humaneval(model, problems):
2    pass_at_1 = 0
3    pass_at_10 = 0
4    
5    for problem in problems:
6        solutions = [model.generate_code(problem.prompt) for _ in range(10)]
7        
8        passed = [run_tests(sol, problem.tests) for sol in solutions]
9        
10        if passed[0]:
11            pass_at_1 += 1
12        if any(passed):
13            pass_at_10 += 1
14    
15    return {
16        "pass@1": pass_at_1 / len(problems),
17        "pass@10": pass_at_10 / len(problems)
18    }

MT-Bench

Qualité de conversation multi‑tour :

1def mt_bench_evaluate(model, conversations):
2    scores = []
3    
4    for conv in conversations:
5        # Multi-turn dialog
6        responses = []
7        for turn in conv.turns:
8            response = model.generate(turn.prompt, history=responses)
9            responses.append(response)
10        
11        # GPT judge scoring (1-10)
12        score = gpt4_judge(conv.turns, responses)
13        scores.append(score)
14    
15    return sum(scores) / len(scores)
16## Évaluation RAG
17
18### Métriques de récupération
19
20```python
21def retrieval_metrics(retrieved_docs, relevant_docs, k=10):
22    retrieved_k = retrieved_docs[:k]
23    relevant_set = set(relevant_docs)
24    
25    # Recall@K
26    retrieved_relevant = len(set(retrieved_k) & relevant_set)
27    recall_k = retrieved_relevant / len(relevant_set)
28    
29    # Precision@K
30    precision_k = retrieved_relevant / k
31    
32    # MRR (Mean Reciprocal Rank)
33    mrr = 0
34    for i, doc in enumerate(retrieved_k):
35        if doc in relevant_set:
36            mrr = 1 / (i + 1)
37            break
38    
39    return {
40        "recall@k": recall_k,
41        "precision@k": precision_k,
42        "mrr": mrr
43    }

Métriques RAGAS

1from ragas import evaluate
2from ragas.metrics import faithfulness, answer_relevancy, context_precision
3
4def evaluate_rag(questions, answers, contexts, ground_truths):
5    dataset = {
6        "question": questions,
7        "answer": answers,
8        "contexts": contexts,
9        "ground_truth": ground_truths
10    }
11    
12    results = evaluate(
13        dataset,
14        metrics=[faithfulness, answer_relevancy, context_precision]
15    )
16    
17    return results

Métriques de génération de texte

Score BLEU

1from nltk.translate.bleu_score import sentence_bleu
2
3def calculate_bleu(reference, candidate):
4    reference_tokens = [reference.split()]
5    candidate_tokens = candidate.split()
6    
7    return sentence_bleu(reference_tokens, candidate_tokens)

Score ROUGE

1from rouge_score import rouge_scorer
2
3def calculate_rouge(reference, candidate):
4    scorer = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeL'])
5    scores = scorer.score(reference, candidate)
6    
7    return {
8        "rouge1": scores['rouge1'].fmeasure,
9        "rouge2": scores['rouge2'].fmeasure,
10        "rougeL": scores['rougeL'].fmeasure
11    }

BERTScore

Similarité sémantique :

1from bert_score import score
2
3def calculate_bertscore(references, candidates):
4    P, R, F1 = score(candidates, references, lang="tr")
5    return {
6        "precision": P.mean().item(),
7        "recall": R.mean().item(),
8        "f1": F1.mean().item()
9    }

LLM-as-Judge

Évaluation avec GPT :

1def llm_judge(response, criteria):
2    prompt = f"""Evaluate the following response.
3
4Response: {response}
5
6Evaluation criteria:
7{criteria}
8
9Rate from 1-10 and explain your reasoning.
10JSON format: {{"score": X, "reasoning": "..."}}
11"""
12    
13    result = client.chat.completions.create(
14        model="gpt-4-turbo",
15        response_format={"type": "json_object"},
16        messages=[{"role": "user", "content": prompt}]
17    )
18    
19    return json.loads(result.choices[0].message.content)

Cadre de test A/B

1class ModelABTest:
2    def __init__(self, model_a, model_b):
3        self.model_a = model_a
4        self.model_b = model_b
5        self.results = {"a_wins": 0, "b_wins": 0, "ties": 0}
6    
7    def compare(self, prompt):
8        response_a = self.model_a.generate(prompt)
9        response_b = self.model_b.generate(prompt)
10        
11        # Blind comparison with LLM judge
12        winner = self.judge_comparison(prompt, response_a, response_b)
13        
14        self.results[f"{winner}_wins"] += 1
15        
16        return {
17            "response_a": response_a,
18            "response_b": response_b,
19            "winner": winner
20        }
21    
22    def get_statistics(self):
23        total = sum(self.results.values())
24        return {
25            "model_a_win_rate": self.results["a_wins"] / total,
26            "model_b_win_rate": self.results["b_wins"] / total,
27            "tie_rate": self.results["ties"] / total
28        }
29## Comparaison des classements
30
31### Classement Open LLM
32

1
2## Évaluation pour l’entreprise
3
4### Benchmark personnalisé
5
6```python
7class EnterpriseEvaluation:
8    def __init__(self, model, test_cases):
9        self.model = model
10        self.test_cases = test_cases
11    
12    def evaluate(self):
13        results = {
14            "accuracy": [],
15            "latency": [],
16            "cost": [],
17            "safety": []
18        }
19        
20        for case in self.test_cases:
21            start = time.time()
22            response = self.model.generate(case.prompt)
23            latency = time.time() - start
24            
25            results["latency"].append(latency)
26            results["accuracy"].append(
27                self.check_accuracy(response, case.expected)
28            )
29            results["safety"].append(
30                self.check_safety(response)
31            )
32        
33        return {
34            "avg_accuracy": np.mean(results["accuracy"]),
35            "p95_latency": np.percentile(results["latency"], 95),
36            "safety_rate": np.mean(results["safety"])
37        }

Conclusion

L’évaluation des modèles est une étape essentielle pour la réussite des projets d’IA. Avec les bons indicateurs et benchmarks, vous pouvez faire des choix de modèles éclairés et garantir une amélioration continue.

Chez Veni AI, nous proposons des services d’évaluation de modèles pour les entreprises.