言語タスクの指標：パープレキシティ、BLEU

Python で学ぶ LLM の入門

Jasmin Ludolf

Senior Data Science Content Developer, DataCamp

LLMのタスクと評価指標

言語タスクの評価指標

パープレキシティ

モデルが次の単語を正確かつ自信を持って予測する能力
パープレキシティが低い = 信頼度が高い

input_text = "Latest research findings in Antarctica show"

generated_text = "Latest research findings in Antarctica show that the ice sheet 
is melting faster than previously thought."


# Encode the prompt, generate text and decode it
input_text_ids = tokenizer.encode(input_text, return_tensors="pt")
output = model.generate(input_text_ids, max_length=20)
generated_text = tokenizer.decode(output[0], skip_special_tokens=True)

パープレキシティの出力

perplexity = evaluate.load("perplexity", module_type="metric")
results = perplexity.compute(predictions=generated_text, model_id="gpt2")

print(results)

{'perplexities': [245.63299560546875, 520.3106079101562, ....], 
 'mean_perplexity': 2867.7229790460497}

print(results["mean_perplexity"])

2867.7229790460497

ベースライン結果と比較

BLEU

翻訳の質を人手による参照訳と比較して測定
予測: LLMの出力物
参照: 人的参照

bleu = evaluate.load("bleu")


input_text = "Latest research findings in Antarctica show"
references = [["Latest research findings in Antarctica show significant ice loss due to 
               climate change.", "Latest research findings in Antarctica show that the ice 
               sheet is melting faster than previously thought."]]
generated_text = "Latest research findings in Antarctica show that the ice sheet is melting 
               faster than previously thought."

BLEU の出力

results = bleu.compute(predictions=[generated_text], references=references)
print(results)

{'bleu': 1.0, 
 'precisions': [1.0, 1.0, 1.0, 1.0], 
 'brevity_penalty': 1.0, 
 'length_ratio': 1.2142857142857142, 
 'translation_length': 17, 
 'reference_length': 14}

0-1スコア：1に近いほど高い類似度

練習しましょう！

Python で学ぶ LLM の入門