Métriques d’évaluation et références pour les LLM

Welcome! As developers working with Large Language Models (LLMs), it's crucial to know how to measure their performance. But how do we objectively say one LLM output is 'better' than another?

This lesson explores various metrics and benchmarks used to quantitatively assess the quality, relevance, and accuracy of LLM outputs.

LLMs are powerful, but they can sometimes:

• Hallucinate: Make up facts or provide incorrect information.
• Exhibit Bias: Reflect biases present in their training data.
• Be Inconsistent: Give different answers to similar prompts.
• Lack Relevance: Provide outputs that don't directly answer the prompt.

Evaluation helps us identify these issues, track improvements, and ensure our LLM applications are reliable.

The 'gold standard' for evaluating LLM outputs is often human judgment. Human evaluators can assess nuances like creativity, coherence, factual accuracy, and tone that automated metrics might miss.

However, human evaluation is:

• Slow: Requires significant time.
• Expensive: Involves paying human annotators.
• Subjective: Can vary between evaluators.

While invaluable, it's not always scalable for large datasets or continuous monitoring.

To overcome the limitations of human evaluation, developers rely on automated metrics. These are algorithms that compare an LLM's output to a 'ground truth' or reference answer and assign a score.

Automated metrics are:

• Fast: Can process large volumes quickly.
• Cost-effective: Once implemented, they are cheap to run.
• Consistent: Provide objective, reproducible scores.

While not perfect, they offer a scalable way to track performance.

For tasks like summarization or machine translation where an LLM generates text, BLEU (Bilingual Evaluation Understudy) and ROUGE (Recall-Oriented Understudy for Gisting Evaluation) are popular metrics.

• BLEU: Measures precision by counting matching n-grams (sequences of words) between the generated text and reference text. Higher scores mean more overlap.
• ROUGE: Measures recall, focusing on how many n-grams from the reference text appear in the generated text. Useful for summarization.

These metrics are good for comparing text similarity but don't always capture meaning or fluency perfectly.

Perplexity is a common metric used to evaluate language models themselves, rather than specific task outputs. It measures how well a probability model predicts a sample.

• A lower perplexity score indicates the model is 'less surprised' by the text, meaning it predicts the sequence of words more accurately.
• It's often used to assess a model's fluency and its ability to generate natural-sounding language.

Think of it as how confident the model is about the next word it will generate.

For tasks like information extraction, sentiment analysis (classification), or named entity recognition, metrics like Precision, Recall, and F1 Score are more suitable.

• Precision: How many of the items the LLM identified were actually correct? (e.g., of all entities identified, how many were true entities?)
• Recall: How many of the relevant items did the LLM identify? (e.g., of all true entities, how many did the LLM find?)
• F1 Score: The harmonic mean of Precision and Recall, providing a single score that balances both.

These are excellent for evaluating an LLM's ability to perform specific, structured tasks.

Beyond individual metrics, benchmarks are standardized collections of datasets and tasks designed to rigorously test LLM capabilities across various domains.

Examples include:

• MMLU (Massive Multitask Language Understanding): Tests knowledge in 57 subjects (e.g., history, law, math).
• GLUE (General Language Understanding Evaluation): A collection of 9 NLU tasks.
• HELM (Holistic Evaluation of Language Models): A broad framework evaluating models on robustness, fairness, and efficiency across many scenarios.

Benchmarks allow for fair comparisons between different LLMs.

While complex metrics exist, you can start with simple custom evaluation functions. Here's a Python example that checks if an LLM's response contains a specific keyword. This is useful for ensuring constraints or specific information are present.

Try running this example:

We've covered several metrics used to evaluate LLM outputs. It's important to choose the right metric for the task at hand.

Which of these metrics are commonly used to evaluate the quality of generated text by an LLM, specifically by comparing it to one or more reference texts?

Great job! In this lesson, you learned about the importance of evaluating LLM outputs and the different methods available:

• Human Evaluation: The gold standard but not scalable.
• Automated Metrics: Scalable and consistent. We looked at:
  • BLEU & ROUGE: For comparing generated text to references.
  • Perplexity: For assessing a language model's fluency.
  • F1 Score: For classification and extraction tasks.
• Benchmarks: Standardized tests (MMLU, GLUE) for broad LLM comparison.

Understanding these tools is key to building robust and reliable LLM applications!