Code Readability Evaluation

Evaluate the readability of code generated by LLM applications using Phoenix’s evaluation framework.

This tutorial shows how to classify code as readable or unreadable using benchmark datasets with ground-truth labels.

Key Takeaways:

• Download and prepare benchmark datasets for code readability evaluation
• Compare different LLM models (GPT-4, GPT-3.5, GPT-4 Turbo) for classification accuracy
• Analyze results with confusion matrices and detailed reports
• Get explanations for LLM classifications to understand decision-making

---

Notebook Walkthrough

We will go through key code snippets on this page. To follow the full tutorial, check out the full notebook.

Google Colab

colab.research.google.com

Download Benchmark Dataset

dataset_name = "openai_humaneval_with_readability"
df = download_benchmark_dataset(task="code-readability-classification", dataset_name=dataset_name)

Configure Evaluation

N_EVAL_SAMPLE_SIZE = 10
df = df.sample(n=N_EVAL_SAMPLE_SIZE).reset_index(drop=True)
df = df.rename(columns={"prompt": "input", "solution": "output"})

Run Code Readability Classification

Run readability classifications against a subset of the data.

from phoenix.evals import LLM, ClassificationEvaluator, async_evaluate_dataframe

CODE_READABILITY_PROMPT_TEMPLATE = """
You are evaluating whether a piece of code is readable or not.

[BEGIN DATA]
************
[Code]: {output}
************
[END DATA]

Is the code readable? Respond with "readable" or "unreadable".
"""

llm = LLM(provider="openai", model="gpt-4")
readability_evaluator = ClassificationEvaluator(
    name="code_readability",
    prompt_template=CODE_READABILITY_PROMPT_TEMPLATE,
    llm=llm,
    choices={"readable": 1.0, "unreadable": 0.0},
)

evals_df = await async_evaluate_dataframe(dataframe=df, evaluators=[readability_evaluator], concurrency=10)
readability_classifications = evals_df["code_readability_score"].str["label"].tolist()

Evaluate Results and Plot Confusion Matrix

Evaluate the predictions against human-labeled ground-truth readability labels.

true_labels = df["readable"].map({True: "readable", False: "unreadable"}).tolist()
choices = ["readable", "unreadable"]
print(classification_report(true_labels, readability_classifications, labels=choices))
confusion_matrix = ConfusionMatrix(
    actual_vector=true_labels, predict_vector=readability_classifications, classes=choices
)
confusion_matrix.plot(
    cmap=plt.colormaps["Blues"],
    number_label=True,
    normalized=True,
)

Get Explanations

When evaluating a dataset for readability, it can be useful to know why the LLM classified text as readable or not. The following code block runs the classifier with explanations included so that we can inspect why the LLM made the classification it did. There is a speed tradeoff since more tokens are being generated but it can be highly informative when troubleshooting.

readability_classifications_df = await async_evaluate_dataframe(
    dataframe=df.sample(n=5),
    evaluators=[readability_evaluator],
    concurrency=10,
)
readability_classifications_df["label"] = readability_classifications_df["code_readability_score"].str["label"]
readability_classifications_df["explanation"] = readability_classifications_df[
    "code_readability_score"
].str["explanation"]

Compare Models

Run the same evaluation with different models:

# GPT-3.5
llm_gpt35 = LLM(provider="openai", model="gpt-3.5-turbo")

# GPT-4 Turbo
llm_gpt4turbo = LLM(provider="openai", model="gpt-4-turbo-preview")