lm-eval-task/utils.py

import re
import evaluate as hf_evaluate
compute_ = hf_evaluate.load("code_eval")
    

try:
    test_cases = ["assert add(2, 3)==5"]
    candidates = [["def add(a,b): return a*b"]]
    results = compute_.compute(references=test_cases, predictions=candidates, k=[1])
except Exception as e:
    raise e



def pass_at_k_process(doc, results):
    global compute_
    
    result = {"pass@1": 0}

    # We support a single prediction currently
    prediction = results[0][0]
    if prediction.strip() == "" or prediction is None:
        return result
    
    assert doc["test_code"] is not None

    # Compute
    res = compute_.compute(
        references=[doc["test_code"]],
        predictions=[[prediction]],
        k=[1],
    )
    result["pass@1"] = res[0]['pass@1']

    return result

def build_predictions_chat_ds1000(
    resps: list[list[str]], docs: list[dict]
) -> list[list[str]]:
    out = list()
    for resp, _ in zip(resps, docs):
        out.append(list())
        for r in resp:
            filtered = extract_longest_def_solve_code_block(r)
            if filtered is not None:
                out[-1].append(filtered['content'])
            else:
                out[-1].append(r)
    return out


def extract_longest_def_solve_code_block(text: str) -> dict | None:
    """
    Extracts the longest code block from text.
    
    A code block starts with ```language (e.g., ```python) and ends with ```.
    Returns dict with 'language', 'content', and 'length', or None if none found.
    """
    # Matches ```language followed by content until ```
    pattern = r'```([a-zA-Z0-9]+)\s*(.*?)\s*```'
    matches = re.findall(pattern, text, re.DOTALL)
    if not matches:
        return None
    
    # Filter only those including the solve definition
    matches = [(lang, content) for lang, content in matches if 'def solve' in content]
    if len(matches) == 0:
        return None
          
    # Select longest by content length
    longest = max(matches, key=lambda x: len(x[1].strip()))
    language, content = longest
    
    return {
        'language': language,
        'content': content,
        'length': len(content)
    }