bc_eval.py

import dataclasses
import itertools
import os
import re
import tempfile
from collections import defaultdict
from pathlib import Path

import datasets
import evaluate
import numpy as np
from tqdm import tqdm

from .execution import execute_predictions

STDOUT_PARSE_REGEX = re.compile(r"^TEST-(.+)\.\.\.(.+)$", flags=re.MULTILINE)

_CITATION = """\
@article{orlanski2023measuring,
  title={Measuring The Impact Of Programming Language Distribution},
  author={Orlanski, Gabriel and Xiao, Kefan and Garcia, Xavier and Hui, Jeffrey and Howland, Joshua and Malmaud, Jonathan and Austin, Jacob and Singh, Rishah and Catasta, Michele},
  journal={arXiv preprint arXiv:2302.01973},
  year={2023}
}
"""

_DESCRIPTION = """\
This metric implements the evaluation harness for datasets translated with the BabelCode framework as described in the paper "Measuring The Impact Of Programming Language Distribution" (https://arxiv.org/abs/2302.01973).
"""


_KWARGS_DESCRIPTION = """
Calculates how many predictions per question pass a set of tests for the given problem.

Args:
    predictions: The list of predictions for each question to execute.
    languages: The language to use for each question.
    question_dicts: The information for each question.
    k: number of code candidates to consider in the evaluation (Default: [1, 10, 100])
    num_workers: number of workers used to evaluate the candidate programs (Default: 4).
    language_timeout: Timeouts to use for each language. If it is not set, will default to the one in the question dict (Default: None).
Returns:
    pass_at_k: dict with pass rates for each k
    results: dict with granular results of each unittest
Examples:
    >>> bc_eval = evaluate.load("bc_eval")
    >>> predictions = [["def add(a,b):\n\treturn a+b", "def add(a,b):\n\treturn a-b"]]
    >>> languages = ["Python"]
    >>> question_dicts = [{"test_code": "...", "entry_fn_name": "add","entry_cls_name":"Solution", "test_case_ids":["0","1"],"test_list":"..."}]
    >>> pass_at_k, results = code_eval.compute(predictions=predictions,languages=languages, question_dicts=question_dicts, k=[1, 2])
    >>> print(pass_at_k)
    {'pass@1': 0.5, 'pass@2': 1.0}
"""


_WARNING = """
################################################################################
                                  !!!WARNING!!!
################################################################################
The "bc_eval" metric executes untrusted model-generated code in Python.
Although it is highly unlikely that model-generated code will do something
overtly malicious in response to this test suite, model-generated code may act
destructively due to a lack of model capability or alignment.
Users are strongly encouraged to sandbox this evaluation suite so that it
does not perform destructive actions on their host or network. For more
information on how OpenAI sandboxes its code, see the paper "Evaluating Large
Language Models Trained on Code" (https://arxiv.org/abs/2107.03374).
Once you have read this disclaimer and taken appropriate precautions,
set the environment variable HF_ALLOW_CODE_EVAL="1". Within Python you can to this
with:
>>> import os
>>> os.environ["HF_ALLOW_CODE_EVAL"] = "1"
################################################################################\
"""

_QUESTION_INFO_KEYS = {
    "entry_fn_name",
    "entry_cls_name",
    "test_code",
    "test_list",
    "test_case_ids",
}


def make_file_and_command(qid, idx, pred, question, working_dir, timeout_override=None):
    file_name = f"pred.{question['extension']}"
    pred_dir = working_dir.joinpath(idx)
    pred_dir.mkdir(parents=True)
    pred_file = pred_dir.joinpath(file_name)
    with pred_file.open("w") as f:
        code = question["test_code"]
        code = question["test_code"].replace("PLACEHOLDER_CODE_BODY", pred)
        code = code.replace("PLACEHOLDER_FN_NAME", question["entry_fn_name"])
        code = code.replace("PLACEHOLDER_CLS_NAME", question["entry_cls_name"])
        f.write(code)

    commands = []
    for cmd, t in zip(question["commands"], question["timeouts"]):
        commands.append(
            {
                "timeout": t if timeout_override is None else timeout_override,
                "command": [c if c != "__FILENAME__" else file_name for c in cmd],
            }
        )

    return {"qid": qid, "idx": idx, "commands": commands, "cwd": pred_dir}


def _write_preds(
    preds,
    languages,
    language_timeout,
    question_dicts,
    tmp_dir,
):
    commands = []
    question_id_to_dict = {}

    for pred_list, l, q_dict in tqdm(
        zip(preds, languages, question_dicts), desc="Setup", total=len(preds)
    ):
        qid = len(question_id_to_dict)
        q_dict["language"] = l
        question_id_to_dict[qid] = q_dict
        for p in pred_list:
            commands.append(
                make_file_and_command(
                    qid=qid,
                    idx=str(len(commands)),
                    pred=p,
                    question=q_dict,
                    timeout_override=language_timeout.get(l),
                    working_dir=tmp_dir,
                )
            )

    return question_id_to_dict, commands


@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class BabelCodeEval(evaluate.Metric):
    def _info(self):
        list_keys = ["timeouts", "commands", "test_case_ids"]
        question_info_type = {
            k: datasets.Value(dtype="string")
            for k in _QUESTION_INFO_KEYS
            if k not in list_keys
        }
        question_info_type["test_case_ids"] = datasets.Value("string")
        question_info_type["commands"] = datasets.Sequence(datasets.Value("string"))
        question_info_type["timeouts"] = datasets.Sequence(datasets.Value("int32"))

        return evaluate.MetricInfo(
            # This is the description that will appear on the metrics page.
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            # This defines the format of each prediction and reference
            features=datasets.Features(
                {
                    "predictions": datasets.Sequence(datasets.Value("string")),
                    "languages": datasets.Value("string"),
                    "question_dicts": question_info_type,
                }
            ),
            homepage="https://github.com/google-research/babelcode",
            codebase_urls=["https://github.com/google-research/babelcode"],
            reference_urls=["https://github.com/google-research/babelcode"],
        )

    def _compute(
        self,
        predictions,
        languages,
        question_dicts,
        k=[1, 10, 100],
        num_workers=4,
        language_timeout=None,
    ):
        """Returns the scores"""

        if os.getenv("HF_ALLOW_CODE_EVAL", 0) != "1":
            raise ValueError(_WARNING)

        language_timeout = language_timeout or {}

        with tempfile.TemporaryDirectory() as tmp_dir:
            working_dir = Path(tmp_dir)
            question_map, pred_commands = _write_preds(
                preds=predictions,
                languages=languages,
                language_timeout=language_timeout,
                question_dicts=question_dicts,
                tmp_dir=working_dir,
            )

            results = execute_predictions(
                pred_commands,
                num_workers=num_workers,
                max_task_per_child=5,
                garbage_collection_freq=500,
            )

        all_results, q_passes, q_pct = _eval_predictions(results, question_map)

        assert len(q_passes) == len(q_pct)
        metrics = {}
        for lang in q_passes:
            metrics.update(
                _calculate_metrics(lang, q_passes[lang], q_pct[lang], k_vals=k)
            )
        return metrics, all_results


def _eval_single_pred(result, test_ids, num_expected_commands):
    test_case_results = {k: "MISSING" for k in test_ids}
    if len(result["results"]) != num_expected_commands:
        return "HAD_ERROR", 0, test_case_results

    last_result = result["results"][-1]
    if last_result.timed_out:
        return "TIMED_OUT", 0, test_case_results
    elif last_result.return_code != 0:
        return "HAD_ERROR", 0, test_case_results
    elif not last_result.stdout:
        return "HAD_ERROR", 0, test_case_results

    for match in STDOUT_PARSE_REGEX.findall(last_result.stdout):
        idx, test_result = match
        if idx in test_ids:
            if test_case_results[idx] != "MISSING":
                return "UNKNOWN_ERROR", 0, test_case_results
            test_case_results[idx] = test_result.strip()

    did_test_fail = False
    had_error = False
    num_passed = 0
    for r in test_case_results.values():
        if r == "PASSED":
            num_passed += 1
        elif r == "FAILED":
            did_test_fail = True
        else:
            had_error = True

    if had_error:
        return "HAD_ERROR", num_passed, test_case_results
    if did_test_fail:
        return "FAILED", num_passed, test_case_results

    return "PASSED", num_passed, test_case_results


def _eval_predictions(pred_results, question_map):
    out = []
    question_results = defaultdict(lambda: defaultdict(list))
    question_pct_pass = defaultdict(lambda: defaultdict(list))

    for p in pred_results:
        question = question_map[p["qid"]]
        test_cases = question["test_case_ids"]
        num_expected_commands = len(question["commands"])

        outcome, num_passed, test_case_results = _eval_single_pred(
            p, test_ids=test_cases, num_expected_commands=num_expected_commands
        )

        p["results"] = [dataclasses.asdict(r) for r in p["results"]]
        p["test_cases"] = test_case_results
        p["outcome"] = outcome

        lang = question["language"]
        question_results[lang][p["qid"]].append(num_passed == len(test_case_results))
        question_pct_pass[lang][p["qid"]].append(num_passed / len(test_case_results))

        out.append(p)

    return out, question_results, question_pct_pass


def _calculate_metrics(lang, q_passed, q_pcts, k_vals):
    assert len(q_passed) == len(q_pcts)

    num_samples = np.zeros(len(q_passed))
    num_correct = np.zeros(len(q_passed))
    pcts_passed = np.zeros(len(q_passed))
    for i, (k, v) in enumerate(q_passed.items()):
        num_samples[i] = len(v)
        num_correct[i] = sum(v)
        pcts_passed[i] = np.mean(q_pcts[k])

    out = {
        f"{lang}/pass@{k}": estimate_pass_at_k(num_samples, num_correct, k).mean()
        for k in k_vals
    }
    out[f"{lang}/mean_pct_pass"] = np.mean(pcts_passed)

    return out


def estimate_pass_at_k(num_samples, num_correct, k):
    """Estimates pass@k of each problem and returns them in an array."""

    def estimator(n: int, c: int, k: int) -> float:
        """Calculates 1 - comb(n - c, k) / comb(n, k)."""
        if n - c < k:
            return 1.0
        return 1.0 - np.prod(1.0 - k / np.arange(n - c + 1, n + 1))

    if isinstance(num_samples, int):
        num_samples_it = itertools.repeat(num_samples, len(num_correct))
    else:
        assert len(num_samples) == len(num_correct)
        num_samples_it = iter(num_samples)

    return np.array(
        [estimator(int(n), int(c), k) for n, c in zip(num_samples_it, num_correct)]
    )