|
| 1 | +""" |
| 2 | +StudentEval is a dataset of 1,749 prompts for 48 problems, authored by 80 |
| 3 | +students who have only completed a one-semester Python programming class. |
| 4 | +Unlike many other benchmarks, it has multiple prompts per problem and multiple |
| 5 | +attempts by the same participant. |
| 6 | +
|
| 7 | +Web page: https://huggingface.co/datasets/wellesley-easel/StudentEval |
| 8 | +""" |
| 9 | + |
| 10 | +from bigcode_eval.base import Task |
| 11 | +from datasets import load_dataset |
| 12 | +from multiprocessing import cpu_count |
| 13 | +from concurrent.futures import ThreadPoolExecutor |
| 14 | +from tqdm import tqdm |
| 15 | +import tempfile |
| 16 | +import pandas as pd |
| 17 | +import numpy as np |
| 18 | +import subprocess |
| 19 | + |
| 20 | +_CITATION = """\ |
| 21 | +@misc{babe2023studenteval, |
| 22 | + title={StudentEval: A Benchmark of Student-Written Prompts for Large Language Models of Code}, |
| 23 | + author={Hannah McLean Babe and Sydney Nguyen and Yangtian Zi and Arjun Guha and Molly Q Feldman and Carolyn Jane Anderson}, |
| 24 | + year={2023}, |
| 25 | + eprint={2306.04556}, |
| 26 | + archivePrefix={arXiv}, |
| 27 | + primaryClass={cs.LG} |
| 28 | +}""" |
| 29 | + |
| 30 | +EXECUTION_TIMEOUT = 15 |
| 31 | + |
| 32 | + |
| 33 | +# Source: Chen at al. Evaluating Large Language Models of Code. 2021 |
| 34 | +def _estimator(n: int, c: int, k: int) -> float: |
| 35 | + """ |
| 36 | + Calculates 1 - comb(n - c, k) / comb(n, k). |
| 37 | + """ |
| 38 | + assert c <= n, "c must be less than n" |
| 39 | + if n - c < k: |
| 40 | + return 1.0 |
| 41 | + return 1.0 - np.prod(1.0 - k / np.arange(n - c + 1, n + 1)) |
| 42 | + |
| 43 | + |
| 44 | +def _run_assembled_program(item): |
| 45 | + """ |
| 46 | + Runs the program with a timeout. The result dictionary has a "success" key |
| 47 | + that is 1 on success and 0 on failure. It also includes keys necessary to |
| 48 | + group results (problem, prompt, and group) and report results for each |
| 49 | + subset of StudentEval. |
| 50 | + """ |
| 51 | + with tempfile.NamedTemporaryFile(mode="w", suffix=".py") as f: |
| 52 | + f.write(item["program"]) |
| 53 | + f.flush() |
| 54 | + try: |
| 55 | + result = subprocess.run( |
| 56 | + ["python3", f.name], |
| 57 | + timeout=EXECUTION_TIMEOUT, |
| 58 | + stdout=subprocess.DEVNULL, |
| 59 | + stderr=subprocess.DEVNULL, |
| 60 | + stdin=subprocess.DEVNULL, |
| 61 | + ) |
| 62 | + exit_code = result.returncode |
| 63 | + except subprocess.TimeoutExpired: |
| 64 | + exit_code = 1 |
| 65 | + return { |
| 66 | + "problem": item["problem"], |
| 67 | + "prompt": item["prompt"], |
| 68 | + "group": item["group"], |
| 69 | + "success": 1 if exit_code == 0 else 0, |
| 70 | + } |
| 71 | + |
| 72 | + |
| 73 | +def _get_group(item): |
| 74 | + """ |
| 75 | + These boolean flags are mutually exclusive in the dataset. We turn them into a |
| 76 | + a string for easy grouping with Pandas. |
| 77 | + """ |
| 78 | + if item["is_first_success"]: |
| 79 | + return "First Success" |
| 80 | + if item["is_last_success"]: |
| 81 | + return "Last Success" |
| 82 | + if item["is_first_failure"]: |
| 83 | + return "First Failure" |
| 84 | + if item["is_last_failure"]: |
| 85 | + return "Last Failure" |
| 86 | + return None |
| 87 | + |
| 88 | + |
| 89 | +class StudentEval(Task): |
| 90 | + DATASET_PATH = "wellesley-easel/StudentEval" |
| 91 | + |
| 92 | + def __init__(self): |
| 93 | + self.stop_words = ["\ndef", "\nclass", "\nif", "\nprint"] |
| 94 | + self.requires_execution = True |
| 95 | + self.dataset = load_dataset(path=self.DATASET_PATH) |
| 96 | + # NOTE(Arjun Guha): Avoiding .filter so that we don't get a datasets |
| 97 | + # cache item on disk. |
| 98 | + self.dataset = [ |
| 99 | + item for item in self.dataset["test"] if _get_group(item) is not None |
| 100 | + ] |
| 101 | + |
| 102 | + def get_dataset(self): |
| 103 | + return self.dataset |
| 104 | + |
| 105 | + def get_prompt(self, doc): |
| 106 | + return doc["prompt"].rstrip() |
| 107 | + |
| 108 | + # For a task with tests, the reference solution is the suite of tests. |
| 109 | + def get_reference(self, doc): |
| 110 | + return { |
| 111 | + "prompt": doc["prompt"], |
| 112 | + "assertions": doc["assertions"], |
| 113 | + "problem": doc["problem"], |
| 114 | + "group": _get_group(doc), |
| 115 | + } |
| 116 | + |
| 117 | + def postprocess_generation(self, generation, idx): |
| 118 | + """Defines the postprocessing for a LM generation. |
| 119 | + :param generation: str |
| 120 | + code generation from LM |
| 121 | + :param idx: int |
| 122 | + index of doc in the dataset to which the generation belongs |
| 123 | + (not used for Humaneval-Task) |
| 124 | + """ |
| 125 | + prompt = self.get_prompt(self.dataset[idx]) |
| 126 | + generation = generation[len(prompt) :] |
| 127 | + return prompt + self._stop_at_stop_token(generation, self.stop_words) |
| 128 | + |
| 129 | + def process_results(self, generations, references): |
| 130 | + """Takes the list of LM generations and evaluates them against ground truth references, |
| 131 | + returning the metric for the generations. |
| 132 | + :param generations: list(list(str)) |
| 133 | + list of lists containing generations |
| 134 | + :param references: list({ "assertions": list(str), "problem": str }) |
| 135 | + list of reference solutions |
| 136 | + """ |
| 137 | + |
| 138 | + worklist = [] |
| 139 | + for generations, reference in zip(generations, references): |
| 140 | + # NOTE(Arjun Guha): This can be more efficient. At low temperature, we get lots of |
| 141 | + # repeated completions. So, this will end up running the same program repeatedly. |
| 142 | + # The original StudentEval code runs each generation once. |
| 143 | + for generation in generations: |
| 144 | + item = { |
| 145 | + "program": generation + "\n\n" + reference["assertions"], |
| 146 | + "prompt": reference["prompt"], |
| 147 | + "problem": reference["problem"], |
| 148 | + "group": reference["group"], |
| 149 | + } |
| 150 | + worklist.append(item) |
| 151 | + |
| 152 | + with ThreadPoolExecutor(max_workers=cpu_count() - 1) as executor: |
| 153 | + results_df = pd.DataFrame( |
| 154 | + list( |
| 155 | + tqdm( |
| 156 | + executor.map(_run_assembled_program, worklist), |
| 157 | + total=len(worklist), |
| 158 | + ) |
| 159 | + ) |
| 160 | + ) |
| 161 | + |
| 162 | + # Calculate pass@1 for each prompt |
| 163 | + results_df = results_df.groupby(["problem", "prompt", "group"]).agg( |
| 164 | + c=("success", np.sum), n=("success", "count") |
| 165 | + ) |
| 166 | + results_df.reset_index(inplace=True) |
| 167 | + results_df["pass1"] = results_df.apply( |
| 168 | + lambda row: _estimator(row["n"], row["c"], 1), axis=1 |
| 169 | + ) |
| 170 | + |
| 171 | + # Calculate mean pass@1 for each group |
| 172 | + results_df = results_df.groupby(["group"]).agg(pass1=("pass1", np.mean)) |
| 173 | + |
| 174 | + # Turn into JSON |
| 175 | + results_df.reset_index(inplace=True) |
| 176 | + results_df = results_df.to_dict(orient="records") |
| 177 | + return results_df |
0 commit comments