File size: 6,165 Bytes
6fcd376
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
import ast
import inspect
import json
import multiprocessing
import sys
from concurrent.futures import ProcessPoolExecutor, as_completed

from tqdm import tqdm

from evalplus.data.humaneval import (
    HUMANEVAL_PLUS_VERSION,
    get_human_eval_plus,
    get_human_eval_plus_hash,
)
from evalplus.eval import is_floats
from evalplus.eval._special_oracle import _poly
from evalplus.evaluate import get_groundtruth

HUMANEVAL_TEST_TEMPLATE = """\
{imports}

{aux_fn}

def check(candidate):
    inputs = {inputs}
    results = {results}
    for i, (inp, exp) in enumerate(zip(inputs, results)):
        {assertion}
"""

HUMANEVAL_CROSSCHECK_TEMPLATE = """\
{aux_fn}

{ref_func}

def check(candidate):
    inputs = {inputs}
    for i, inp in enumerate(inputs):
        assertion(candidate(*inp), ref_func(*inp), {atol})
"""

ASSERTION_FN = f"""\
import numpy as np

{inspect.getsource(is_floats)}

def assertion(out, exp, atol):
    exact_match = out == exp

    if atol == 0 and is_floats(exp):
        atol = 1e-6
    if not exact_match and atol != 0:
        np.testing.assert_allclose(out, exp, atol=atol)
    else:
        assert exact_match
"""


def synthesize_test_code(task_id, entry_point, inputs, results, ref_func, atol):
    # dataset size optimization for large outputs
    if entry_point in (
        "tri",
        "string_sequence",
        "starts_one_ends",
        "make_a_pile",
        "special_factorial",
        "all_prefixes",
    ):
        return task_id, HUMANEVAL_CROSSCHECK_TEMPLATE.format(
            aux_fn=ASSERTION_FN,
            inputs=inputs,
            ref_func=ref_func.replace(f" {entry_point}(", " ref_func("),
            atol=atol,
        )

    # default settings
    imports = set()
    aux_fn = ASSERTION_FN
    assertion = f"assertion(candidate(*inp), exp, {atol})"

    # special case: poly
    if entry_point == "find_zero":
        imports.add("import math")
        aux_fn = inspect.getsource(_poly) + "\n"
        assertion = f"assert _poly(*candidate(*inp), inp) <= {atol}"

    return task_id, HUMANEVAL_TEST_TEMPLATE.format(
        imports="\n".join(imports),
        aux_fn=aux_fn,
        inputs=inputs,
        results=results,
        assertion=assertion,
    )


def deduplicate(inputs, results):
    assert len(inputs) == len(results)
    unique_input_strs = set([f"{x}" for x in inputs])

    new_inputs, new_results = [], []
    for inp, res in zip(inputs, results):
        inp_str = f"{inp}"
        if inp_str in unique_input_strs:
            new_inputs.append(inp)
            new_results.append(res)
            unique_input_strs.remove(inp_str)

    return new_inputs, new_results


def main():
    import argparse

    parser = argparse.ArgumentParser()
    parser.add_argument("--debug-tasks", nargs="+", default=[], type=int)

    args = parser.parse_args()

    if hasattr(sys, "set_int_max_str_digits"):
        sys.set_int_max_str_digits(int(10e8))

    plus_problems = get_human_eval_plus(mini=False)
    dataset_hash = get_human_eval_plus_hash()

    compatible_problems = {}
    expected_outputs = get_groundtruth(plus_problems, dataset_hash, [])

    # debugging: monitoring test code size
    id2bytes = {}

    n_workers = max(1, multiprocessing.cpu_count() // 4)
    with ProcessPoolExecutor(max_workers=n_workers) as executor:
        futures = []
        for task_id, plus_form in tqdm(plus_problems.items()):
            if args.debug_tasks and int(task_id.split("/")[-1]) not in args.debug_tasks:
                continue

            compatible_form = {}
            compatible_form["task_id"] = task_id
            compatible_form["prompt"] = plus_form["prompt"]
            compatible_form["canonical_solution"] = plus_form["canonical_solution"]
            compatible_form["entry_point"] = plus_form["entry_point"]
            compatible_problems[task_id] = compatible_form

            inputs = plus_form["base_input"] + plus_form["plus_input"]
            results = (
                expected_outputs[task_id]["base"] + expected_outputs[task_id]["plus"]
            )

            inputs, results = deduplicate(inputs, results)

            assert len(inputs) == len(results)
            atol = plus_form["atol"]

            simplified_prompt = ""
            for line in compatible_form["prompt"].split("\n"):
                if not line:
                    continue
                if '"""' in line or "'''" in line:
                    break
                simplified_prompt += line + "\n"

            futures.append(
                executor.submit(
                    synthesize_test_code,
                    task_id,
                    compatible_form["entry_point"],
                    inputs,
                    results,
                    simplified_prompt + compatible_form["canonical_solution"],
                    atol,
                )
            )

        for future in tqdm(as_completed(futures), total=len(plus_problems)):
            task_id, test_code = future.result()
            # syntax check of test_code
            ast.parse(test_code)
            id2bytes[task_id] = len(test_code.encode("utf-8"))
            compatible_problems[task_id]["test"] = test_code

    # print the top-10 largest test code
    print("Top-10 largest test code comes from problems (in megabytes):")
    for task_id, size in sorted(id2bytes.items(), key=lambda x: x[1], reverse=True)[
        :10
    ]:
        print(f"{task_id}:\t{size / 1024 / 1024:.2f}mb")

    if args.debug_tasks:
        for problem in compatible_problems.values():
            print("--- debugging:", problem["task_id"])
            print(problem["prompt"] + problem["canonical_solution"])
            test_code = problem["test"]
            if len(test_code) <= 2048 + 512:
                print(test_code)
            else:
                print(problem["test"][:1024], "...")
                print("...", problem["test"][-1024:])
    else:
        with open(f"HumanEvalPlus-OriginFmt-{HUMANEVAL_PLUS_VERSION}.jsonl", "w") as f:
            for problem in compatible_problems.values():
                f.write(json.dumps(problem) + "\n")


if __name__ == "__main__":
    main()