File size: 8,844 Bytes
fdf190d |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 |
import ast
import inspect
import json
import multiprocessing
import sys
from concurrent.futures import ProcessPoolExecutor, as_completed
from traceback import print_exc
from rich.console import Console
from rich.syntax import Syntax
from tqdm import tqdm
from evalplus.data.mbpp import (
MBPP_PLUS_VERSION,
get_mbpp,
get_mbpp_plus,
get_mbpp_plus_hash,
)
from evalplus.eval import is_floats
from evalplus.eval._special_oracle import (
MBPP_OUTPUT_NOT_NONE_TASKS,
MBPP_OUTPUT_SET_EQ_TASKS,
_digit_distance_nums,
_surface_Area,
)
from evalplus.evaluate import get_groundtruth
MBPP_TEST_TEMPLATE = """\
{imports}
{aux_fn}
inputs = {inputs}
results = {results}
for i, (inp, exp) in enumerate(zip(inputs, results)):
{assertion}
"""
MBPP_CROSSCHECK_TEMPLATE = """\
import numpy as np
from math import inf
{aux_fn}
{ref_func}
inputs = {inputs}
for i, inp in enumerate(inputs):
assertion({entry_point}(*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:
assert np.allclose(out, exp, rtol=1e-07, atol=atol)
else:
assert exact_match, f"out: {{out}}, exp: {{exp}}"
"""
def synthesize_test_code(task_id, entry_point, inputs, results, ref_func, atol):
# dataset size optimization for large outputs
if entry_point in ("combinations_colors", "freq_count", "get_coordinates"):
return task_id, MBPP_CROSSCHECK_TEMPLATE.format(
aux_fn=ASSERTION_FN,
inputs=inputs,
ref_func=ref_func.replace(f" {entry_point}(", " ref_func("),
entry_point=entry_point,
atol=atol,
)
# default settings
imports = set(["import numpy as np", "from math import inf"])
aux_fn = ASSERTION_FN
assertion = f"assertion({entry_point}(*inp), exp, {atol})"
# ================================================ #
# ============== special oracles ================= #
if entry_point in MBPP_OUTPUT_SET_EQ_TASKS:
aux_fn = f"""\
{inspect.getsource(is_floats)}
def assertion(out, exp, atol):
if atol == 0 and is_floats(exp):
atol = 1e-6
out = set(out)
exp = set(exp)
if out != exp and atol != 0:
assert np.allclose(out, exp, rtol=1e-07, atol=atol)
else:
assert out == exp, f"out: {{out}}, exp: {{exp}}"
"""
elif entry_point in MBPP_OUTPUT_NOT_NONE_TASKS:
aux_fn = f"""\
def assertion(out, exp, atol):
if isinstance(out, bool):
exact_match = out == exp
else:
exact_match = exp == (out is not None)
"""
elif entry_point == "surface_Area":
imports.add("import math")
aux_fn = (
inspect.getsource(_surface_Area)
+ "\n"
+ """\
def assertion(out, exp_0, exp_1, atol):
assert abs(out - exp_0) <= atol or abs(out - exp_1) <= atol
"""
)
assertion = f"assertion(surface_Area(*inp), exp, _surface_Area(*inp), {atol})"
elif entry_point == "digit_distance_nums":
aux_fn = (
inspect.getsource(_digit_distance_nums)
+ "\n"
+ """\
def assertion(out, exp_0, exp_1, atol):
assert out == exp_0 or out == exp_1
"""
)
assertion = f"assertion(digit_distance_nums(*inp), exp, _digit_distance_nums(*inp), {atol})"
# ============== special oracles ================= #
# ================================================ #
test_code = MBPP_TEST_TEMPLATE.format(
imports="\n".join(imports),
aux_fn=aux_fn,
inputs=inputs,
results=results,
entry_point=entry_point,
assertion=assertion,
)
return task_id, test_code
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()
console = Console()
if hasattr(sys, "set_int_max_str_digits"):
sys.set_int_max_str_digits(int(10e8))
plus_problems = get_mbpp_plus(mini=False)
dataset_hash = get_mbpp_plus_hash()
original_mbpp = get_mbpp()
compatible_problems = {}
expected_outputs = get_groundtruth(
plus_problems, dataset_hash, MBPP_OUTPUT_NOT_NONE_TASKS
)
# 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()):
# expected MBPP task_id is numbers directly
# i.e., "666" instead of "Mbpp/666"
# But in EvalPlus the task_id is "Mbpp/666"
task_id_int = int(task_id.split("/")[-1])
if args.debug_tasks and task_id_int not in args.debug_tasks:
continue
compatible_form = {
"task_id": task_id_int,
"code": plus_form["canonical_solution"],
"prompt": original_mbpp[str(task_id_int)]["prompt"],
"source_file": original_mbpp[str(task_id_int)]["source_file"],
"test_imports": original_mbpp[str(task_id_int)]["test_imports"],
"test_list": original_mbpp[str(task_id_int)]["test_list"],
}
compatible_problems[task_id_int] = compatible_form
inputs = (
plus_form["base_input"] + plus_form["plus_input"]
if len(plus_form["plus_input"]) > 0
else plus_form["base_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"]
futures.append(
executor.submit(
synthesize_test_code,
task_id_int,
plus_form["entry_point"],
inputs,
results,
compatible_form["code"],
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)
# ground-truth check
task = plus_problems[f"Mbpp/{task_id}"]
exec_code = (
task["prompt"] + "\n" + task["canonical_solution"] + "\n" + test_code
)
# run the code in a subprocess
def test():
try:
exec(exec_code, globals())
except Exception:
print_exc()
raise
p = multiprocessing.Process(target=test)
p.start()
p.join(timeout=20)
assert not p.is_alive(), f"Timeout for Mbpp/{task_id}!"
p.terminate()
p.join()
if p.exitcode != 0:
console.print(Syntax(exec_code, "python", line_numbers=True))
raise RuntimeError(f"Error for Mbpp/{task_id}")
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('"""\n' + problem["prompt"] + '\n"""\n' + problem["code"])
test_code = problem["test"]
if len(test_code) <= 1024:
print(test_code)
else:
print(problem["test"][:1024], "...")
print("...", problem["test"][-1024:])
else:
with open(f"MbppPlus-OriginFmt-{MBPP_PLUS_VERSION}.jsonl", "w") as f:
for problem in compatible_problems.values():
f.write(json.dumps(problem) + "\n")
if __name__ == "__main__":
main()
|