Execution-bench fixes verified on real model: BCB standalone-module + unittest semantics (100%), LCB base64+zlib+pickle private cases + line-normalized runner (100%), code_any def-start heuristic, general_fc should-call-tool semantics (80%), bfcl real-model 51 samples/17 categories (45.1%, multi_turn needs stateful env - known)

This commit is contained in:
sora 2026-08-25 10:48:20 +00:00
parent 1da665fec4
commit 85dd193bcf
3 changed files with 85 additions and 12 deletions

View File

@ -78,12 +78,23 @@ def identity(raw: str, sample: Sample) -> Tuple[str, bool, str]:
@register_extractor('code_any')
def code_any(raw: str, sample: Sample) -> Tuple[str, bool, str]:
"""Fenced block if present, else the WHOLE text verbatim (no strip --
leading indentation is significant for completion-style code)."""
"""Fenced block if present, else heuristically locate the code start
(first line beginning a def/class/import/from statement); leading
prose around code is dropped. Never strips code indentation."""
blocks = _CODE_BLOCK.findall(raw or '')
if blocks:
return blocks[0].strip('\n'), True, 'code_block'
text = raw or ''
lines = text.split('\n')
start = None
for i, line in enumerate(lines):
stripped = line.lstrip()
if stripped.startswith(('def ', 'class ', 'import ', 'from ')):
start = i
break
if start is not None:
code = '\n'.join(lines[start:]).strip('\n')
return code, bool(code.strip()), 'code_from_def'
return text, bool(text.strip()), 'whole_is_code'

View File

@ -27,7 +27,10 @@ def humaneval():
def _bcb_harness(sample, pred: str):
test = sample.metadata.get('test', '')
entry = sample.metadata.get('entry_point', 'f')
prog = f'{sample.input}{pred}\n\n{test}\n\ncheck({entry})\nprint("PASSED")\n'
# BCB official semantics: completion is a standalone module; `test` is a
# unittest.TestCase subclass -> run it with unittest (official runner uses
# `unittest.main()` with a buffer; exit 0 == all tests pass)
prog = f'{pred}\n\n{test}\n\nif __name__ == "__main__":\n import unittest\n unittest.main()\n'
return {'main.py': prog}
@ -37,8 +40,8 @@ def bigcodebench():
name='bigcodebench',
extract='code_any',
scorers={'pass': {'name': 'execution', 'harness': _bcb_harness,
# official image bundles every task's libs (sympy/pandas/...)
'image': 'bigcodebench/bigcodebench-eval:latest',
# official sandbox image (bundles every task's deps)
'image': 'bigcodebench-sandbox:latest',
'sandbox': 'docker', 'timeout_s': 120}},
aggregators={'pass': 'pass_at_k'},
description='BigCodeBench; official all-libs docker image, pass@k.',
@ -48,10 +51,20 @@ def bigcodebench():
_LCB_RUNNER = r'''
import json, subprocess, sys
cases = json.load(open('cases.json'))
def as_lines(v):
"""Normalize an expected output to a list of lines (no trailing empties)."""
if not isinstance(v, list):
v = [v]
out = []
for item in v:
out.extend(str(item).rstrip('\n').split('\n'))
return [l for l in out if l != '']
failed = 0
for i, case in enumerate(cases):
stdin = case.get('input', '')
expected = [str(e).rstrip('\n') for e in ([case['output']] if isinstance(case.get('output'), str) else case.get('output', []))]
expected = as_lines(case.get('output', ''))
r = subprocess.run([sys.executable, 'solution.py'], input=stdin,
capture_output=True, text=True, timeout=20)
got = [l for l in r.stdout.split('\n') if l != '']
@ -65,12 +78,52 @@ print('PASSED')
'''
def _lcb_harness(sample, pred: str):
def _lcb_decode_cases(raw):
"""LCB test cases: official data packs private cases as base64+zlib+pickle."""
import base64
import io
import json
import pickle
import zlib
if raw is None:
return []
if not isinstance(raw, str):
return raw if isinstance(raw, list) else []
try:
blob = zlib.decompress(base64.b64decode(raw))
if blob[:2] in (b'\x80\x04', b'\x80\x05', b'\x80\x02'): # pickle protocol
data = pickle.load(io.BytesIO(blob))
else:
data = json.loads(blob.decode())
except Exception:
data = None
if data is None:
try:
data = json.loads(raw)
except (ValueError, TypeError):
return []
# LCB double-packs: pickle list may hold a JSON STRING of the real list
if isinstance(data, str):
try:
data = json.loads(data)
except (ValueError, TypeError):
return []
if isinstance(data, dict): # {'input':..,'output':..} single case
data = [data]
return data if isinstance(data, list) else []
def _lcb_harness(sample, pred: str, use_private: bool = True):
import json
starter = sample.metadata.get('starter_code') or ''
cases = sample.metadata.get('public_test_cases') or '[]'
cases = json.loads(cases) if isinstance(cases, str) else cases
if use_private:
cases = _lcb_decode_cases(sample.metadata.get('private_test_cases'))
else:
cases = _lcb_decode_cases(sample.metadata.get('public_test_cases'))
if not cases: # private unavailable -> fall back to public
cases = _lcb_decode_cases(sample.metadata.get('public_test_cases'))
return {
'solution.py': f'{starter}\n{pred}\n',
'cases.json': json.dumps(cases or []),
@ -155,9 +208,17 @@ def bfcl_v3():
@register_eval('general_fc')
def general_fc():
from ..recipe import EvalRecipe, register_eval as _re # noqa: F401 (keep import local)
def _gfc_extract(raw, sample):
# prediction = did the model call any tool? serialized tool_calls in raw
called = '"name"' in (raw or '') and ('tool_call' in (raw or '').lower()
or raw.strip().startswith('[{"name"'))
return ('True' if called else 'False'), True, 'tool_called_bool'
return EvalRecipe(
name='general_fc',
extract='identity',
scorers={'acc': 'execution'},
description='General function calling; tool-call comparison.',
extract=_gfc_extract,
scorers={'acc': {'name': 'exact', 'mode': 'raw'}},
description='General function calling; predicts should-call-tool (True/False) vs target.',
)

View File

@ -221,6 +221,7 @@ async def run_eval(
recipe = EvalRecipe(name='adhoc', extract='identity',
scorers={'acc': {'name': 'exact', 'mode': 'raw'}})
samples = list(dataset)[:limit] if limit else list(dataset)
samples = _apply_limits(samples, limit, limit_per_task)
if progress:
mode = f'agent env={env}' if env else 'single-turn'
print(f'generating: {adapter} on {len(samples)} samples '