File size: 7,662 Bytes
66cf27f 440d35c 66cf27f 5f867e3 66cf27f df6fa5a 66cf27f 5f867e3 66cf27f 5f867e3 66cf27f 430d13c 66cf27f 76eacb1 5f867e3 440d35c df6fa5a edce79f df6fa5a | 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 | import os
import pytest
from workbench.environment import WorkbenchEnvironment
from workbench.models import WorkbenchAction
from workbench.tasks import ALL_TASK_IDS, REGISTRY, TASK_IDS, build_task
SEEDS = range(int(os.environ.get("WB_SEEDS", "25")))
@pytest.mark.parametrize("task_id", ALL_TASK_IDS)
def test_oracle_scores_one_and_wrong_answers_do_not(task_id):
if task_id.startswith("swe-repo-"):
pytest.skip("graded on workspace files; see test_repo_repair")
for seed in SEEDS:
task = build_task(task_id, seed)
assert task.grade(task.oracle) == 1.0, (task_id, seed, task.oracle)
assert task.grade("") == 0.0
assert task.grade("totally wrong") == 0.0
for wrong in task.wrong:
assert task.grade(wrong) < 1.0, (task_id, seed, wrong, task.oracle)
assert len(task.instruction) < 800
assert sum(len(c) for c in task.files.values()) < 60_000
@pytest.mark.parametrize("task_id", ALL_TASK_IDS)
def test_same_seed_is_deterministic(task_id):
a, b = build_task(task_id, 7), build_task(task_id, 7)
assert a.files == b.files and a.oracle == b.oracle and a.instruction == b.instruction
def test_episode_tools_and_budget():
env = WorkbenchEnvironment()
obs = env.reset(task_id="fin-loan-balance", seed=3)
assert not obs.done and "loan.txt" in obs.output
obs = env.step(WorkbenchAction(tool="read", path="loan.txt", lines=3))
assert "Principal" in obs.output
obs = env.step(WorkbenchAction(tool="read", path="loan.txt", start=0, lines=500))
assert "Principal" in obs.output and not obs.done
obs = env.step(WorkbenchAction(tool="grep", pattern="APR"))
assert "loan.txt:" in obs.output
obs = env.step(WorkbenchAction(tool="python", code="print(open('loan.txt').read().count('\\n'))"))
assert obs.output.strip().isdigit()
obs = env.step(WorkbenchAction(tool="read", path="../etc/passwd"))
assert obs.output.startswith("Error")
oracle = build_task("fin-loan-balance", 3).oracle
obs = env.step(WorkbenchAction(tool="submit", answer=oracle))
assert obs.done and obs.reward == 1.0
env.reset(task_id="math-grid-paths", seed=1)
for _ in range(20):
obs = env.step(WorkbenchAction(tool="ls"))
assert obs.done and obs.reward == 0.0 and obs.remaining_steps == 0
env.close()
def test_python_timeout():
env = WorkbenchEnvironment()
env.reset(task_id="math-grid-paths", seed=1)
obs = env.step(WorkbenchAction(tool="python", code="while True: pass"))
assert "killed" in obs.output or "CPU" in obs.output or obs.output
assert not obs.done
env.close()
def test_grep_catastrophic_regex_times_out_without_hanging():
env = WorkbenchEnvironment()
env.reset(task_id="sci-longest-orf", seed=1)
obs = env.step(WorkbenchAction(tool="grep", pattern="([ACGT]+)+X"))
assert obs.output.startswith("Error") and not obs.done
obs = env.step(WorkbenchAction(tool="grep", pattern="ATG", path="sequence.fasta"))
assert "matches" in obs.output
env.close()
def test_thousands_separators_and_cents_rule():
from workbench.grading import parts_score, close, exact
assert parts_score("APAC,10,450.25", [exact("APAC"), close(10450.25, abs_tol=0.01)]) == 1.0
from workbench.tasks.office import _usd_cents
assert _usd_cents(133.93, 1.12) == 15000 # 150.0016 USD rounds to the cap and complies
assert _usd_cents(45.46, 1.1) == 5001
def test_submit_that_raises_still_ends_episode():
env = WorkbenchEnvironment()
env.reset(task_id="math-grid-paths", seed=1)
env._task.grade = lambda answer: 1 / 0
obs = env.step(WorkbenchAction(tool="submit", answer="3"))
assert obs.done and obs.reward == 0.0
env.close()
def test_level_ids_and_plain_ids_agree():
for t in TASK_IDS:
assert build_task(t, 3).oracle == build_task(f"{t}@1", 3).oracle
if getattr(REGISTRY[t], "max_level", 2) >= 2:
assert build_task(f"{t}-l2", 3).oracle == build_task(f"{t}@2", 3).oracle
@pytest.mark.skipif("swe-repo-repair" not in TASK_IDS, reason="vendored libraries not present")
@pytest.mark.parametrize("level", [1, 2, 3, 4])
def test_repo_repair_pristine_fixes_and_broken_does_not(level):
for seed in range(int(os.environ.get("WB_REPO_SEEDS", "3"))):
task = build_task(f"swe-repo-repair@{level}", seed)
pristine = task.meta["pristine"]
broken = {k: v for k, v in task.files.items() if k in pristine}
assert broken != pristine
assert task.grade(pristine) == 1.0, (level, seed, task.meta["library"])
assert task.grade(broken) == 0.0, (level, seed)
if level in (2, 3): # fixing one of the two bugs gives half
first = task.meta["mutants"][0]
half = dict(broken)
half[first["path"]] = pristine[first["path"]] if sum(m["path"] == first["path"] for m in task.meta["mutants"]) == 1 else half[first["path"]]
if half != broken:
assert task.grade(half) == 0.5, (seed, task.meta["library"])
def test_write_and_edit_tools():
env = WorkbenchEnvironment()
env.reset(task_id="fin-loan-balance", seed=3)
obs = env.step(WorkbenchAction(tool="write", path="notes/a.txt", content="alpha beta\n"))
assert "Wrote" in obs.output
obs = env.step(WorkbenchAction(tool="edit", path="notes/a.txt", old="beta", new="gamma"))
assert "Edited" in obs.output
obs = env.step(WorkbenchAction(tool="read", path="notes/a.txt"))
assert "alpha gamma" in obs.output
obs = env.step(WorkbenchAction(tool="edit", path="notes/a.txt", old="zeta", new="x"))
assert obs.output.startswith("Error")
obs = env.step(WorkbenchAction(tool="write", path="../escape.txt", content="x"))
assert obs.output.startswith("Error")
env.close()
def test_task_id_levels():
from workbench.tasks import is_valid
assert is_valid("ind-oee@2") and is_valid("ind-oee-l2") and not is_valid("ind-oee@3")
assert not is_valid("ind-oee@x") and not is_valid("nope@1")
if "swe-repo-repair" in TASK_IDS:
assert is_valid("swe-repo-repair-l4") and not is_valid("swe-repo-repair@5")
@pytest.mark.skipif("swe-repo-repair" not in TASK_IDS, reason="vendored libraries not present")
def test_repo_library_families():
libs = [t for t in TASK_IDS if t.startswith("swe-repo-") and t != "swe-repo-repair"]
assert len(libs) == 8
for fam in libs:
task = build_task(f"{fam}-l3", 0)
assert fam.endswith(task.meta["library"].replace("_", "-"))
assert task.grade(task.meta["pristine"]) == 1.0
@pytest.mark.skipif("swe-repo-repair" not in TASK_IDS, reason="vendored libraries not present")
def test_repo_repair_grader_resists_tampering():
task = build_task("swe-repo-toolz-l4", 1)
pristine = task.meta["pristine"]
broken = {k: v for k, v in task.files.items() if k in pristine}
first = task.meta["mutants"][0]["path"]
# a deleted or unparseable library file scores 0, even with everything else fixed
assert task.grade({k: v for k, v in pristine.items() if k != first}) == 0.0
assert task.grade(dict(pristine, **{first: pristine[first] + "\n)(\n"})) == 0.0
# comments and blank lines do not matter
assert task.grade({k: "# reviewed\n" + v + "\n\n" for k, v in pristine.items()}) == 1.0
# any change outside the broken functions scores 0, e.g. patching pytest from __init__
init = next(k for k in pristine if k.endswith("__init__.py"))
assert task.grade(dict(pristine, **{init: pristine[init] + "\nimport _pytest\n"})) == 0.0
# the broken state scores 0; one of three fixed scores one third
assert task.grade(broken) == 0.0
|