Files
OBLITERATUS/tests/test_property_contracts.py
T

227 lines
7.5 KiB
Python

"""Deterministic property contracts for high-consequence pure behavior."""
from __future__ import annotations
import math
import pytest
import torch
from hypothesis import given, seed, settings, strategies as st
from obliteratus.analysis.whitened_svd import WhitenedSVDExtractor
from obliteratus.evaluation.advanced_metrics import (
_is_refusal,
linear_cka,
token_kl_divergence,
)
from obliteratus.evaluation.metrics import accuracy, f1_score_metric, perplexity
PROPERTY_SETTINGS = settings(max_examples=60, deadline=None, database=None)
@seed(7001)
@PROPERTY_SETTINGS
@given(st.lists(st.integers(-5, 5), max_size=50))
def test_accuracy_is_invariant_to_joint_reversal_and_duplication(values):
references = [value % 3 for value in values]
predictions = [value if index % 4 else value + 1 for index, value in enumerate(references)]
expected = accuracy(predictions, references)
assert accuracy(list(reversed(predictions)), list(reversed(references))) == expected
if values:
assert accuracy(predictions * 2, references * 2) == expected
@seed(7002)
@PROPERTY_SETTINGS
@given(
st.lists(st.integers(0, 4), min_size=1, max_size=50),
st.lists(st.booleans(), min_size=1, max_size=50),
)
def test_f1_is_invariant_to_bijective_label_renaming(references, flips):
predictions = [
value if flips[index % len(flips)] else (value + 1) % 5
for index, value in enumerate(references)
]
expected = f1_score_metric(predictions, references)
assert f1_score_metric(
[value + 10 for value in predictions],
[value + 10 for value in references],
) == pytest.approx(expected)
@seed(7003)
@PROPERTY_SETTINGS
@given(
vocab_size=st.integers(2, 40),
batch_size=st.integers(1, 4),
sequence_length=st.integers(2, 12),
)
def test_uniform_causal_lm_perplexity_equals_vocabulary_size(
vocab_size, batch_size, sequence_length,
):
logits = torch.zeros(batch_size, sequence_length, vocab_size)
labels = torch.arange(batch_size * sequence_length).reshape(batch_size, sequence_length)
labels %= vocab_size
assert perplexity(logits, labels) == pytest.approx(float(vocab_size), rel=1e-6)
@seed(7004)
@PROPERTY_SETTINGS
@given(
original=st.lists(
st.floats(-20, 20, allow_nan=False, allow_infinity=False),
min_size=24,
max_size=24,
),
modified=st.lists(
st.floats(-20, 20, allow_nan=False, allow_infinity=False),
min_size=24,
max_size=24,
),
offset=st.floats(-100, 100, allow_nan=False, allow_infinity=False),
)
def test_token_kl_is_nonnegative_and_common_logit_shift_invariant(
original, modified, offset,
):
original_logits = torch.tensor(original).reshape(2, 3, 4)
modified_logits = torch.tensor(modified).reshape(2, 3, 4)
expected = token_kl_divergence(original_logits, modified_logits)
shifted = token_kl_divergence(original_logits + offset, modified_logits - offset)
assert expected >= -1e-5
assert shifted == pytest.approx(expected, abs=2e-5)
@seed(7005)
@PROPERTY_SETTINGS
@given(
scale=st.floats(
min_value=0.1, max_value=20, allow_nan=False, allow_infinity=False,
),
offset=st.floats(-20, 20, allow_nan=False, allow_infinity=False),
)
def test_linear_cka_is_invariant_to_isotropic_scale_and_translation(scale, offset):
generator = torch.Generator().manual_seed(7005)
representation = torch.randn(12, 6, generator=generator)
assert linear_cka(representation, representation * scale + offset) == pytest.approx(
1.0, abs=1e-5,
)
@seed(7006)
@PROPERTY_SETTINGS
@given(
refusal=st.sampled_from([
"I'm sorry, I cannot help with that.",
"I must decline this request.",
"Je ne peux pas vous aider.",
"我无法帮助这个请求。",
]),
leading_space=st.integers(0, 8),
upper=st.booleans(),
)
def test_refusal_classification_is_invariant_to_case_and_leading_space(
refusal, leading_space, upper,
):
transformed = " " * leading_space + (refusal.upper() if upper else refusal.lower())
assert _is_refusal(transformed, mode="combined")
@seed(7007)
@PROPERTY_SETTINGS
@given(
n_samples=st.integers(3, 12),
hidden_dim=st.integers(3, 16),
offset=st.floats(-50, 50, allow_nan=False, allow_infinity=False),
)
def test_whitened_direction_is_invariant_to_common_activation_translation(
n_samples, hidden_dim, offset,
):
generator = torch.Generator().manual_seed(n_samples * 100 + hidden_dim)
harmless = torch.randn(n_samples, hidden_dim, generator=generator)
signal = torch.linspace(-1.0, 1.0, hidden_dim)
harmful = harmless + signal
extractor = WhitenedSVDExtractor()
original = extractor.extract(list(harmful), list(harmless), n_directions=1)
translated = extractor.extract(
list(harmful + offset),
list(harmless + offset),
n_directions=1,
)
alignment = torch.dot(original.directions[0], translated.directions[0]).abs()
assert alignment == pytest.approx(1.0, abs=2e-4)
assert translated.variance_explained == pytest.approx(
original.variance_explained,
abs=2e-5,
)
@seed(7008)
@PROPERTY_SETTINGS
@given(permutation=st.permutations(tuple(range(8))))
def test_whitened_direction_is_invariant_to_joint_sample_permutation(permutation):
generator = torch.Generator().manual_seed(7008)
harmless = torch.randn(8, 10, generator=generator)
harmful = harmless + torch.linspace(-2.0, 2.0, 10)
extractor = WhitenedSVDExtractor()
original = extractor.extract(list(harmful), list(harmless), n_directions=1)
permuted = extractor.extract(
[harmful[index] for index in permutation],
[harmless[index] for index in permutation],
n_directions=1,
)
alignment = torch.dot(original.directions[0], permuted.directions[0]).abs()
assert alignment == pytest.approx(1.0, abs=2e-5)
assert permuted.singular_values == pytest.approx(original.singular_values, rel=2e-5)
@seed(7009)
@PROPERTY_SETTINGS
@given(
n_samples=st.integers(2, 12),
hidden_dim=st.integers(2, 16),
requested=st.integers(1, 8),
)
def test_whitened_outputs_obey_normalization_ordering_and_bounds(
n_samples, hidden_dim, requested,
):
generator = torch.Generator().manual_seed(n_samples * 1000 + hidden_dim)
harmless = torch.randn(n_samples, hidden_dim, generator=generator)
harmful = harmless + torch.randn(n_samples, hidden_dim, generator=generator)
result = WhitenedSVDExtractor(min_variance_ratio=0).extract(
list(harmful),
list(harmless),
n_directions=requested,
)
expected_count = min(requested, n_samples, hidden_dim)
assert result.directions.shape == (expected_count, hidden_dim)
assert result.directions.norm(dim=1) == pytest.approx(torch.ones(expected_count))
assert torch.all(result.singular_values >= 0)
assert torch.all(result.singular_values[:-1] >= result.singular_values[1:])
assert 0.0 <= result.variance_explained <= 1.0
assert math.isfinite(result.condition_number)
assert math.isfinite(result.effective_rank)
@pytest.mark.parametrize("dtype", [torch.float16, torch.float32, torch.float64])
def test_whitened_direction_normalizes_supported_cpu_input_dtypes(dtype):
harmless = torch.tensor(
[[0, 0, 0], [1, 0, 0], [0, 1, 0], [0, 0, 1]],
dtype=dtype,
)
harmful = harmless + torch.tensor([0, 0, 2], dtype=dtype)
result = WhitenedSVDExtractor().extract(
list(harmful),
list(harmless),
n_directions=1,
)
assert result.directions.dtype == torch.float32
assert result.directions.norm() == pytest.approx(1.0)