File size: 5,132 Bytes
53c10a4 | 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 | # Copyright 2026 The OraRL Authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import numpy as np
import pytest
import torch
from orarl.algorithm import (
CorrectionConfig,
PostSelectionReference,
apply_post_selection_correction,
capture_pre_selection_references,
correct_post_selection_group,
)
class FakeDataProto:
def __init__(self, batch, non_tensor_batch, meta_info=None):
self.batch = batch
self.non_tensor_batch = non_tensor_batch
self.meta_info = {} if meta_info is None else meta_info
def __getitem__(self, rows):
return FakeDataProto(
{key: value[rows] for key, value in self.batch.items()},
{key: np.asarray(value)[rows] for key, value in self.non_tensor_batch.items()},
self.meta_info,
)
def test_oracle_projection_preserves_zero_mean_and_nonnegative_anchor():
active = torch.tensor([0.8, 0.2, -0.1, 0.1])
oracle = torch.tensor([False, False, False, True])
corrected, metrics = correct_post_selection_group(
active,
oracle,
reference=PostSelectionReference(
policy_rms=0.5,
sigma_policy=0.2,
policy_rows=3,
),
config=CorrectionConfig(rms_match=False),
)
assert torch.allclose(
corrected,
torch.tensor([0.5, -0.1, -0.4, 0.0]),
atol=1e-6,
)
assert abs(float(corrected.sum())) < 1e-6
assert corrected[oracle].item() >= 0.0
assert metrics["oracle_sign_projection"] == 1.0
def test_rms_matching_only_downscales():
active = torch.tensor([0.5, -0.3, -0.2, 0.4])
oracle = torch.tensor([False, False, False, True])
corrected, metrics = correct_post_selection_group(
active,
oracle,
reference=PostSelectionReference(
policy_rms=0.2,
sigma_policy=0.1,
policy_rows=3,
),
)
assert abs(float(corrected.mean())) < 1e-6
assert math.isclose(
float(torch.sqrt(torch.mean(corrected.square()))),
0.2,
abs_tol=2e-6,
)
assert 0.25 <= metrics["rms_scale"] <= 1.0
small, small_metrics = correct_post_selection_group(
torch.tensor([-0.1, 0.1, 0.05]),
torch.tensor([False, False, True]),
reference=PostSelectionReference(
policy_rms=10.0,
sigma_policy=1.0,
policy_rows=2,
),
)
assert small_metrics["rms_scale"] == 1.0
assert torch.sqrt(torch.mean(small.square())) <= torch.tensor(0.1)
def test_small_reward_spread_skips_rms_scaling():
active = torch.tensor([0.0, 0.0, 0.0, 0.4])
oracle = torch.tensor([False, False, False, True])
corrected, metrics = correct_post_selection_group(
active,
oracle,
reference=PostSelectionReference(
policy_rms=0.0,
sigma_policy=0.0,
policy_rows=3,
),
)
assert torch.allclose(corrected, torch.tensor([-0.1, -0.1, -0.1, 0.3]))
assert metrics["small_sigma_fallback"] == 1.0
assert metrics["rms_scale"] == 1.0
def test_reference_capture_and_batch_correction_use_preselection_policy_rows():
advantages = torch.tensor(
[
[-1.0, -1.0],
[0.1, 0.1],
[0.2, 0.2],
[0.3, 0.3],
[0.2, 0.2],
]
)
data = FakeDataProto(
batch={
"advantages": advantages,
"response_mask": torch.ones_like(advantages),
"token_level_scores": torch.tensor(
[
[0.0, 0.1],
[0.0, 0.2],
[0.0, 0.3],
[0.0, 0.4],
[0.0, 1.0],
]
),
},
non_tensor_batch={
"uid": np.asarray(["group"] * 5, dtype=object),
"is_oracle_row": np.asarray(
[False, False, False, False, True],
dtype=bool,
),
"row_id": np.arange(5),
},
)
references = capture_pre_selection_references(data)
expected_rms = math.sqrt((1.0 + 0.01 + 0.04 + 0.09) / 4.0)
assert references["group"].policy_rows == 4
assert references["group"].policy_rms == pytest.approx(expected_rms)
selected = data[[0, 3, 4]]
metrics = apply_post_selection_correction(selected, references)
active = selected.batch["advantages"][:, 0]
assert abs(float(active.mean())) < 1e-6
assert active[-1].item() >= 0.0
assert metrics["orarl/post_selection_groups"] == 1.0
assert metrics["orarl/post_selection_rms_scale"] <= 1.0
|