Spaces:
Sleeping
Sleeping
File size: 9,866 Bytes
cdf485e f3f7bc4 cdf485e f3f7bc4 cdf485e f3f7bc4 cdf485e f3f7bc4 cdf485e f3f7bc4 cdf485e f3f7bc4 cdf485e f3f7bc4 cdf485e f3f7bc4 cdf485e f3f7bc4 cdf485e 7574c9a cdf485e f3f7bc4 7574c9a cdf485e 7574c9a cdf485e f3f7bc4 cdf485e 7574c9a cdf485e f3f7bc4 cdf485e 7574c9a cdf485e 7574c9a cdf485e 7574c9a cdf485e 7574c9a cdf485e b9664a2 | 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 | """
Tests for PreferenceLab environment.
Run: pytest tests/ -v
"""
import sys
import os
sys.path.insert(0, os.path.dirname(os.path.dirname(__file__)))
import pytest
from models import (
PairwiseAction, LikertAction, ConsistencyAction,
PairwiseObservation, LikertObservation, ConsistencyObservation,
)
from server.environment import (
PreferenceLabEnvironment,
grade_pairwise, grade_likert, grade_consistency,
)
# ── Grader unit tests ─────────────────────────────────────────
class TestPairwiseGrader:
def test_correct_choice_scores_1(self):
action = PairwiseAction(choice="A")
example = {"gold_label": "A", "source": "test"}
reward, info = grade_pairwise(action, example)
assert reward == 0.99
assert info["verdict"] == "correct"
def test_wrong_choice_scores_0(self):
action = PairwiseAction(choice="B")
example = {"gold_label": "A", "source": "test"}
reward, info = grade_pairwise(action, example)
assert reward == 0.01
assert info["verdict"] == "incorrect"
def test_skip_scores_partial(self):
action = PairwiseAction(choice="skip")
example = {"gold_label": "A", "source": "test"}
reward, info = grade_pairwise(action, example)
assert reward == 0.3
def test_tie_scores_low(self):
action = PairwiseAction(choice="tie")
example = {"gold_label": "A", "source": "test"}
reward, info = grade_pairwise(action, example)
assert reward == 0.1
def test_reward_in_range(self):
for choice in ["A", "B", "tie", "skip"]:
action = PairwiseAction(choice=choice)
reward, _ = grade_pairwise(action, {"gold_label": "A", "source": "test"})
assert 0.0 < reward < 1.0
class TestLikertGrader:
def test_perfect_scores_reward_1(self):
action = LikertAction(helpfulness=5, honesty=5, harmlessness=5, instruction_following=5)
example = {
"gold_scores": {"helpfulness": 5, "honesty": 5, "harmlessness": 5, "instruction_following": 5},
"source": "test",
}
reward, info = grade_likert(action, example)
assert reward == 0.99
def test_worst_scores_reward_0(self):
action = LikertAction(helpfulness=1, honesty=1, harmlessness=1, instruction_following=1)
example = {
"gold_scores": {"helpfulness": 5, "honesty": 5, "harmlessness": 5, "instruction_following": 5},
"source": "test",
}
reward, info = grade_likert(action, example)
assert reward == 0.01
def test_partial_error_gives_partial_reward(self):
action = LikertAction(helpfulness=4, honesty=4, harmlessness=4, instruction_following=4)
example = {
"gold_scores": {"helpfulness": 5, "honesty": 5, "harmlessness": 5, "instruction_following": 5},
"source": "test",
}
reward, info = grade_likert(action, example)
assert 0.0 < reward < 1.0
def test_reward_always_in_range(self):
import random
for _ in range(20):
action = LikertAction(
helpfulness=random.randint(1, 5),
honesty=random.randint(1, 5),
harmlessness=random.randint(1, 5),
instruction_following=random.randint(1, 5),
)
example = {
"gold_scores": {
"helpfulness": random.randint(1, 5),
"honesty": random.randint(1, 5),
"harmlessness": random.randint(1, 5),
"instruction_following": random.randint(1, 5),
}
}
reward, _ = grade_likert(action, example)
assert 0.0 < reward < 1.0, f"Reward out of range: {reward}"
class TestConsistencyGrader:
def test_perfect_ranking_scores_1(self):
action = ConsistencyAction(ranking=["A", "B", "C", "D"])
example = {"gold_ranking": ["A", "B", "C", "D"], "source": "test"}
reward, info = grade_consistency(action, example)
assert reward == 0.99
def test_reversed_ranking_scores_low(self):
action = ConsistencyAction(ranking=["D", "C", "B", "A"])
example = {"gold_ranking": ["A", "B", "C", "D"], "source": "test"}
reward, info = grade_consistency(action, example)
# Transitivity score = 0.5 (ranking is still a valid total order)
# Quality score = 0.0 (worst possible Kendall tau = -1 → normalized to 0)
# Total = 0.5 — strictly less than perfect score of 1.0
assert reward < 1.0
assert info["quality_score"] == 0.0
def test_invalid_ids_scores_low(self):
action = ConsistencyAction(ranking=["A", "B", "C", "X"])
example = {"gold_ranking": ["A", "B", "C", "D"], "source": "test"}
reward, info = grade_consistency(action, example)
assert reward == 0.01
assert info["has_invalid_ids"] is True
def test_reward_always_in_range(self):
import itertools
import random
ids = ["A", "B", "C", "D"]
gold = ["A", "B", "C", "D"]
for perm in itertools.permutations(ids):
action = ConsistencyAction(ranking=list(perm))
example = {"gold_ranking": gold, "source": "test"}
reward, _ = grade_consistency(action, example)
assert 0.0 < reward < 1.0, f"Reward out of range: {reward} for {perm}"
def test_graders_not_always_same_score(self):
"""Critical: graders must NOT always return the same score."""
action_correct = ConsistencyAction(ranking=["A", "B", "C", "D"])
action_wrong = ConsistencyAction(ranking=["D", "C", "B", "A"])
example = {"gold_ranking": ["A", "B", "C", "D"], "source": "test"}
r1, _ = grade_consistency(action_correct, example)
r2, _ = grade_consistency(action_wrong, example)
assert r1 != r2, "Grader must return different scores for different inputs!"
# ── Environment integration tests ─────────────────────────────
class TestPreferenceLabEnvironment:
def setup_method(self):
self.env = PreferenceLabEnvironment()
def test_reset_returns_observation(self):
obs = self.env.reset()
assert obs is not None
assert hasattr(obs, "prompt")
assert hasattr(obs, "reward")
assert hasattr(obs, "done")
def test_reset_pairwise_returns_pairwise_obs(self):
obs = self.env.reset(task_type="pairwise")
assert isinstance(obs, PairwiseObservation)
assert obs.response_a != ""
assert obs.response_b != ""
def test_reset_likert_returns_likert_obs(self):
obs = self.env.reset(task_type="likert")
assert isinstance(obs, LikertObservation)
assert obs.response != ""
assert obs.rubric != ""
def test_reset_consistency_returns_consistency_obs(self):
obs = self.env.reset(task_type="consistency")
assert isinstance(obs, ConsistencyObservation)
assert obs.response_a != ""
assert obs.response_d != ""
def test_step_pairwise(self):
self.env.reset(task_type="pairwise")
action = PairwiseAction(choice="A")
obs = self.env.step(action)
assert isinstance(obs, PairwiseObservation)
assert 0.0 < obs.reward < 1.0
assert isinstance(obs.done, bool)
def test_step_likert(self):
self.env.reset(task_type="likert")
action = LikertAction(helpfulness=4, honesty=4, harmlessness=5, instruction_following=4)
obs = self.env.step(action)
assert isinstance(obs, LikertObservation)
assert 0.0 < obs.reward < 1.0
def test_step_consistency(self):
self.env.reset(task_type="consistency")
action = ConsistencyAction(ranking=["A", "B", "C", "D"])
obs = self.env.step(action)
assert isinstance(obs, ConsistencyObservation)
assert 0.0 < obs.reward < 1.0
def test_episode_terminates_after_max_steps(self):
self.env.reset(task_type="pairwise")
done = False
steps = 0
while not done:
obs = self.env.step(PairwiseAction(choice="A"))
done = obs.done
steps += 1
assert steps <= 10, "Episode ran too long!"
assert done is True
def test_state_returns_metadata(self):
self.env.reset(seed=42, task_type="pairwise")
state = self.env.state
assert "episode_id" in state.model_dump()
assert "step_count" in state.model_dump()
assert "task_type" in state.model_dump()
assert state.seed == 42
def test_reproducible_with_seed(self):
obs1 = self.env.reset(seed=123, task_type="pairwise")
obs2 = self.env.reset(seed=123, task_type="pairwise")
assert obs1.prompt == obs2.prompt
assert obs1.response_a == obs2.response_a
def test_rewards_vary_across_actions(self):
"""Ensure graders do NOT always return the same reward (disqualification check)."""
rewards = set()
for _ in range(5):
self.env.reset(task_type="pairwise")
obs_a = self.env.step(PairwiseAction(choice="A"))
self.env.reset(task_type="pairwise")
obs_b = self.env.step(PairwiseAction(choice="B"))
rewards.add(obs_a.reward)
rewards.add(obs_b.reward)
assert len(rewards) > 1, "Grader always returns the same score — DISQUALIFICATION!"
def test_all_three_tasks_run(self):
for task in ["pairwise", "likert", "consistency"]:
obs = self.env.reset(task_type=task)
assert obs is not None
state = self.env.state
assert state.task_type == task
|