File size: 7,827 Bytes
087e4a0 | 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 | from __future__ import annotations
import json
import tempfile
import unittest
from pathlib import Path
from hcl.optimizer import Optimizer
ROOT = Path(__file__).resolve().parents[1]
class PaperAlignmentTest(unittest.TestCase):
def test_main_stream_order_and_sample_limits(self) -> None:
text_stream = json.loads(
(ROOT / "configs/taskstream_textual_main_250_50_500.json").read_text()
)
multimodal_stream = json.loads(
(ROOT / "configs/taskstream_multimodal_main_250_50_500.json").read_text()
)
self.assertEqual(
text_stream["paper_order"],
["musique", "proofwriter", "gsm8k", "hotpotqa"],
)
self.assertEqual(
multimodal_stream["paper_order"],
["coco_detection", "coco_caption", "refcoco_grounding", "vqav2"],
)
self.assertTrue(all("anchor" in task["splits"] for task in text_stream["tasks"]))
self.assertTrue(
all("anchor" in task["splits"] for task in multimodal_stream["tasks"])
)
for filename in (
"deepseek_flash_reasoning_hcl_stability_250_50_500.json",
"deepseek_flash_reasoning_hcl_plasticity_250_50_500.json",
"qwen36_27b_coco_gpu1_hcl_stability_250_50_500.json",
"qwen36_27b_coco_gpu1_hcl_plasticity_250_50_500.json",
):
config = json.loads((ROOT / "configs" / filename).read_text())
train = config["task_flow"][0]
self.assertEqual(train["train_limit_per_task"], 250)
self.assertEqual(train["validation_limit_per_task"], 50)
self.assertEqual(train["test_limit_per_task"], 500)
self.assertEqual(config["optimizer"]["min_format_compliance_rate"], 1.0)
self.assertGreater(config["optimizer"]["min_primary_score_delta"], 0.0)
def test_main_profiles_differ_only_in_historical_loss_budget(self) -> None:
pairs = (
(
"deepseek_flash_reasoning_hcl_stability_250_50_500.json",
"deepseek_flash_reasoning_hcl_plasticity_250_50_500.json",
),
(
"qwen36_27b_coco_gpu1_hcl_stability_250_50_500.json",
"qwen36_27b_coco_gpu1_hcl_plasticity_250_50_500.json",
),
)
for stability_name, plasticity_name in pairs:
stability = json.loads((ROOT / "configs" / stability_name).read_text())[
"optimizer"
]
plasticity = json.loads((ROOT / "configs" / plasticity_name).read_text())[
"optimizer"
]
self.assertEqual(stability.pop("historical_loss_budget"), 0)
self.assertIsNone(plasticity.pop("historical_loss_budget"))
self.assertEqual(stability, plasticity)
def test_explicit_historical_budget_gates_anchor_loss(self) -> None:
candidates = [{"candidate_id": "candidate", "validation_errors": []}]
current = {
"current_task_metrics": {
"primary_score": 0.5,
"correct": 5,
"format_compliance_rate": 1.0,
}
}
results = {
"candidate": {
"metrics": {},
"current_task_metrics": {
"primary_score": 0.6,
"correct": 6,
"format_compliance_rate": 1.0,
},
"historical_anchor_metrics": {
"forget_count": 1,
"recovered_count": 0,
"total": 10,
"forget_rate": 0.1,
},
}
}
with tempfile.TemporaryDirectory() as directory:
stability = Optimizer(
record_dir=Path(directory) / "stability",
selection_objective="plasticity",
historical_loss_budget=0,
min_primary_score_delta=1e-9,
min_correct_gain=0,
min_format_compliance_rate=1.0,
)
plasticity = Optimizer(
record_dir=Path(directory) / "plasticity",
selection_objective="plasticity",
historical_loss_budget=None,
min_primary_score_delta=1e-9,
min_correct_gain=0,
min_format_compliance_rate=1.0,
)
self.assertIsNone(stability.select_best_candidate(candidates, results, current))
self.assertIsNotNone(
plasticity.select_best_candidate(candidates, results, current)
)
def test_budget_sweep_matches_paper_protocol(self) -> None:
expected_budgets = {"b0": 0, "b1": 1, "b3": 3, "binf": None}
common_optimizer = None
for label, budget in expected_budgets.items():
path = ROOT / "configs" / (
f"deepseek_v4_flash_textual_budget_{label}_300_80_80_600.json"
)
config = json.loads(path.read_text())
self.assertEqual(config["model"]["paper_model_name"], "DeepSeek-V4-Flash")
self.assertEqual(config["model"]["temperature"], 0.0)
self.assertEqual(config["model"]["thinking"], {"type": "disabled"})
optimizer = dict(config["optimizer"])
self.assertEqual(optimizer.pop("historical_loss_budget"), budget)
if common_optimizer is None:
common_optimizer = optimizer
else:
self.assertEqual(optimizer, common_optimizer)
self.assertEqual(config["memory"]["anchor_capacity_per_task"], 80)
train = config["task_flow"][0]
self.assertEqual(300 // train["batchsize"], 10)
self.assertEqual(train["validation_limit_per_task"], 80)
self.assertEqual(train["test_limit_per_task"], 600)
stream = json.loads(
(ROOT / "configs/taskstream_textual_budget_300_80_80_600.json").read_text()
)
self.assertEqual(
stream["paper_order"],
["musique", "proofwriter", "gsm8k", "hotpotqa"],
)
self.assertTrue(all("anchor" in task["splits"] for task in stream["tasks"]))
def test_current_score_tie_prefers_lower_historical_loss(self) -> None:
candidates = [
{"candidate_id": "higher_loss", "validation_errors": []},
{"candidate_id": "lower_loss", "validation_errors": []},
]
current = {
"current_task_metrics": {
"primary_score": 0.5,
"correct": 5,
"format_compliance_rate": 1.0,
}
}
results = {}
for candidate_id, loss in (("higher_loss", 2), ("lower_loss", 1)):
results[candidate_id] = {
"metrics": {},
"current_task_metrics": {
"primary_score": 0.6,
"correct": 6,
"format_compliance_rate": 1.0,
},
"historical_anchor_metrics": {
"forget_count": loss,
"recovered_count": 0,
"total": 80,
"forget_rate": loss / 80,
},
}
with tempfile.TemporaryDirectory() as directory:
optimizer = Optimizer(
record_dir=directory,
selection_objective="plasticity",
historical_loss_budget=3,
min_primary_score_delta=1e-9,
min_correct_gain=0,
min_format_compliance_rate=1.0,
)
selected = optimizer.select_best_candidate(candidates, results, current)
self.assertEqual(selected["candidate_id"], "lower_loss")
if __name__ == "__main__":
unittest.main()
|