File size: 6,334 Bytes
4be6a52 | 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 | """Selection uses small fake checkpoints and synthetic reports, never model loads."""
import importlib.util
import json
import math
import sys
from dataclasses import asdict
from pathlib import Path
import pytest
from stackcraft.engine import new_game
from stackcraft.evaluation import summarize_positions
from stackcraft.players import Decision, observe
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "scripts"))
try:
SPEC = importlib.util.spec_from_file_location(
"stackcraft_select_cli", ROOT / "scripts/select_checkpoint.py"
)
assert SPEC is not None and SPEC.loader is not None
CLI = importlib.util.module_from_spec(SPEC)
SPEC.loader.exec_module(CLI)
finally:
sys.path.pop(0)
@pytest.fixture
def rows(monkeypatch):
observation = observe(new_game(70))
records = [
{
"id": f"validation-{index}",
"action_id": observation.legal_actions[0].id,
"observation": asdict(observation),
}
for index in range(215)
]
monkeypatch.setattr(CLI, "load_validation", lambda path: (records, "manifest"))
return records
def candidate(root, rows, epoch, target_probability, *, failure=False):
checkpoint = root / f"epoch-{epoch:02d}"
checkpoint.mkdir()
metadata = {
"extra": {
"epoch": epoch,
"config": CLI.TRAINING_CONFIG,
"dataset_manifest_sha256": "manifest",
"test_trajectories_used": False,
"validation_used_for_training": False,
"source_hashes": {"training": "same-source"},
"dataset_counts": {"validation": 215},
"dataset_split_sha256": {"validation": "same-validation"},
}
}
(checkpoint / "training_config.json").write_text(json.dumps(metadata))
(checkpoint / "joint_head.safetensors").write_bytes(b"FAKE-NOT-A-MODEL")
hashes = CLI.checkpoint_hashes(checkpoint)
validation = root / f"validation-{epoch}"
validation.mkdir()
predictions = {}
events = []
for index, row in enumerate(rows):
target = row["action_id"]
options = [action["id"] for action in row["observation"]["legal_actions"]]
probs = {
option: target_probability
if option == target
else (1 - target_probability) / (len(options) - 1)
for option in options
}
decision = Decision(target, probs)
event = {"id": row["id"], "target_action_id": target}
if failure and index == 0:
event["error"] = {"type": "RuntimeError", "message": "test"}
predictions[row["id"]] = None
else:
event["decision"] = asdict(decision)
predictions[row["id"]] = decision
events.append(event)
predictions_path = validation / "trained-positions.jsonl"
predictions_path.write_text("".join(json.dumps(event) + "\n" for event in events))
metrics = summarize_positions(rows, predictions)
report = {
"mode": "positions",
"split": "validation",
"positions": 215,
"dataset_manifest_sha256": "manifest",
"checkpoint_sha256": hashes,
"provenance": {
"source_hashes": {"evaluation": "same"},
"encoding_version": CLI.ENCODING_VERSION,
},
"players": {
"trained": {
"metrics": metrics,
"runtime_config": {"head_dtype": "float32", "max_length": 4096},
"predictions_file": predictions_path.name,
"predictions_sha256": CLI.file_hash(predictions_path),
}
},
}
report_path = validation / "report.json"
report_path.write_text(json.dumps(report))
return checkpoint, report_path
@pytest.mark.parametrize(
"probabilities,selected", [((0.2, 0.4), "epoch-02"), ((0.4, 0.4), "epoch-01")]
)
def test_lowest_nll_and_earlier_exact_tie_preserve_both_reports(
tmp_path, rows, probabilities, selected
):
pairs = [
candidate(tmp_path, rows, epoch, probability)
for epoch, probability in enumerate(probabilities, 1)
]
output = tmp_path / "selection"
result = CLI.select_checkpoint(pairs, tmp_path / "data", output)
assert result["selected_key"] == selected
assert result["validation_metrics"]["mean_nll"] == pytest.approx(-math.log(max(probabilities)))
assert len(result["candidates"]) == 2
for epoch in (1, 2):
assert (output / f"epoch-{epoch:02d}/trained-positions.jsonl").is_file()
assert (output / "selection-audit.json").is_file()
assert result["max_pieces"] == 200
assert len(result["test_seeds"]) == 200
def test_inference_error_makes_candidate_ineligible_even_with_better_remaining_nll(tmp_path, rows):
pairs = [candidate(tmp_path, rows, 1, 0.9, failure=True), candidate(tmp_path, rows, 2, 0.2)]
result = CLI.select_checkpoint(pairs, tmp_path / "data", tmp_path / "selection")
assert result["selected_key"] == "epoch-02"
assert "validation inference errors" in result["candidates"][0]["ineligibility_reasons"]
def test_both_nonfinite_candidates_leave_evidence_without_test_selection(tmp_path, rows):
pairs = [candidate(tmp_path, rows, epoch, 0.0) for epoch in (1, 2)]
output = tmp_path / "selection"
with pytest.raises(ValueError, match="both epoch candidates"):
CLI.select_checkpoint(pairs, tmp_path / "data", output)
assert (output / "selection-audit.json").is_file()
assert not (output / "selection.json").exists()
@pytest.mark.parametrize("corruption", ["metrics", "dataset", "extra_candidate", "provenance"])
def test_corrupted_or_unregistered_evidence_is_rejected(tmp_path, rows, corruption):
pairs = [candidate(tmp_path, rows, epoch, 0.4) for epoch in (1, 2)]
report = json.loads(pairs[1][1].read_text())
if corruption == "metrics":
report["players"]["trained"]["metrics"]["mean_nll"] = 0.0
elif corruption == "dataset":
report["dataset_manifest_sha256"] = "other"
elif corruption == "provenance":
report["provenance"]["source_hashes"] = {"evaluation": "changed"}
else:
pairs.append(pairs[0])
pairs[1][1].write_text(json.dumps(report))
with pytest.raises(ValueError):
CLI.select_checkpoint(pairs, tmp_path / "data", tmp_path / "selection")
|