File size: 11,110 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 | """Select only the two preregistered epochs using complete validation evidence."""
from __future__ import annotations
import argparse
import hashlib
import json
import shutil
from pathlib import Path
from typing import Any
from evaluate_clef import (
ENCODING_VERSION,
FINAL_TEST_SEEDS,
checkpoint_hashes,
file_hash,
validate_selection,
validation_ineligibility,
)
from stackcraft.data import audit_dataset, canonical_json
from stackcraft.evaluation import summarize_positions
from stackcraft.players import Decision
RULE = (
"lowest finite mean target NLL over all 215 validation positions; "
"exact tie chooses earlier epoch"
)
PLAYERS = ["base", "base-fp32", "trained", "random", "heuristic"]
TRAINING_CONFIG = {
"mode": "lora",
"epochs": 2,
"seed": 42,
"rank": 4,
"learning_rate": 1e-5,
"batch_size": 1,
"gradient_accumulation": 8,
"max_length": 4096,
"optimizer": "AdamW",
"weight_decay": 0.01,
"clip_gradient_norm": 1.0,
"label_smoothing": 0.05,
"brier_weight": 0.1,
}
def load_validation(dataset: Path) -> tuple[list[dict[str, Any]], str]:
manifest = json.loads((dataset / "manifest.json").read_text())
rows = {
split: [json.loads(line) for line in (dataset / f"{split}.jsonl").read_text().splitlines()]
for split in ("train", "validation")
}
audit_dataset(rows, manifest)
if len(rows["validation"]) != 215:
raise ValueError("selection requires all 215 frozen validation positions")
return rows["validation"], file_hash(dataset / "manifest.json")
def inspect_candidate(
checkpoint: Path,
report_path: Path,
epoch: int,
rows: list[dict[str, Any]],
manifest_sha256: str,
) -> dict[str, Any]:
hashes = checkpoint_hashes(checkpoint)
metadata = json.loads((checkpoint / "training_config.json").read_text())
report = json.loads(report_path.read_text())
extra = metadata.get("extra", {})
if extra.get("epoch") != epoch or extra.get("dataset_manifest_sha256") != manifest_sha256:
raise ValueError("candidate epoch/dataset differs from preregistered study")
if any(extra.get("config", {}).get(key) != value for key, value in TRAINING_CONFIG.items()):
raise ValueError("candidate training configuration differs from preregistration")
if (
extra.get("test_trajectories_used") is not False
or extra.get("validation_used_for_training") is not False
):
raise ValueError("candidate training metadata must exclude test and validation use")
if (
report.get("checkpoint_sha256") != hashes
or report.get("dataset_manifest_sha256") != manifest_sha256
):
raise ValueError("validation report is not bound to this checkpoint and dataset")
if (
report.get("positions") != 215
or report.get("mode") != "positions"
or report.get("split") != "validation"
):
raise ValueError("candidate report must cover the complete validation split")
trained = report.get("players", {}).get("trained")
if not isinstance(trained, dict):
raise ValueError("candidate report lacks trained predictions")
predictions_path = report_path.parent / trained["predictions_file"]
if file_hash(predictions_path) != trained["predictions_sha256"]:
raise ValueError("raw prediction hash differs from validation report")
events = [json.loads(line) for line in predictions_path.read_text().splitlines()]
if [event.get("id") for event in events] != [row["id"] for row in rows]:
raise ValueError("raw predictions omit, duplicate, reorder or add validation positions")
predictions = {}
for row, event in zip(rows, events, strict=True):
if event.get("target_action_id") != row["action_id"]:
raise ValueError("prediction target differs from frozen teacher label")
predictions[row["id"]] = None if "error" in event else Decision(**event["decision"])
recomputed = summarize_positions(rows, predictions)
if trained.get("metrics") != recomputed:
raise ValueError("reported validation metrics differ from raw predictions")
reasons = validation_ineligibility(report, metadata)
return {
"key": f"epoch-{epoch:02d}",
"epoch": epoch,
"checkpoint_sha256": hashes,
"metrics": recomputed,
"ineligibility_reasons": reasons,
"checkpoint": checkpoint,
"report_path": report_path,
"report": report,
"metadata": metadata,
}
def select_checkpoint(
candidates: list[tuple[Path, Path]], dataset: Path, output: Path
) -> dict[str, Any]:
if len(candidates) != 2:
raise ValueError("exactly epoch 01 and epoch 02 candidates are required")
if output.exists():
raise FileExistsError("selection output already exists")
rows, manifest_hash = load_validation(dataset)
inspected = [
inspect_candidate(checkpoint, report, index + 1, rows, manifest_hash)
for index, (checkpoint, report) in enumerate(candidates)
]
first, second = inspected
for key in ("source_hashes", "config", "dataset_split_sha256", "dataset_counts"):
if first["metadata"]["extra"].get(key) != second["metadata"]["extra"].get(key):
raise ValueError(f"candidate training provenance differs: {key}")
for key in ("source_hashes", "installed_versions", "encoding_version", "base_revision"):
if first["report"].get("provenance", {}).get(key) != second["report"].get(
"provenance", {}
).get(key):
raise ValueError(f"candidate evaluation provenance differs: {key}")
runtime = [
candidate["report"]["players"]["trained"]["runtime_config"] for candidate in inspected
]
for key in (
"model_id",
"base_revision",
"encoding_version",
"source_sha256",
"max_length",
"dtype",
"device",
"head_dtype",
"versions",
):
if runtime[0].get(key) != runtime[1].get(key):
raise ValueError(f"candidate evaluation runtime differs: {key}")
output.mkdir(parents=True)
preserved: list[dict[str, Any]] = []
for candidate in inspected:
folder = output / candidate["key"]
folder.mkdir()
shutil.copyfile(candidate["report_path"], folder / "report.json")
shutil.copyfile(
candidate["checkpoint"] / "training_config.json", folder / "training_config.json"
)
for player in candidate["report"]["players"].values():
relative = Path(player["predictions_file"])
if relative.is_absolute() or ".." in relative.parts:
raise ValueError("prediction evidence must use a contained relative path")
source = candidate["report_path"].parent / relative
if file_hash(source) != player["predictions_sha256"]:
raise ValueError("diagnostic prediction file hash mismatch")
target = folder / relative
target.parent.mkdir(parents=True, exist_ok=True)
shutil.copyfile(source, target)
preserved.append(
{
"key": candidate["key"],
"epoch": candidate["epoch"],
"checkpoint_sha256": candidate["checkpoint_sha256"],
"checkpoint_origin": str(candidate["checkpoint"].resolve()),
"checkpoint_metadata": {
"path": f"{candidate['key']}/training_config.json",
"sha256": file_hash(folder / "training_config.json"),
},
"validation_evidence": {
"path": f"{candidate['key']}/report.json",
"sha256": file_hash(folder / "report.json"),
},
"metrics": candidate["metrics"],
"ineligibility_reasons": candidate["ineligibility_reasons"],
}
)
eligible = [candidate for candidate in preserved if not candidate["ineligibility_reasons"]]
audit = {
"decision_rule": RULE,
"dataset_manifest_sha256": manifest_hash,
"candidates": preserved,
"status": "selected" if eligible else "no-eligible-candidate",
}
(output / "selection-audit.json").write_text(
json.dumps(audit, indent=2, allow_nan=False) + "\n"
)
if not eligible:
raise ValueError(
"both epoch candidates are ineligible; evidence preserved, no test selection"
)
selected = min(
eligible, key=lambda candidate: (candidate["metrics"]["mean_nll"], candidate["epoch"])
)
selection = {
"schema_version": 1,
"selection_split": "validation",
"selected_key": selected["key"],
"decision_rule": RULE,
"checkpoint_sha256": selected["checkpoint_sha256"],
"validation_metrics": selected["metrics"],
"validation_evidence": selected["validation_evidence"],
"candidates": preserved,
"dataset_manifest_sha256": manifest_hash,
"test_seeds": list(FINAL_TEST_SEEDS),
"test_seeds_sha256": hashlib.sha256(
canonical_json(list(FINAL_TEST_SEEDS)).encode()
).hexdigest(),
"max_pieces": 200,
"encoding_version": ENCODING_VERSION,
"max_length": 4096,
"players": PLAYERS,
"head_dtypes": {"base": "bfloat16", "base-fp32": "float32", "trained": "float32"},
"bootstrap_samples": 10000,
"bootstrap_seed": 2026,
"max_error_rate": 0.0,
}
selection_path = output / "selection.json"
selection_path.write_text(json.dumps(selection, indent=2, allow_nan=False) + "\n")
args = argparse.Namespace(
final_test=True,
seeds=FINAL_TEST_SEEDS,
max_pieces=200,
players=PLAYERS,
max_length=4096,
selection_file=selection_path,
checkpoint=Path(selected["checkpoint_origin"]),
)
validate_selection(args, selected["checkpoint_sha256"])
return selection
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
"--candidate01", nargs=2, type=Path, metavar=("CHECKPOINT", "REPORT"), required=True
)
parser.add_argument(
"--candidate02", nargs=2, type=Path, metavar=("CHECKPOINT", "REPORT"), required=True
)
parser.add_argument("--dataset", type=Path, default=Path("data/study-v1"))
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args(argv)
try:
selection = select_checkpoint(
[tuple(args.candidate01), tuple(args.candidate02)], args.dataset, args.output
)
except (ValueError, OSError, KeyError) as error:
parser.error(str(error))
print(
json.dumps(
{
"selected_key": selection["selected_key"],
"mean_nll": selection["validation_metrics"]["mean_nll"],
"output": str(args.output),
}
)
)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|