File size: 12,385 Bytes
784f572 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 | """Audit C64 conditional generation, evaluation, and timing artifacts."""
from __future__ import annotations
from dataclasses import dataclass
import json
import math
from pathlib import Path
from typing import Any, Iterable, Mapping
from dlb.aggregate import (
IncompleteMatrixError,
_finite,
_json,
_provenance,
_write_csv,
)
from dlb.conditional_matrix import (
ConditionalMatrixTask,
build_conditional_matrix,
conditional_unsupported_inventory,
)
from dlb.io import expected_conditional_schedule, validate_conditional_samples
from dlb.registry import ExperimentRegistry, load_registry
QUALITY_METRICS = (
"conditional_generative_perplexity",
"mauve_suffix",
"sample_entropy",
"sample_entropy_delta",
"self_bleu",
"prefix_exact_match",
)
ALL_METRICS = QUALITY_METRICS + ("generation_seconds_per_sample",)
CONDITIONAL_AGGREGATE_SCHEMA = "dlb-conditional-results-summary-v1"
DATASET_CONTRACTS = {
"lm1b": {"sequence_length": 128, "vocab_size": 30_522},
"owt": {"sequence_length": 1024, "vocab_size": 50_257},
}
@dataclass(frozen=True)
class ConditionalAggregateReport:
rows: tuple[dict[str, Any], ...]
failures: tuple[dict[str, Any], ...]
unsupported: tuple[dict[str, Any], ...]
expected_tasks: int
complete: bool
output_dir: Path | None = None
def _finite_signed(value: object) -> float:
if type(value) not in {int, float} or isinstance(value, bool):
raise ValueError("metric value must be a finite number")
result = float(value)
if not math.isfinite(result):
raise ValueError("metric value is not finite")
return result
def _lookup_metric(metrics: Mapping[str, Any], name: str) -> float:
if name == "conditional_generative_perplexity":
value = metrics.get("conditional_generative_perplexity")
if not isinstance(value, Mapping) or "perplexity" not in value:
raise ValueError("missing conditional generative perplexity")
return _finite(value["perplexity"], positive=True)
if name == "mauve_suffix":
value = metrics.get("mauve_suffix")
if not isinstance(value, Mapping) or "score" not in value:
raise ValueError("missing MAUVE suffix score")
score = _finite(value["score"])
if score > 1:
raise ValueError("MAUVE suffix score must be in [0, 1]")
return score
if name == "sample_entropy":
value = metrics.get("sample_entropy")
generated = value.get("generated") if isinstance(value, Mapping) else None
if not isinstance(generated, Mapping) or "mean_entropy" not in generated:
raise ValueError("missing generated sample entropy")
return _finite(generated["mean_entropy"])
if name == "sample_entropy_delta":
value = metrics.get("sample_entropy")
if not isinstance(value, Mapping) or "generated_minus_reference" not in value:
raise ValueError("missing sample entropy delta")
return _finite_signed(value["generated_minus_reference"])
if name == "self_bleu":
value = metrics.get("self_bleu")
if not isinstance(value, Mapping) or "score" not in value:
raise ValueError("missing Self-BLEU score")
score = _finite(value["score"])
if score > 1:
raise ValueError("Self-BLEU must be in [0, 1]")
return score
if name == "prefix_exact_match":
value = metrics.get("prefix_exact_match")
if not isinstance(value, Mapping) or "rate" not in value:
raise ValueError("missing prefix exact-match rate")
rate = _finite(value["rate"])
if rate != 1.0:
raise ValueError("prefix exact-match rate must be 100%")
return rate
raise ValueError(f"unknown metric {name}")
def _validate_task(task: ConditionalMatrixTask) -> dict[str, Any]:
contract = DATASET_CONTRACTS[task.dataset]
sample_dir = Path(task.sample_dir)
samples = sample_dir / "samples.jsonl"
metadata_path = sample_dir / "run_metadata.json"
metrics_path = Path(task.metrics_path)
timing_path = Path(task.timing_path)
validate_conditional_samples(
samples,
expected=task.sample_count,
schedule=expected_conditional_schedule(),
sequence_length=int(contract["sequence_length"]),
vocab_size=int(contract["vocab_size"]),
)
metadata = _json(metadata_path)
if metadata.get("status") != "succeeded":
raise ValueError("run metadata is not marked succeeded")
identity = metadata.get("identity")
if not isinstance(identity, Mapping) or identity.get("generation_mode") != "conditional_prefix":
raise ValueError("run metadata is not a conditional-prefix run")
provenance = _provenance(metadata, task)
metrics_document = _json(metrics_path)
if metrics_document.get("sample_count") != task.sample_count:
raise ValueError("metrics sample_count does not match matrix task")
if metrics_document.get("protocol") != "c64_zs_v1":
raise ValueError("metrics document has the wrong conditional protocol")
metric_document = metrics_document.get("metrics")
if not isinstance(metric_document, Mapping):
raise ValueError("metrics document has no metrics object")
quality = {name: _lookup_metric(metric_document, name) for name in QUALITY_METRICS}
timing_document = _json(timing_path)
if timing_document.get("schema") != "dlb-generation-timing-v1":
raise ValueError("timing document has an unsupported schema")
timing = timing_document.get("timing")
if not isinstance(timing, Mapping):
raise ValueError("timing document has no timing object")
protocol = {
"mode": "primary_latency",
"warmups": 5,
"repeats": 32,
"batch_size": 1,
"num_timed_samples": 32,
}
if any(timing.get(key) != value for key, value in protocol.items()):
raise ValueError("timing document does not use the pinned latency protocol")
seconds_value = _finite(timing.get("seconds_per_sample"))
timing_metadata = timing_document.get("metadata")
if not isinstance(timing_metadata, Mapping):
raise ValueError("timing document has no metadata object")
_provenance(timing_metadata, task)
if timing_metadata.get("generation_mode") != "conditional_prefix":
raise ValueError("timing metadata is not bound to conditional-prefix generation")
return {
"task_id": task.task_id,
"model": task.model,
"dataset": task.dataset,
"category": task.category,
"steps": task.steps,
"sample_count": task.sample_count,
"seed": task.seed,
"protocol": task.protocol,
"provenance": task.provenance,
**quality,
"generation_seconds_per_sample": seconds_value,
"source_sha256": provenance["source_sha256"],
"config_sha256": provenance["config_sha256"],
"checkpoint_sha256": provenance["checkpoint_sha256"],
"checkpoint_lock_id": provenance["checkpoint_lock_id"],
"checkpoint_teacher_family": provenance["checkpoint_teacher_family"],
"adapter_identity": provenance["adapter_identity"],
"environment": (
provenance["environment"]
if isinstance(provenance["environment"], str)
else json.dumps(provenance["environment"], sort_keys=True)
),
"conditioning_manifest_sha256": task.conditioning_manifest_sha256,
}
def _failure(task: ConditionalMatrixTask, error: Exception) -> dict[str, Any]:
return {
"status": "failed",
"task_id": task.task_id,
"model": task.model,
"dataset": task.dataset,
"category": task.category,
"steps": task.steps,
"reason": str(error),
}
def _write_outputs(output_dir: Path, report: ConditionalAggregateReport) -> None:
output_dir = output_dir.absolute()
output_dir.mkdir(parents=True, exist_ok=True)
rows = list(report.rows)
long_rows = [
{
"task_id": row["task_id"],
"model": row["model"],
"dataset": row["dataset"],
"category": row["category"],
"steps": row["steps"],
"sample_count": row["sample_count"],
"seed": row["seed"],
"protocol": row["protocol"],
"provenance": row["provenance"],
"metric": metric,
"value": row[metric],
"source_sha256": row["source_sha256"],
"config_sha256": row["config_sha256"],
"checkpoint_sha256": row["checkpoint_sha256"],
"checkpoint_lock_id": row["checkpoint_lock_id"],
"adapter_identity": row["adapter_identity"],
"environment": row["environment"],
"conditioning_manifest_sha256": row["conditioning_manifest_sha256"],
}
for row in rows
for metric in ALL_METRICS
]
long_columns = (
"task_id",
"model",
"dataset",
"category",
"steps",
"sample_count",
"seed",
"protocol",
"provenance",
"metric",
"value",
"source_sha256",
"config_sha256",
"checkpoint_sha256",
"checkpoint_lock_id",
"adapter_identity",
"environment",
"conditioning_manifest_sha256",
)
wide_columns = (
"task_id",
"model",
"dataset",
"category",
"steps",
"sample_count",
"seed",
"protocol",
*ALL_METRICS,
"provenance",
"source_sha256",
"config_sha256",
"checkpoint_sha256",
"checkpoint_lock_id",
"checkpoint_teacher_family",
"adapter_identity",
"environment",
"conditioning_manifest_sha256",
)
_write_csv(output_dir / "conditional_results_long.csv", long_rows, long_columns)
_write_csv(output_dir / "conditional_results_wide.csv", rows, wide_columns)
failures = list(report.failures) + [
{**item, "status": "unsupported"} for item in report.unsupported
]
_write_csv(
output_dir / "conditional_failures.csv",
failures,
("status", "task_id", "model", "dataset", "category", "steps", "reason"),
)
(output_dir / "README.md").write_text(
"\n".join(
[
f"# Conditional Baseline Results ({CONDITIONAL_AGGREGATE_SCHEMA})",
"",
f"- complete: `{str(report.complete).lower()}`",
f"- expected supported tasks: `{report.expected_tasks}`",
f"- valid result rows: `{len(report.rows)}`",
f"- failures: `{len(report.failures)}`",
f"- unsupported cells: `{len(report.unsupported)}`",
"",
]
),
encoding="utf-8",
)
def aggregate_conditional(
root: Path,
*,
strict: bool = True,
partial: bool = False,
output_dir: Path | None = None,
registry: ExperimentRegistry | None = None,
) -> ConditionalAggregateReport:
"""Validate all conditional tasks and optionally publish summary tables."""
root = Path(root).resolve()
registry = registry or load_registry(root / "configs" / "experiments.yaml")
tasks = build_conditional_matrix(registry, root=root)
rows: list[dict[str, Any]] = []
failures: list[dict[str, Any]] = []
for task in tasks:
try:
rows.append(_validate_task(task))
except (OSError, ValueError, TypeError) as error:
failures.append(_failure(task, error))
unsupported = conditional_unsupported_inventory(registry)
complete = not failures and len(rows) == len(tasks)
report = ConditionalAggregateReport(
rows=tuple(rows),
failures=tuple(failures),
unsupported=tuple(unsupported),
expected_tasks=len(tasks),
complete=complete,
output_dir=Path(output_dir).resolve() if output_dir is not None else None,
)
if strict and not partial and not complete:
raise IncompleteMatrixError(failures)
if output_dir is not None:
_write_outputs(Path(output_dir), report)
return report
__all__ = [
"ALL_METRICS",
"CONDITIONAL_AGGREGATE_SCHEMA",
"ConditionalAggregateReport",
"QUALITY_METRICS",
"aggregate_conditional",
]
|