File size: 16,576 Bytes
6c5f29f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 | """Train a non-oracle utility writer and evaluate it on natural packages.
This is the deployable-writer diagnostic for OracleMem. Training may use oracle
coverage labels on train packages, but test-time selection uses only visible
candidate metadata through ``EstimatedUtilityModel.predict``. The reported
ratios are still scored against exact finite-package optima.
"""
from __future__ import annotations
import argparse
from collections import defaultdict
import json
import math
import sys
from pathlib import Path
from typing import Any, Mapping, Sequence
REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from oraclemem.evaluate import (
LEARNED_ESTIMATOR_PROFILE,
LOCAL_LEARNED_ESTIMATOR_MODEL,
OracleMemInstance,
aggregate_results,
evaluate_instance,
generate_named_distribution,
objective_value,
train_feature_utility_estimator,
)
from llm_memory_validation.evaluate_human_style_examples import (
build_instance as build_human_instance,
evaluate_human_package,
load_examples,
parse_tokens,
)
from llm_memory_validation.run_mem0_natural_baseline import (
load_package,
package_instance,
resolved_queries,
write_json,
)
DEFAULT_METHODS = (
"opt",
"oracle_gvt",
"estimated_gvt",
"estimated_utility",
"memgpt_tiered",
"amem_graph",
"amac_admission",
"mem0_extract",
"density_only",
"greedy",
"fact_only",
"summary_only",
"recency_raw",
"no_tombstone_opt",
)
def build_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(
description=(
"Train a visible-feature OracleMem utility estimator on synthetic "
"and model-annotated natural packages, then test on a human-edited "
"finite package with exact OPT scoring."
)
)
parser.add_argument(
"--human-examples-jsonl",
default="llm_memory_validation/human_style_examples/examples_100.jsonl",
help="Human-edited JSONL package used for held-out evaluation.",
)
parser.add_argument(
"--train-natural-package-dir",
action="append",
default=["llm_memory_validation/oraclemem_natural_200_gemini_v2/coverage_package"],
help=(
"Natural coverage package directory to use for train labels. "
"Can be supplied multiple times. Defaults to Natural-200."
),
)
parser.add_argument(
"--train-natural-limit",
type=int,
default=None,
help="Optional per-package cap on natural train queries.",
)
parser.add_argument(
"--natural-train-weight",
type=int,
default=1,
help=(
"Integer replication weight for allowed natural train instances. "
"This changes estimator fitting only; manifests report weighted and "
"unweighted counts."
),
)
parser.add_argument(
"--tune-natural-train-weight",
action="store_true",
help=(
"Choose natural-train weight and ridge from train-only validation "
"labels before fitting the final estimator."
),
)
parser.add_argument(
"--candidate-natural-train-weights",
default="1,2,3,5,8,10,15,20,30,50",
help="Comma or space separated natural weights for train-only tuning.",
)
parser.add_argument(
"--candidate-ridges",
default="0.05,0.25,1.0,2.0",
help="Comma or space separated ridge values for train-only tuning.",
)
parser.add_argument(
"--validation-natural-stride",
type=int,
default=5,
help="Use every Nth allowed natural train instance as train-only validation.",
)
parser.add_argument(
"--validation-synthetic-fraction",
type=float,
default=0.20,
help="Fraction of synthetic train seeds reserved for train-only validation.",
)
parser.add_argument(
"--validation-synthetic-budgets",
default="4,6",
help="Synthetic validation budgets used only for hyperparameter selection.",
)
parser.add_argument(
"--validation-natural-budgets",
default="30,60,100",
help="Natural validation budgets used only for hyperparameter selection.",
)
parser.add_argument(
"--n-synthetic-train-seeds",
type=int,
default=200,
help="Use synthetic train seeds 0..N-1. Set 0 to disable synthetic train data.",
)
parser.add_argument(
"--synthetic-distributions",
default="base,update_chain,temporal_interval,scope_shift_v2,density_trap_v2",
help="Comma or space separated synthetic train distributions.",
)
parser.add_argument(
"--normal-count",
type=int,
default=3,
help="Synthetic normal fact count.",
)
parser.add_argument(
"--update-count",
type=int,
default=2,
help="Synthetic update/tombstone pair count.",
)
parser.add_argument(
"--budgets",
default="150,300,600,1000",
help="Comma or space separated held-out test budgets.",
)
parser.add_argument(
"--methods",
default=",".join(DEFAULT_METHODS),
help="Comma or space separated evaluation methods.",
)
parser.add_argument(
"--eval-coverage-package-dir",
action="append",
default=["llm_memory_validation/natural_adjudicated_100_gemini_flash/coverage_package"],
help=(
"Held-out coverage package directory to evaluate with exact package OPT. "
"Can be supplied multiple times. Defaults to the adjudicated natural package."
),
)
parser.add_argument(
"--skip-coverage-eval",
action="store_true",
help="Evaluate only the human-style examples package.",
)
parser.add_argument(
"--eval-coverage-limit",
type=int,
default=None,
help="Optional per-held-out coverage-package query cap.",
)
parser.add_argument(
"--eval-coverage-budgets",
default="30,60,100",
help="Comma or space separated held-out coverage-package budgets.",
)
parser.add_argument(
"--eval-coverage-methods",
default=(
"opt,oracle_gvt,estimated_gvt,estimated_utility,memgpt_tiered,"
"amem_graph,amac_admission,mem0_extract,density_only,summary_only,"
"fact_only,recency_raw"
),
help="Comma or space separated methods for held-out coverage-package evaluation.",
)
parser.add_argument(
"--allow-natural-train-overlap",
action="store_true",
help=(
"Do not exclude held-out coverage-package query ids from natural train "
"packages. The default is safer and excludes overlaps."
),
)
parser.add_argument(
"--estimator-ridge",
type=float,
default=0.25,
help="Ridge penalty for the visible-feature linear estimator.",
)
parser.add_argument(
"--estimated-noise-scale",
type=float,
default=0.0,
help="Optional deterministic noise scale applied to learned predictions.",
)
parser.add_argument(
"--estimated-noise-seed",
type=int,
default=0,
help="Seed for deterministic learned-estimator prediction noise.",
)
parser.add_argument(
"--out-dir",
default="llm_memory_validation/learned_writer_deployable_noapi",
help="Output directory.",
)
return parser
def synthetic_train_instances(
*,
n_seeds: int,
distributions: Sequence[str],
normal_count: int,
update_count: int,
) -> list[OracleMemInstance]:
if n_seeds <= 0:
return []
return [
generate_named_distribution(
distribution,
seed,
normal_count=normal_count,
update_count=update_count,
)
for distribution in distributions
for seed in range(n_seeds)
]
def natural_train_instances(
package_dirs: Sequence[str],
*,
limit: int | None,
exclude_query_ids: set[str] | None = None,
) -> tuple[list[OracleMemInstance], list[dict[str, Any]]]:
instances: list[OracleMemInstance] = []
manifest_rows: list[dict[str, Any]] = []
exclude_query_ids = set(exclude_query_ids or ())
for package_dir_text in package_dirs:
package_dir = Path(package_dir_text)
data = load_package(package_dir)
all_queries = resolved_queries(data, limit)
excluded = [
query
for query in all_queries
if str(query.get("query_id", "")) in exclude_query_ids
]
queries = [
query
for query in all_queries
if str(query.get("query_id", "")) not in exclude_query_ids
]
before = len(instances)
for query in queries:
instance = package_instance(data, query)
if instance.candidates and any(weight > 0 for weight in instance.unit_weights.values()):
instances.append(instance)
manifest_rows.append(
{
"package_dir": str(package_dir),
"resolved_queries_before_exclusion": len(all_queries),
"excluded_query_ids": sorted(str(query["query_id"]) for query in excluded),
"excluded_query_count": len(excluded),
"resolved_queries": len(queries),
"usable_instances": len(instances) - before,
}
)
return instances, manifest_rows
def coverage_eval_query_ids(package_dirs: Sequence[str], *, limit: int | None) -> dict[str, list[str]]:
query_ids: dict[str, list[str]] = {}
for package_dir_text in package_dirs:
package_dir = Path(package_dir_text)
data = load_package(package_dir)
query_ids[str(package_dir)] = [
str(query.get("query_id", ""))
for query in resolved_queries(data, limit)
]
return query_ids
def weighted_train_instances(
synthetic_instances: Sequence[OracleMemInstance],
natural_instances: Sequence[OracleMemInstance],
*,
natural_weight: int,
) -> list[OracleMemInstance]:
weight = max(0, int(natural_weight))
return [*synthetic_instances, *(list(natural_instances) * weight)]
def estimator_coefficients(model: Any, limit: int = 25) -> list[dict[str, float | str]]:
rows = [
{"feature": name, "weight": float(weight), "abs_weight": abs(float(weight))}
for name, weight in zip(model.feature_names, model.weights)
]
rows.sort(key=lambda row: (-float(row["abs_weight"]), str(row["feature"])))
return rows[:limit]
def write_transfer_report(
out_dir: Path,
*,
train_manifest: Mapping[str, Any],
summary: Mapping[str, Any],
) -> None:
lines = [
"# Learned Writer Transfer Report",
"",
"This run trains a local visible-feature utility estimator on train-only oracle labels and evaluates held-out memory-writing decisions against exact finite-package OPT.",
"",
"## Train Data",
"",
f"- Synthetic train instances: {train_manifest['synthetic_train_instances']}",
f"- Natural train instances: {train_manifest['natural_train_instances']}",
f"- Total train instances: {train_manifest['total_train_instances']}",
f"- Train candidates: {train_manifest['train_candidate_count']}",
f"- Ridge: {train_manifest['estimator_ridge']}",
f"- Test package: `{train_manifest['human_examples_jsonl']}`",
"",
"## Claim Boundary",
"",
"- Oracle coverage is used to create train labels only.",
"- Held-out estimated-writer decisions use visible candidate metadata only.",
"- The human-edited test package is schema-valid and exact-scored, but it is not an inter-annotator agreement study.",
"",
"## Held-Out Package Ratios",
"",
]
methods = sorted(summary.get("methods", []))
by_budget = {}
for row in summary.get("by_budget_method", []):
by_budget.setdefault(int(row["budget"]), {})[str(row["method"])] = row
for budget in sorted(by_budget):
lines.append(f"### Budget {budget}")
for method in methods:
row = by_budget[budget].get(method)
if row is None:
continue
lines.append(
"- `{method}`: ratio_to_opt={ratio:.3f}, objective={objective:.3f}, cost={cost:.1f}".format(
method=method,
ratio=float(row.get("mean_ratio_to_opt", 0.0)),
objective=float(row.get("mean_objective", 0.0)),
cost=float(row.get("mean_selected_cost", 0.0)),
)
)
lines.append("")
(out_dir / "REPORT.md").write_text("\n".join(lines), encoding="utf-8")
def main(argv: Sequence[str] | None = None) -> int:
args = build_parser().parse_args(argv)
out_dir = Path(args.out_dir)
out_dir.mkdir(parents=True, exist_ok=True)
synthetic_distributions = parse_tokens(args.synthetic_distributions)
synthetic_instances = synthetic_train_instances(
n_seeds=args.n_synthetic_train_seeds,
distributions=synthetic_distributions,
normal_count=args.normal_count,
update_count=args.update_count,
)
natural_instances, natural_manifest = natural_train_instances(
args.train_natural_package_dir,
limit=args.train_natural_limit,
)
train_instances = [*synthetic_instances, *natural_instances]
if not train_instances:
raise ValueError("at least one synthetic or natural train instance is required")
estimator = train_feature_utility_estimator(
train_instances,
train_distributions=(
*(f"synthetic:{name}" for name in synthetic_distributions),
*(f"natural:{Path(path).name}" for path in args.train_natural_package_dir),
),
train_seeds=tuple(range(max(0, args.n_synthetic_train_seeds))),
estimator_model=LOCAL_LEARNED_ESTIMATOR_MODEL,
estimator_profile=LEARNED_ESTIMATOR_PROFILE,
ridge=args.estimator_ridge,
noise_scale=args.estimated_noise_scale,
noise_seed=args.estimated_noise_seed,
)
human_examples_path = Path(args.human_examples_jsonl)
human_rows = load_examples(human_examples_path)
human_instance = build_human_instance(human_rows)
budgets = tuple(int(token) for token in parse_tokens(args.budgets))
methods = parse_tokens(args.methods)
results = evaluate_human_package(
human_instance,
budgets,
methods,
estimator_model=estimator.estimator_model,
estimator_profile=estimator.estimator_profile,
estimator_state=estimator,
)
paths = write_benchmark_outputs(results, out_dir)
write_human_report(out_dir, human_examples_path, human_rows, results)
train_manifest = {
"human_examples_jsonl": str(human_examples_path),
"synthetic_train_distributions": list(synthetic_distributions),
"synthetic_train_seeds": list(range(max(0, args.n_synthetic_train_seeds))),
"synthetic_train_instances": len(synthetic_instances),
"natural_train_packages": natural_manifest,
"natural_train_instances": len(natural_instances),
"total_train_instances": len(train_instances),
"train_candidate_count": sum(len(instance.candidates) for instance in train_instances),
"estimator_model": estimator.estimator_model,
"estimator_profile": estimator.estimator_profile,
"estimator_ridge": args.estimator_ridge,
"estimated_noise_scale": args.estimated_noise_scale,
"estimated_noise_seed": args.estimated_noise_seed,
"top_coefficients": estimator_coefficients(estimator),
"decision_features": "visible candidate metadata only at held-out test time",
"oracle_coverage_used_for_training": True,
"oracle_coverage_used_for_test_decision": False,
**paths,
}
write_json(out_dir / "train_manifest.json", train_manifest)
summary = json.loads((out_dir / "summary.json").read_text(encoding="utf-8"))
write_transfer_report(out_dir, train_manifest=train_manifest, summary=summary)
print(json.dumps(train_manifest, indent=2, sort_keys=True))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|