File size: 20,409 Bytes
21c7db9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 | #!/usr/bin/env python3
"""Generate sweep summaries, charts, and anti-hacking checks for HF training."""
from __future__ import annotations
import argparse
from collections import Counter
import json
from pathlib import Path
from typing import Any
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt # noqa: E402
ROOT = Path(__file__).resolve().parents[1]
REWARD_MIN = 0.001
REWARD_MAX = 0.999
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Summarize PolyGuard HF training sweeps.")
parser.add_argument("--sweep-dir", default="outputs/reports/sweeps")
parser.add_argument("--plot-dir", default="outputs/plots")
parser.add_argument("--output", default="outputs/reports/hf_sweep_summary.json")
parser.add_argument("--anti-hacking-output", default="outputs/reports/anti_hacking_overfit_report.json")
parser.add_argument(
"--mode",
choices=["full", "sft-baseline"],
default="full",
help="Report mode. SFT baseline mode treats GRPO artifacts as optional.",
)
return parser.parse_args()
def _read_json(path: Path) -> dict[str, Any]:
if not path.exists():
return {}
try:
payload = json.loads(path.read_text(encoding="utf-8"))
except json.JSONDecodeError:
return {}
return payload if isinstance(payload, dict) else {}
def _read_history(path: Path) -> list[dict[str, Any]]:
if not path.exists():
return []
try:
payload = json.loads(path.read_text(encoding="utf-8"))
except json.JSONDecodeError:
return []
return [row for row in payload if isinstance(row, dict)] if isinstance(payload, list) else []
def _read_jsonl(path: Path) -> list[dict[str, Any]]:
if not path.exists():
return []
rows: list[dict[str, Any]] = []
with path.open("r", encoding="utf-8") as handle:
for line in handle:
line = line.strip()
if not line:
continue
try:
payload = json.loads(line)
except json.JSONDecodeError:
continue
if isinstance(payload, dict):
rows.append(payload)
return rows
def _as_float(value: Any, default: float = 0.0) -> float:
try:
return float(value)
except (TypeError, ValueError):
return default
def _is_reward_value(value: Any) -> bool:
if isinstance(value, bool) or not isinstance(value, int | float):
return False
number = float(value)
return REWARD_MIN <= number <= REWARD_MAX and round(number, 3) == number
def _scan_reward_payload(payload: Any, failures: list[str], path: str) -> None:
if isinstance(payload, dict):
for key, value in payload.items():
next_path = f"{path}.{key}" if path else str(key)
if key in {"reward", "env_reward", "avg_reward", "avg_env_reward"} or key.endswith("_score"):
if not _is_reward_value(value):
failures.append(f"{next_path}={value!r}")
elif key in {"reward_breakdown", "primary_reward_channels", "avg_reward_components", "avg_primary_reward_channels"}:
if isinstance(value, dict):
for sub_key, sub_value in value.items():
if not _is_reward_value(sub_value):
failures.append(f"{next_path}.{sub_key}={sub_value!r}")
else:
failures.append(f"{next_path}=not_dict")
else:
_scan_reward_payload(value, failures, next_path)
elif isinstance(payload, list):
for idx, item in enumerate(payload):
_scan_reward_payload(item, failures, f"{path}[{idx}]")
def _history_series(history: list[dict[str, Any]], names: tuple[str, ...]) -> tuple[list[int], list[float]]:
xs: list[int] = []
ys: list[float] = []
for idx, row in enumerate(history, start=1):
for name in names:
if name in row:
xs.append(idx)
ys.append(_as_float(row.get(name)))
break
return xs, ys
def _plot_placeholder(path: Path, title: str) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
fig, ax = plt.subplots(figsize=(8, 4.5))
ax.text(0.5, 0.5, "No completed sweep data yet", ha="center", va="center", fontsize=12)
ax.set_axis_off()
ax.set_title(title)
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
def _bar_chart(path: Path, title: str, labels: list[str], series: dict[str, list[float]], ylabel: str = "Reward") -> None:
if not labels or not series:
_plot_placeholder(path, title)
return
path.parent.mkdir(parents=True, exist_ok=True)
fig, ax = plt.subplots(figsize=(10, 5.2))
width = 0.8 / max(1, len(series))
x_positions = list(range(len(labels)))
offsets = [(-0.4 + (idx + 0.5) * width) for idx in range(len(series))]
for offset, (name, values) in zip(offsets, series.items(), strict=False):
ax.bar([x + offset for x in x_positions], values, width=width, label=name)
ax.set_title(title)
ax.set_ylabel(ylabel)
ax.set_xticks(x_positions)
ax.set_xticklabels(labels, rotation=20, ha="right")
ax.set_ylim(0, max(1.0, max((max(vals) for vals in series.values() if vals), default=1.0) * 1.15))
ax.legend()
ax.grid(axis="y", alpha=0.24)
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
def _line_chart(
path: Path,
title: str,
curves: dict[str, tuple[list[int], list[float]]],
ylabel: str,
) -> None:
curves = {key: value for key, value in curves.items() if value[0] and value[1]}
if not curves:
_plot_placeholder(path, title)
return
path.parent.mkdir(parents=True, exist_ok=True)
fig, ax = plt.subplots(figsize=(10, 5.2))
for label, (xs, ys) in curves.items():
ax.plot(xs, ys, marker="o", linewidth=1.6, markersize=3.5, label=label)
ax.set_title(title)
ax.set_xlabel("Logged step")
ax.set_ylabel(ylabel)
ax.grid(alpha=0.24)
ax.legend()
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
def _safe_model_label(model_id: str, fallback: str) -> str:
if model_id:
return model_id.split("/")[-1].replace("-Instruct", "")
return fallback
def _bandit_chart_label(label: str) -> str:
if "bandit" in label.lower():
return label
if "qwen" in label.lower():
return f"{label} + Bandits"
return label
def _summarize_run(run_dir: Path, *, mode: str) -> dict[str, Any]:
sft = _read_json(run_dir / "sft_trl_run.json")
grpo = _read_json(run_dir / "grpo_trl_run.json")
sft_inference = _read_json(run_dir / "postsave_inference_sft.json")
grpo_inference = _read_json(run_dir / "postsave_inference_grpo.json")
error = _read_json(run_dir / "error.json")
sft_history = _read_history(run_dir / "sft_history.json")
grpo_history = _read_history(run_dir / "grpo_history.json")
reward_rows = _read_jsonl(run_dir / "grpo_reward_components.jsonl")
sft_only = mode == "sft-baseline"
model_id = str(grpo.get("model_id") or sft.get("model_id") or error.get("model_id") or run_dir.name)
fallback_detected = any(
"fallback" in str(payload.get("backend", "")).lower()
or str(payload.get("model_source", "")).lower() == "fallback_policy"
for payload in ([sft, sft_inference] if sft_only else [sft, grpo, sft_inference, grpo_inference])
)
reward_failures: list[str] = []
if not sft_only:
_scan_reward_payload(grpo, reward_failures, "grpo")
_scan_reward_payload(sft_inference, reward_failures, "sft_inference")
if not sft_only:
_scan_reward_payload(grpo_inference, reward_failures, "grpo_inference")
for idx, row in enumerate(reward_rows):
_scan_reward_payload(row, reward_failures, f"reward_log[{idx}]")
legal_count = sum(1 for row in reward_rows if row.get("legal") is True)
reward_count = len(reward_rows)
exploit_count = sum(
1
for row in reward_rows
if any(
marker in str(row.get("termination_reason", "")).lower()
for marker in ["cheat", "exploit", "abuse", "timeout", "invalid"]
)
)
selected = [str(row.get("selected_candidate_id") or row.get("generated_candidate_id") or "") for row in reward_rows]
selected = [item for item in selected if item]
counts = Counter(selected)
top_candidate_rate = (max(counts.values()) / len(selected)) if selected else 0.0
candidate_diversity = (len(counts) / len(selected)) if selected else 0.0
train_reward = _as_float((grpo.get("reward_summary") or {}).get("avg_reward"))
if sft_only:
holdout_reward = _as_float(sft_inference.get("avg_env_reward"))
train_reward = holdout_reward
else:
holdout_reward = _as_float(grpo_inference.get("avg_env_reward"), train_reward)
train_holdout_gap = round(train_reward - holdout_reward, 3)
validity = _as_float(sft_inference.get("valid_rate") if sft_only else grpo_inference.get("valid_rate"), 0.0)
completed = sft.get("status") == "ok" if sft_only else sft.get("status") == "ok" and grpo.get("status") == "ok"
return {
"run_id": run_dir.name,
"training_mode": mode,
"model_id": model_id,
"label": _safe_model_label(model_id, run_dir.name),
"status": "failed" if error else ("completed" if completed else "incomplete"),
"error": error.get("error", ""),
"sft_backend": sft.get("backend", ""),
"sft_examples": int(sft.get("examples_used", 0) or 0),
"sft_train_loss": _as_float(sft.get("train_loss")),
"sft_runtime": _as_float(sft.get("train_runtime")),
"grpo_backend": grpo.get("backend", ""),
"grpo_records": int(grpo.get("records", 0) or 0),
"grpo_avg_reward": train_reward,
"sft_inference_reward": _as_float(sft_inference.get("avg_env_reward")),
"sft_valid_rate": _as_float(sft_inference.get("valid_rate")),
"sft_latency_seconds": _as_float(sft_inference.get("avg_latency_seconds")),
"grpo_inference_reward": holdout_reward,
"grpo_valid_rate": validity,
"grpo_latency_seconds": _as_float(grpo_inference.get("avg_latency_seconds")),
"train_holdout_gap": train_holdout_gap,
"fallback_detected": fallback_detected,
"reward_range_ok": not reward_failures,
"reward_range_failures": reward_failures[:25],
"exploit_rate": round(exploit_count / reward_count, 3) if reward_count else 0.0,
"legal_rate": round(legal_count / reward_count, 3) if reward_count else 0.0,
"candidate_diversity": round(candidate_diversity, 3),
"top_candidate_rate": round(top_candidate_rate, 3),
"reward_components": (grpo.get("reward_summary") or {}).get("avg_reward_components", {}),
"primary_reward_channels": (grpo.get("reward_summary") or {}).get("avg_primary_reward_channels", {}),
"sft_history": sft_history,
"grpo_history": grpo_history,
"artifact_paths": {
"sft": sft.get("artifact_path", ""),
"grpo": grpo.get("artifact_path", ""),
},
}
def _write_charts(rows: list[dict[str, Any]], plot_dir: Path, *, mode: str) -> dict[str, str]:
completed = [row for row in rows if row["status"] == "completed"]
labels = [_bandit_chart_label(str(row["label"])) for row in completed]
charts = {
"sft_vs_grpo_reward": plot_dir / "sft_vs_grpo_reward.png",
"sft_loss_curves": plot_dir / "sft_loss_curves.png",
"qwen_model_sft_reward": plot_dir / "qwen_model_sft_reward.png",
"qwen_model_sft_loss": plot_dir / "qwen_model_sft_loss.png",
"sft_validity_reward": plot_dir / "sft_validity_reward.png",
"grpo_reward_curves": plot_dir / "grpo_reward_curves.png",
"qwen_model_grpo_reward": plot_dir / "qwen_model_grpo_reward.png",
"reward_component_bars": plot_dir / "reward_component_bars.png",
"anti_cheat_failure_rates": plot_dir / "anti_cheat_failure_rates.png",
"train_holdout_gap": plot_dir / "train_holdout_gap.png",
"inference_validity_reward": plot_dir / "inference_validity_reward.png",
"inference_latency_validity": plot_dir / "inference_latency_validity.png",
}
_bar_chart(
charts["sft_vs_grpo_reward"],
"SFT Baseline vs GRPO + Bandits Policy Reward",
labels,
{
"SFT inference reward": [row["sft_inference_reward"] for row in completed],
"GRPO + Bandits inference reward": [row["grpo_inference_reward"] for row in completed],
},
)
_line_chart(
charts["sft_loss_curves"],
"Qwen + Bandits SFT Training Loss Curves",
{
_bandit_chart_label(str(row["label"])): _history_series(row["sft_history"], ("loss", "train_loss"))
for row in completed
},
ylabel="Loss",
)
_bar_chart(
charts["qwen_model_sft_reward"],
"Qwen + Bandits Model Sweep SFT Reward",
labels,
{"SFT inference reward": [row["sft_inference_reward"] for row in completed]},
)
_bar_chart(
charts["qwen_model_sft_loss"],
"Qwen + Bandits Model Sweep SFT Loss",
labels,
{"SFT train loss": [row["sft_train_loss"] for row in completed]},
ylabel="Loss",
)
_bar_chart(
charts["sft_validity_reward"],
"SFT Inference Validity and Reward",
labels,
{
"SFT valid rate": [row["sft_valid_rate"] for row in completed],
"SFT reward": [row["sft_inference_reward"] for row in completed],
},
ylabel="Rate / reward",
)
_line_chart(
charts["grpo_reward_curves"],
"GRPO + Bandits Reward Curves",
{
_bandit_chart_label(str(row["label"])): _history_series(
row["grpo_history"],
("reward", "rewards/environment_reward_verifier", "mean_reward", "train_reward"),
)
for row in completed
},
ylabel="Reward",
)
_bar_chart(
charts["qwen_model_grpo_reward"],
"Qwen + Bandits Model Sweep GRPO Reward",
labels,
{"GRPO + Bandits train reward": [row["grpo_avg_reward"] for row in completed]},
)
component_names = sorted(
{
key
for row in completed
for key, value in dict(row.get("reward_components") or {}).items()
if isinstance(value, int | float)
}
)
component_means = []
for key in component_names:
values = [_as_float((row.get("reward_components") or {}).get(key)) for row in completed]
component_means.append(round(sum(values) / len(values), 3) if values else 0.0)
_bar_chart(
charts["reward_component_bars"],
"Mean GRPO Reward Components",
component_names,
{"component reward": component_means},
)
_bar_chart(
charts["anti_cheat_failure_rates"],
"Anti-Cheat and Failure Visibility",
labels,
{
"exploit/invalid rate": [row["exploit_rate"] for row in completed],
"illegal rate": [round(1.0 - row["legal_rate"], 3) for row in completed],
"candidate collapse": [row["top_candidate_rate"] for row in completed],
},
ylabel="Rate",
)
_bar_chart(
charts["train_holdout_gap"],
"Train vs Holdout Reward Gap",
labels,
{"train - holdout": [abs(row["train_holdout_gap"]) for row in completed]},
ylabel="Absolute reward gap",
)
_bar_chart(
charts["inference_validity_reward"],
"Inference Validity and Reward",
labels,
{
"GRPO valid rate": [row["grpo_valid_rate"] for row in completed],
"GRPO holdout reward": [row["grpo_inference_reward"] for row in completed],
},
ylabel="Rate / reward",
)
_bar_chart(
charts["inference_latency_validity"],
"Inference Latency and Validity",
labels,
{
"SFT latency sec": [row["sft_latency_seconds"] for row in completed],
"GRPO latency sec": [row["grpo_latency_seconds"] for row in completed],
"GRPO valid rate": [row["grpo_valid_rate"] for row in completed],
},
ylabel="Seconds / rate",
)
chart_index: dict[str, str] = {}
for key, path in charts.items():
try:
chart_index[key] = str(path.relative_to(ROOT))
except ValueError:
chart_index[key] = str(path)
return chart_index
def generate_report(
sweep_dir: Path,
plot_dir: Path,
output_path: Path,
anti_hacking_output: Path,
mode: str = "full",
) -> tuple[dict[str, Any], dict[str, Any]]:
run_dirs = sorted(path for path in sweep_dir.iterdir() if path.is_dir()) if sweep_dir.exists() else []
mode = "sft-baseline" if mode in {"sft", "sft-only", "sft_baseline", "sft-baseline"} else "full"
rows = [_summarize_run(run_dir, mode=mode) for run_dir in run_dirs]
chart_paths = _write_charts(rows, plot_dir, mode=mode)
completed = [row for row in rows if row["status"] == "completed"]
failed = [row for row in rows if row["status"] == "failed"]
warnings: list[str] = []
if not completed:
warnings.append("no_completed_models")
for row in completed:
if row["fallback_detected"]:
warnings.append(f"{row['label']}:fallback_detected")
if not row["reward_range_ok"]:
warnings.append(f"{row['label']}:reward_range_violation")
if mode == "sft-baseline":
if row["sft_valid_rate"] < 0.8:
warnings.append(f"{row['label']}:low_sft_validity")
else:
if row["exploit_rate"] > 0.35:
warnings.append(f"{row['label']}:high_exploit_rate")
if row["top_candidate_rate"] > 0.85 and row["candidate_diversity"] < 0.2:
warnings.append(f"{row['label']}:candidate_collapse")
if row["grpo_valid_rate"] < 0.8:
warnings.append(f"{row['label']}:low_validity")
if abs(row["train_holdout_gap"]) > 0.25:
warnings.append(f"{row['label']}:large_train_holdout_gap")
public_rows = [
{key: value for key, value in row.items() if key not in {"sft_history", "grpo_history"}}
for row in rows
]
summary = {
"status": "ok" if completed else "incomplete",
"training_mode": mode,
"completed_models": len(completed),
"failed_or_skipped_models": len(failed),
"models": public_rows,
"charts": chart_paths,
}
anti_hacking = {
"passed": bool(completed) and not warnings,
"training_mode": mode,
"warnings": warnings,
"completed_models": [row["model_id"] for row in completed],
"failed_or_skipped_models": [{"model_id": row["model_id"], "error": row["error"]} for row in failed],
"checks": {
"reward_bounds": [REWARD_MIN, REWARD_MAX],
"reward_precision": 3,
"fallback_backends_rejected": True,
"exploit_rate_threshold": 0.35,
"train_holdout_gap_threshold": 0.25,
"min_validity_rate": 0.8,
},
}
output_path.parent.mkdir(parents=True, exist_ok=True)
anti_hacking_output.parent.mkdir(parents=True, exist_ok=True)
output_path.write_text(json.dumps(summary, ensure_ascii=True, indent=2), encoding="utf-8")
anti_hacking_output.write_text(json.dumps(anti_hacking, ensure_ascii=True, indent=2), encoding="utf-8")
return summary, anti_hacking
def main() -> None:
args = parse_args()
summary, anti_hacking = generate_report(
sweep_dir=ROOT / args.sweep_dir,
plot_dir=ROOT / args.plot_dir,
output_path=ROOT / args.output,
anti_hacking_output=ROOT / args.anti_hacking_output,
mode=args.mode,
)
print(
json.dumps(
{
"hf_sweep_summary": summary.get("status"),
"completed_models": summary.get("completed_models"),
"anti_hacking_passed": anti_hacking.get("passed"),
},
ensure_ascii=True,
)
)
if __name__ == "__main__":
main()
|