Spaces:
Running
Running
File size: 18,383 Bytes
c452421 | 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 | from __future__ import annotations
import argparse
import json
from collections import defaultdict
from pathlib import Path
from statistics import mean
from typing import Any, Dict, Iterable, List, Optional
def _load_json(path: Path) -> Dict[str, Any]:
if not path.exists():
return {}
try:
return json.loads(path.read_text(encoding="utf-8"))
except Exception:
return {}
def _load_jsonl(path: Path) -> List[Dict[str, Any]]:
if not path.exists():
return []
rows: List[Dict[str, Any]] = []
for line in path.read_text(encoding="utf-8", errors="ignore").splitlines():
line = line.strip()
if not line:
continue
try:
item = json.loads(line)
except json.JSONDecodeError:
continue
if isinstance(item, dict):
rows.append(item)
return rows
def _get(payload: Dict[str, Any], dotted_key: str, default: Any = None) -> Any:
cur: Any = payload
for part in dotted_key.split("."):
if not isinstance(cur, dict) or part not in cur:
return default
cur = cur[part]
return cur
def _as_float(value: Any, default: float = 0.0) -> float:
try:
if value is None:
return default
return float(value)
except (TypeError, ValueError):
return default
def _ensure_matplotlib():
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
return plt
def _save_placeholder(path: Path, title: str, message: str) -> None:
plt = _ensure_matplotlib()
fig, ax = plt.subplots(figsize=(10, 5.4))
ax.axis("off")
ax.text(0.5, 0.62, title, ha="center", va="center", fontsize=17, fontweight="bold")
ax.text(0.5, 0.42, message, ha="center", va="center", fontsize=11, wrap=True)
fig.tight_layout()
fig.savefig(path, dpi=170)
plt.close(fig)
def _task_groups(rollouts: Iterable[Dict[str, Any]]) -> Dict[str, List[Dict[str, Any]]]:
grouped: Dict[str, List[Dict[str, Any]]] = defaultdict(list)
for row in rollouts:
grouped[str(row.get("task_id") or "unknown")].append(row)
return dict(sorted(grouped.items(), key=lambda item: item[0]))
def _save_keep_drop(path: Path, rollouts: List[Dict[str, Any]]) -> None:
if not rollouts:
_save_placeholder(path, "RFT Keep/Drop By Task", "No RFT rollouts found.")
return
plt = _ensure_matplotlib()
groups = _task_groups(rollouts)
labels = list(groups)
kept = [sum(1 for row in groups[label] if row.get("kept")) for label in labels]
dropped = [len(groups[label]) - kept[index] for index, label in enumerate(labels)]
fig, ax = plt.subplots(figsize=(12, 5.8))
ax.bar(labels, kept, color="#2ca25f", label="kept for RFT")
ax.bar(labels, dropped, bottom=kept, color="#d95f02", label="rejected")
ax.set_title("RFT Rejection Sampling: Kept vs Rejected Rollouts")
ax.set_ylabel("rollouts")
ax.tick_params(axis="x", rotation=25)
ax.grid(True, axis="y", alpha=0.25)
ax.legend()
for i, label in enumerate(labels):
total = len(groups[label])
rate = kept[i] / total if total else 0.0
ax.text(i, kept[i] + dropped[i] + 0.25, f"{rate:.0%}", ha="center", fontsize=9)
fig.tight_layout()
fig.savefig(path, dpi=170)
plt.close(fig)
def _save_score_by_task(path: Path, rollouts: List[Dict[str, Any]], min_score: Optional[float]) -> None:
if not rollouts:
_save_placeholder(path, "RFT Score Distribution", "No RFT rollouts found.")
return
plt = _ensure_matplotlib()
groups = _task_groups(rollouts)
labels = list(groups)
fig, ax = plt.subplots(figsize=(12, 5.8))
for index, label in enumerate(labels):
rows = groups[label]
scores = [_as_float(row.get("score")) for row in rows]
colors = ["#2ca25f" if row.get("kept") else "#d95f02" for row in rows]
xs = [index + ((i % 7) - 3) * 0.025 for i in range(len(rows))]
ax.scatter(xs, scores, c=colors, alpha=0.8, s=36, edgecolors="white", linewidths=0.4)
if min_score is not None:
ax.axhline(min_score, color="#333333", linestyle="--", linewidth=1.4, label=f"keep score >= {min_score:g}")
ax.legend()
ax.set_title("RFT Rollout Scores By Task")
ax.set_ylabel("filter score")
ax.set_xticks(range(len(labels)))
ax.set_xticklabels(labels, rotation=25, ha="right")
ax.grid(True, axis="y", alpha=0.25)
fig.tight_layout()
fig.savefig(path, dpi=170)
plt.close(fig)
def _save_fp_by_task(path: Path, rollouts: List[Dict[str, Any]], max_fp: Optional[float]) -> None:
if not rollouts:
_save_placeholder(path, "RFT False Positive Distribution", "No RFT rollouts found.")
return
plt = _ensure_matplotlib()
groups = _task_groups(rollouts)
labels = list(groups)
fig, ax = plt.subplots(figsize=(12, 5.8))
for index, label in enumerate(labels):
rows = groups[label]
fps = [_as_float(row.get("fp")) for row in rows]
colors = ["#2ca25f" if row.get("kept") else "#d95f02" for row in rows]
xs = [index + ((i % 7) - 3) * 0.025 for i in range(len(rows))]
ax.scatter(xs, fps, c=colors, alpha=0.8, s=36, edgecolors="white", linewidths=0.4)
if max_fp is not None:
ax.axhline(max_fp, color="#333333", linestyle="--", linewidth=1.4, label=f"keep fp <= {max_fp:g}")
ax.legend()
ax.set_title("RFT False Positives By Task")
ax.set_ylabel("false positives / episode")
ax.set_xticks(range(len(labels)))
ax.set_xticklabels(labels, rotation=25, ha="right")
ax.grid(True, axis="y", alpha=0.25)
fig.tight_layout()
fig.savefig(path, dpi=170)
plt.close(fig)
def _save_score_vs_fp(path: Path, rollouts: List[Dict[str, Any]], min_score: Optional[float], max_fp: Optional[float]) -> None:
if not rollouts:
_save_placeholder(path, "RFT Score vs False Positives", "No RFT rollouts found.")
return
plt = _ensure_matplotlib()
groups = _task_groups(rollouts)
palette = ["#1b9e77", "#7570b3", "#e7298a", "#66a61e", "#e6ab02", "#a6761d"]
fig, ax = plt.subplots(figsize=(10.5, 6.2))
for index, (task_id, rows) in enumerate(groups.items()):
kept_rows = [row for row in rows if row.get("kept")]
drop_rows = [row for row in rows if not row.get("kept")]
color = palette[index % len(palette)]
if drop_rows:
ax.scatter(
[_as_float(row.get("fp")) for row in drop_rows],
[_as_float(row.get("score")) for row in drop_rows],
marker="x",
s=50,
color=color,
alpha=0.55,
label=f"{task_id} rejected",
)
if kept_rows:
ax.scatter(
[_as_float(row.get("fp")) for row in kept_rows],
[_as_float(row.get("score")) for row in kept_rows],
marker="o",
s=60,
color=color,
edgecolors="black",
linewidths=0.4,
label=f"{task_id} kept",
)
if min_score is not None:
ax.axhline(min_score, color="#111111", linestyle="--", linewidth=1.2)
if max_fp is not None:
ax.axvline(max_fp, color="#111111", linestyle="--", linewidth=1.2)
ax.set_title("RFT Filter Boundary: Keep High Score, Low False Positives")
ax.set_xlabel("false positives / episode")
ax.set_ylabel("filter score")
ax.grid(True, alpha=0.25)
ax.legend(fontsize=7, ncol=2)
fig.tight_layout()
fig.savefig(path, dpi=170)
plt.close(fig)
def _save_timeline(path: Path, rollouts: List[Dict[str, Any]]) -> None:
if not rollouts:
_save_placeholder(path, "RFT Rollout Timeline", "No RFT rollouts found.")
return
plt = _ensure_matplotlib()
xs = list(range(1, len(rollouts) + 1))
scores = [_as_float(row.get("score")) for row in rollouts]
kept_x = [xs[i] for i, row in enumerate(rollouts) if row.get("kept")]
kept_y = [scores[i] for i, row in enumerate(rollouts) if row.get("kept")]
drop_x = [xs[i] for i, row in enumerate(rollouts) if not row.get("kept")]
drop_y = [scores[i] for i, row in enumerate(rollouts) if not row.get("kept")]
rolling_keep = []
for index in range(len(rollouts)):
start = max(0, index - 9)
window = rollouts[start : index + 1]
rolling_keep.append(sum(1 for row in window if row.get("kept")) / len(window))
fig, ax = plt.subplots(figsize=(12, 5.8))
ax.plot(xs, scores, color="#6b7280", linewidth=1.1, alpha=0.65, label="score")
ax.scatter(kept_x, kept_y, color="#2ca25f", s=45, label="kept")
ax.scatter(drop_x, drop_y, color="#d95f02", marker="x", s=42, label="rejected")
ax2 = ax.twinx()
ax2.plot(xs, rolling_keep, color="#2563eb", linewidth=2, label="rolling keep rate")
ax.set_title("RFT Rollout Timeline")
ax.set_xlabel("generated rollout")
ax.set_ylabel("filter score")
ax2.set_ylabel("rolling keep rate")
ax.grid(True, axis="y", alpha=0.25)
lines, labels = ax.get_legend_handles_labels()
lines2, labels2 = ax2.get_legend_handles_labels()
ax.legend(lines + lines2, labels + labels2, loc="best")
fig.tight_layout()
fig.savefig(path, dpi=170)
plt.close(fig)
def _save_eval_overview(path: Path, eval_report: Dict[str, Any]) -> None:
if not eval_report:
_save_placeholder(path, "Held-Out Eval After RFT", "No eval report provided yet.")
return
plt = _ensure_matplotlib()
metrics = [
("Mean score", "mean_score"),
("Detection", "detection_rate"),
("Risk reduction", "risk_reduction_rate"),
("Worker rehab", "worker_rehabilitation_rate"),
("False positive", "false_positive_rate"),
]
baseline = _get(eval_report, "overall.baseline", {})
candidate = _get(eval_report, "overall.candidate", {})
labels = [label for label, _ in metrics]
base_values = [_as_float(baseline.get(key)) for _, key in metrics]
cand_values = [_as_float(candidate.get(key)) for _, key in metrics]
xs = list(range(len(labels)))
width = 0.38
fig, ax = plt.subplots(figsize=(12, 5.8))
ax.bar([x - width / 2 for x in xs], base_values, width=width, color="#d95f02", label=str(eval_report.get("baseline_label") or "baseline"))
ax.bar([x + width / 2 for x in xs], cand_values, width=width, color="#2ca25f", label=str(eval_report.get("candidate_label") or "candidate"))
ax.set_title("Held-Out Evaluation: Baseline vs RFT Candidate")
ax.set_ylabel("rate / score")
ax.set_xticks(xs)
ax.set_xticklabels(labels, rotation=20, ha="right")
ax.grid(True, axis="y", alpha=0.25)
ax.legend()
fig.tight_layout()
fig.savefig(path, dpi=170)
plt.close(fig)
def _save_eval_task_delta(path: Path, eval_report: Dict[str, Any]) -> None:
per_task = _get(eval_report, "per_task", {})
if not isinstance(per_task, dict) or not per_task:
_save_placeholder(path, "RFT Held-Out Score Delta By Task", "No per-task eval rows found.")
return
labels = []
deltas = []
for task_id, payload in sorted(per_task.items()):
baseline_score = _as_float(_get(payload, "baseline.mean_score"))
candidate_score = _as_float(_get(payload, "candidate.mean_score"))
labels.append(str(task_id))
deltas.append(candidate_score - baseline_score)
plt = _ensure_matplotlib()
colors = ["#2ca25f" if value >= 0 else "#d95f02" for value in deltas]
fig, ax = plt.subplots(figsize=(12, 5.8))
ax.bar(labels, deltas, color=colors)
ax.axhline(0.0, color="#111111", linewidth=1)
ax.set_title("Held-Out Score Delta By Task")
ax.set_ylabel("candidate mean score - baseline mean score")
ax.tick_params(axis="x", rotation=25)
ax.grid(True, axis="y", alpha=0.25)
fig.tight_layout()
fig.savefig(path, dpi=170)
plt.close(fig)
def _write_markdown(
path: Path,
label: str,
rollouts: List[Dict[str, Any]],
kept: List[Dict[str, Any]],
summary: Dict[str, Any],
eval_report: Dict[str, Any],
images: List[str],
) -> None:
total = len(rollouts)
kept_count = len(kept)
keep_rate = kept_count / total if total else 0.0
mean_score_total = mean([_as_float(row.get("score")) for row in rollouts]) if rollouts else 0.0
mean_score_kept = mean([_as_float(row.get("score")) for row in kept]) if kept else 0.0
mean_fp_kept = mean([_as_float(row.get("fp")) for row in kept]) if kept else 0.0
eval_overall = _get(eval_report, "overall", {})
if eval_overall:
intro = (
"This folder is the rejection-sampling fine-tuning proof layer. "
"It shows which model-generated rollouts were accepted, which were rejected, "
"and what the held-out evaluation says after the polish pass."
)
else:
intro = (
"This folder is the rejection-sampling fine-tuning proof layer. "
"It shows which model-generated rollouts were accepted, which were rejected, "
"and which low-false-positive samples were used for the polish pass. "
"Held-out model evaluation was intentionally omitted for this proof pack."
)
lines = [
f"# {label} RFT Proof Pack",
"",
intro,
"",
"## Summary",
"",
f"- Total generated rollouts: `{total}`",
f"- Kept rollouts used for SFT: `{kept_count}`",
f"- Keep rate: `{keep_rate:.1%}`",
f"- Mean rollout score: `{mean_score_total:.3f}`",
f"- Mean kept score: `{mean_score_kept:.3f}`",
f"- Mean kept false positives: `{mean_fp_kept:.2f}`",
]
if summary:
lines.extend([
f"- RFT status: `{_get(summary, 'sft.status', summary.get('status', 'unknown'))}`",
f"- Output adapter: `{_get(summary, 'output.final_dir', summary.get('final_dir', 'see RFT output dir'))}`",
])
if eval_overall:
lines.extend([
"",
"## Held-Out Eval",
"",
f"- Baseline mean score: `{_as_float(eval_overall.get('baseline_mean_score')):.3f}`",
f"- Candidate mean score: `{_as_float(eval_overall.get('candidate_mean_score')):.3f}`",
f"- Mean score delta: `{_as_float(eval_overall.get('mean_score_delta')):.3f}`",
f"- Candidate risk reduction: `{_as_float(eval_overall.get('candidate_risk_reduction_rate')):.1%}`",
f"- Candidate false-positive rate: `{_as_float(eval_overall.get('candidate_false_positive_rate')):.1%}`",
])
lines.extend(["", "## Plots", ""])
for image in images:
title = Path(image).stem.replace("_", " ").title()
lines.extend([f"### {title}", "", f"", ""])
path.write_text("\n".join(lines) + "\n", encoding="utf-8")
def render_rft_proof(
rft_dir: Path,
output_dir: Path,
eval_report_path: Optional[Path],
label: str,
min_score: Optional[float],
max_fp: Optional[float],
) -> Dict[str, Any]:
output_dir.mkdir(parents=True, exist_ok=True)
rollouts = _load_jsonl(rft_dir / "rollouts.jsonl")
kept = [row for row in rollouts if row.get("kept")]
summary = _load_json(rft_dir / "rft_summary.json")
eval_report = _load_json(eval_report_path) if eval_report_path else {}
if min_score is None:
min_score = _as_float(_get(summary, "config.MIN_SCORE"), default=float("nan"))
if min_score != min_score:
min_score = None
if max_fp is None:
max_fp = _as_float(_get(summary, "config.MAX_FP"), default=float("nan"))
if max_fp != max_fp:
max_fp = None
image_names = [
"01_rft_keep_drop_by_task.png",
"02_rft_score_distribution.png",
"03_rft_false_positive_distribution.png",
"04_rft_score_vs_fp_filter.png",
"05_rft_rollout_timeline.png",
"06_rft_eval_overview.png",
"07_rft_eval_task_delta.png",
]
_save_keep_drop(output_dir / image_names[0], rollouts)
_save_score_by_task(output_dir / image_names[1], rollouts, min_score)
_save_fp_by_task(output_dir / image_names[2], rollouts, max_fp)
_save_score_vs_fp(output_dir / image_names[3], rollouts, min_score, max_fp)
_save_timeline(output_dir / image_names[4], rollouts)
_save_eval_overview(output_dir / image_names[5], eval_report)
_save_eval_task_delta(output_dir / image_names[6], eval_report)
manifest = {
"label": label,
"rft_dir": str(rft_dir),
"eval_report_path": str(eval_report_path) if eval_report_path else "",
"total_rollouts": len(rollouts),
"kept_rollouts": len(kept),
"keep_rate": len(kept) / len(rollouts) if rollouts else 0.0,
"images": image_names,
}
(output_dir / "rft_plot_manifest.json").write_text(json.dumps(manifest, indent=2), encoding="utf-8")
_write_markdown(output_dir / "rft_proof.md", label, rollouts, kept, summary, eval_report, image_names)
return manifest
def main() -> None:
parser = argparse.ArgumentParser(description="Render proof plots for a SENTINEL RFT polish run.")
parser.add_argument("--rft-dir", default="/data/sentinel_outputs_rft_phase1_100", help="Directory containing rollouts.jsonl and rft_summary.json.")
parser.add_argument("--eval-report", default="/data/rft_eval/sentinel_held_out_report.json", help="Optional held-out eval JSON report.")
parser.add_argument("--output-dir", default="outputs/rft_phase1_100/plots", help="Where to write PNG plots and markdown.")
parser.add_argument("--label", default="Phase 1 + RFT", help="Label used in the markdown report.")
parser.add_argument("--min-score", type=float, default=None, help="Override score threshold line.")
parser.add_argument("--max-fp", type=float, default=None, help="Override false-positive threshold line.")
args = parser.parse_args()
eval_report = Path(args.eval_report) if args.eval_report else None
manifest = render_rft_proof(
rft_dir=Path(args.rft_dir),
output_dir=Path(args.output_dir),
eval_report_path=eval_report,
label=args.label,
min_score=args.min_score,
max_fp=args.max_fp,
)
print(json.dumps(manifest, indent=2))
if __name__ == "__main__":
main()
|