Spaces:
Running on Zero
Running on Zero
File size: 21,517 Bytes
0dd6c2f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 | import ast
from pathlib import Path
from typing import Any
import matplotlib.pyplot as plt
import numpy as np
from matplotlib.ticker import MaxNLocator
from linalg_zero.generator.models import DifficultyCategory, Question, Task
from linalg_zero.shared.utils import get_logger
logger = get_logger(__name__)
def print_statistics_summary(statistics: dict[str, Any]) -> None: # pragma: no cover
"""Print a formatted summary of dataset statistics."""
if not statistics:
logger.info("No statistics available.")
return
logger.info("=" * 50)
logger.info("DATASET STATISTICS SUMMARY")
logger.info("=" * 50)
# Overall statistics
overall_min = statistics.get("overall_min")
overall_max = statistics.get("overall_max")
overall_min_abs = statistics.get("overall_min_abs")
logger.info(f"Overall Range: {overall_min} to {overall_max}")
logger.info(f"Overall Min Absolute: {overall_min_abs}")
# Per-step statistics
per_step = statistics.get("per_step", {})
if per_step:
logger.info("Per-Step Statistics:")
for step_idx in sorted(per_step.keys()):
step_stats = per_step[step_idx]
logger.info(
f" Step {step_idx}: min={step_stats.get('min')}, max={step_stats.get('max')}, min_abs={step_stats.get('min_abs')}, count={step_stats.get('count')}"
)
# Per-problem-type statistics
per_problem_type = statistics.get("per_problem_type", {})
if per_problem_type:
logger.info("Per-Problem-Type Statistics:")
for problem_type, type_stats in per_problem_type.items():
logger.info(
f" {problem_type}: min={type_stats.get('min')}, max={type_stats.get('max')}, min_abs={type_stats.get('min_abs')}, count={type_stats.get('count')}"
)
# Per-question statistics
per_question = statistics.get("per_question", [])
if per_question:
logger.info(f"Per-Question Statistics: {len(per_question)} questions analyzed")
# Show first few questions as examples
for i, q_stats in enumerate(per_question[:3]):
logger.info(
f" Q{i + 1}: min={q_stats.get('min')}, max={q_stats.get('max')}, min_abs={q_stats.get('min_abs')}, count={q_stats.get('count')}"
)
if len(per_question) > 3:
logger.info(f" ... and {len(per_question) - 3} more questions")
logger.info("=" * 50)
def _extract_numeric_values_from_object(obj: Any) -> list[float]:
"""Recursively extract numeric values (as floats) from an arbitrary object."""
values: list[float] = []
if isinstance(obj, int | float):
values.append(float(obj))
return values
if isinstance(obj, complex):
raise TypeError(f"Complex number found: {obj}")
if isinstance(obj, list | tuple):
for item in obj:
values.extend(_extract_numeric_values_from_object(item))
return values
if isinstance(obj, dict):
for v in obj.values():
values.extend(_extract_numeric_values_from_object(v))
return values
return values
def compute_stepwise_value_statistics(questions: list[Question]) -> dict[str, Any]:
"""Scan stepwise results from all questions and compute statistics.
Returns a dictionary with:
- overall_min: float | None
- overall_max: float | None
- overall_min_abs: float | None
- per_question: list of {index, min, max, min_abs, count}
- per_step: dict[int, {min, max, min_abs, count}] aggregated across all questions by step index
- per_problem_type: dict[str, {min, max, min_abs, count}] aggregated by problem type
- all_values: flat list[float] of every numeric value encountered across all steps/questions
"""
overall_min: float | None = None
overall_max: float | None = None
overall_min_abs: float | None = None
per_question: list[dict[str, Any]] = []
per_step: dict[int, dict[str, Any]] = {}
per_problem_type: dict[str, dict[str, Any]] = {}
all_values: list[float] = []
for q_index, question in enumerate(questions):
q_min: float | None = None
q_max: float | None = None
q_min_abs: float | None = None
q_count: int = 0
# Resolve problem type key once per question
pt_key = getattr(question.problem_type, "value", str(question.problem_type))
for step_index, step in enumerate(question.stepwise):
# Parse the step result into a Python object
result_str = step.get("result")
if result_str is None:
raise ValueError(f"Step {step_index} has no result")
parsed = ast.literal_eval(result_str)
# Extract numeric values
numeric_values = _extract_numeric_values_from_object(parsed)
if not numeric_values:
# Initialize per-step entry with zero count if not present
if step_index not in per_step:
per_step[step_index] = {"min": None, "max": None, "min_abs": None, "count": 0}
continue
step_min = min(numeric_values)
step_max = max(numeric_values)
step_min_abs = min(abs(v) for v in numeric_values)
step_count = len(numeric_values)
# Aggregate raw values
all_values.extend(float(v) for v in numeric_values)
# Update overall stats
overall_min = step_min if overall_min is None else min(overall_min, step_min)
overall_max = step_max if overall_max is None else max(overall_max, step_max)
overall_min_abs = step_min_abs if overall_min_abs is None else min(overall_min_abs, step_min_abs)
# Update question stats
q_min = step_min if q_min is None else min(q_min, step_min)
q_max = step_max if q_max is None else max(q_max, step_max)
q_min_abs = step_min_abs if q_min_abs is None else min(q_min_abs, step_min_abs)
q_count += step_count
# Update per-step aggregated stats
if step_index not in per_step:
per_step[step_index] = {"min": step_min, "max": step_max, "min_abs": step_min_abs, "count": step_count}
else:
ps = per_step[step_index]
ps_min = ps["min"]
ps_max = ps["max"]
ps_min_abs = ps["min_abs"]
ps["min"] = step_min if ps_min is None else min(ps_min, step_min)
ps["max"] = step_max if ps_max is None else max(ps_max, step_max)
ps["min_abs"] = step_min_abs if ps_min_abs is None else min(ps_min_abs, step_min_abs)
ps["count"] += step_count
# Update per-problem-type aggregated stats
if pt_key not in per_problem_type:
per_problem_type[pt_key] = {
"min": step_min,
"max": step_max,
"min_abs": step_min_abs,
"count": step_count,
}
else:
ppt = per_problem_type[pt_key]
ppt_min = ppt["min"]
ppt_max = ppt["max"]
ppt_min_abs = ppt["min_abs"]
ppt["min"] = step_min if ppt_min is None else min(ppt_min, step_min)
ppt["max"] = step_max if ppt_max is None else max(ppt_max, step_max)
ppt["min_abs"] = step_min_abs if ppt_min_abs is None else min(ppt_min_abs, step_min_abs)
ppt["count"] += step_count
per_question.append({"index": q_index, "min": q_min, "max": q_max, "min_abs": q_min_abs, "count": q_count})
return {
"overall_min": overall_min,
"overall_max": overall_max,
"overall_min_abs": overall_min_abs if overall_min_abs is not None else 0.0,
"per_question": per_question,
"per_step": per_step,
"per_problem_type": per_problem_type,
"all_values": all_values,
}
def extract_all_numerical_values(statistics: dict[tuple, dict[str, Any]], use_min_max: bool = False) -> list[float]:
"""Extract all raw stepwise numerical values from all runs.
Uses the aggregated `all_values` emitted by `compute_stepwise_value_statistics` for
each combination rather than only per-question min/max boundaries.
"""
all_values: list[float] = []
for _, stats in statistics.items():
if use_min_max:
values = []
per_question = stats.get("per_question", [])
for q_stats in per_question:
values.append(q_stats["min"])
values.append(q_stats["max"])
else:
values = stats.get("all_values", [])
all_values.extend(values)
for value in all_values:
assert value is not None, "Value is None"
return all_values
def extract_values_by_combination(
statistics: dict[tuple, dict[str, Any]], use_min_max: bool = False
) -> dict[tuple, list[float]]:
"""Extract raw stepwise numerical values grouped by entropy combination."""
values_by_combination: dict[tuple, list[float]] = {}
for combination, stats in statistics.items():
if use_min_max:
values = []
per_question = stats.get("per_question", [])
for q_stats in per_question:
values.append(q_stats["min"])
values.append(q_stats["max"])
else:
values = list(stats.get("all_values", []))
values_by_combination[combination] = values
for key, items in values_by_combination.items():
assert all(value is not None for value in items), f"Values are None for combination {key}"
return values_by_combination
def plot_overall_histogram(
all_values: list[float], target_min: float = -1000, target_max: float = 1000, output_dir: Path | None = None
) -> None:
"""Plot histogram of all numerical values across all runs."""
plt.figure(figsize=(12, 6))
plt.subplot(1, 2, 1)
plt.hist(all_values, bins=50, alpha=0.7, color="skyblue", edgecolor="black")
plt.axvline(target_min, color="red", linestyle="--", label=f"Target Min ({target_min})")
plt.axvline(target_max, color="red", linestyle="--", label=f"Target Max ({target_max})")
plt.xlabel("Numerical Values")
plt.ylabel("Frequency")
plt.title("Distribution of All Numerical Values")
plt.legend()
plt.grid(True, alpha=0.3)
# Log scale version
plt.subplot(1, 2, 2)
# Filter out zero and negative values for log scale
positive_values = [v for v in all_values if v > 0]
if positive_values:
plt.hist(positive_values, bins=50, alpha=0.7, color="lightcoral", edgecolor="black")
plt.axvline(target_max, color="red", linestyle="--", label=f"Target Max ({target_max})")
plt.xlabel("Numerical Values (log scale)")
plt.ylabel("Frequency")
plt.title("Distribution of Positive Values (Log Scale)")
plt.xscale("log")
plt.legend()
plt.grid(True, alpha=0.3)
plt.tight_layout()
filename = "overall_distribution.png"
filepath = output_dir / filename if output_dir else filename
plt.savefig(filepath, dpi=300, bbox_inches="tight")
plt.show()
def plot_combination_comparison(
values_by_combination: dict[tuple, list[float]], max_combinations: int = 12, output_dir: Path | None = None
) -> None:
"""Plot comparison of value distributions across different entropy combinations."""
# Limit to top combinations by number of values
sorted_combinations = sorted(values_by_combination.items(), key=lambda x: len(x[1]), reverse=True)[
:max_combinations
]
_, axes = plt.subplots(3, 4, figsize=(16, 12))
axes = axes.flatten()
for i, (combination, values) in enumerate(sorted_combinations):
if i >= len(axes):
break
ax = axes[i]
if values: # Only plot if we have values
ax.hist(values, bins=20, alpha=0.7, color=plt.colormaps["tab10"](i % 10), edgecolor="black")
ax.yaxis.set_major_locator(MaxNLocator(integer=True))
ax.set_title(f"Combo: {combination}", fontsize=10)
ax.set_xlabel("Values")
ax.set_ylabel("Freq")
ax.grid(True, alpha=0.3)
# Add statistics text
if values:
mean_val = np.mean(values)
std_val = np.std(values)
ax.text(
0.05,
0.95,
f"μ={mean_val:.1f}\no={std_val:.1f}\nn={len(values)}",
transform=ax.transAxes,
verticalalignment="top",
bbox={"boxstyle": "round", "facecolor": "white", "alpha": 0.8},
)
# Hide unused subplots
for i in range(len(sorted_combinations), len(axes)):
axes[i].set_visible(False)
plt.suptitle("Value Distributions by Entropy Combination", fontsize=14)
plt.tight_layout()
filename = "by_combination.png"
filepath = output_dir / filename if output_dir else filename
plt.savefig(filepath, dpi=300, bbox_inches="tight")
plt.show()
def plot_target_compliance(
statistics: dict[tuple, dict[str, Any]],
target_min: float = -1000,
target_max: float = 1000,
output_dir: Path | None = None,
) -> None:
"""Plot how well each combination complies with target ranges."""
compliance_data = []
for combination, stats in statistics.items():
overall_min = stats.get("overall_min")
overall_max = stats.get("overall_max")
if overall_min is None or overall_max is None:
raise ValueError(f"Overall min or max is None for combination {combination}")
# overall_min/overall_max are guaranteed non-None above; avoid falsy-zero filtering
within_range = (target_min <= overall_min) and (overall_max <= target_max)
compliance_data.append({
"combination": str(combination),
"min": overall_min,
"max": overall_max,
"compliant": within_range,
})
if not compliance_data:
logger.warning("No compliance data available for plotting")
return
compliant = [d for d in compliance_data if d["compliant"]]
non_compliant = [d for d in compliance_data if not d["compliant"]]
plt.figure(figsize=(12, 8))
# Plot compliant combinations
if compliant:
plt.scatter(
[d["min"] for d in compliant],
[d["max"] for d in compliant],
c="green",
alpha=0.7,
s=100,
label=f"Compliant ({len(compliant)})",
)
# Plot non-compliant combinations
if non_compliant:
plt.scatter(
[d["min"] for d in non_compliant],
[d["max"] for d in non_compliant],
c="red",
alpha=0.7,
s=100,
label=f"Non-compliant ({len(non_compliant)})",
)
# Add target range box
plt.axhline(target_max, color="blue", linestyle="--", alpha=0.5, label=f"Target Max ({target_max})")
plt.axvline(target_min, color="blue", linestyle="--", alpha=0.5, label=f"Target Min ({target_min})")
plt.xlabel("Overall Minimum Value")
plt.ylabel("Overall Maximum Value")
plt.title("Entropy Combinations: Target Range Compliance")
plt.legend()
plt.grid(True, alpha=0.3)
plt.tight_layout()
filename = "compliance.png"
filepath = output_dir / filename if output_dir else filename
plt.savefig(filepath, dpi=300, bbox_inches="tight")
plt.show()
# Print summary
total_combinations = len(compliance_data)
compliant_count = len(compliant)
logger.info("Target Compliance Summary:")
logger.info(f" Total combinations: {total_combinations}")
logger.info(f" Compliant: {compliant_count} ({compliant_count / total_combinations * 100:.1f}%)")
logger.info(
f" Non-compliant: {total_combinations - compliant_count} ({(total_combinations - compliant_count) / total_combinations * 100:.1f}%)"
)
def extract_report_metadata(
top_choice: dict[str, Any],
problem_type: Task,
entropy_config: tuple[float, float] | dict[Task, tuple[float, float]],
min_value_abs: float,
entropy_jitter: float,
*,
step_size: float,
samples_per_test: int,
target_min_value: float,
target_max_value: float,
) -> dict[str, Any]:
if isinstance(entropy_config, tuple):
# Single-step problem
is_single_step = True
components = [problem_type.name]
difficulty_category = DifficultyCategory.ONE_TOOL_CALL.name
else:
# Multi-step problem - get components from the dict keys
is_single_step = False
components = [task.name for task in entropy_config if isinstance(task, Task)]
# Determine difficulty category based on number of components
num_components = len(components)
if num_components == 2:
difficulty_category = DifficultyCategory.TWO_TOOL_CALLS.name
elif num_components == 3:
difficulty_category = DifficultyCategory.THREE_TOOL_CALLS.name
else:
raise ValueError(f"Unexpected number of components for {problem_type}: {num_components}")
# Validate the ordered combination length matches the number of components
selected_combination = (
list(top_choice["combination"])
if isinstance(top_choice["combination"], list | tuple)
else [top_choice["combination"]]
)
if len(selected_combination) != len(components):
raise ValueError(
f"Mismatch between combination length ({len(selected_combination)}) and components ({len(components)}) for {problem_type.name}"
)
# Optimized within the searched grid if the chosen entropy is strictly below
# the configured upper bound (i.e., the search did not hit the boundary).
optimized = False
if isinstance(entropy_config, tuple):
if len(selected_combination) > 0 and selected_combination[0] < (entropy_config[1] - entropy_jitter):
optimized = True
else:
# Multi-step: mark optimized if ANY component's chosen entropy is strictly
# below its configured upper bound (didn't hit boundary for that component).
component_index = 0
for task in entropy_config:
if isinstance(task, Task) and component_index < len(selected_combination):
if selected_combination[component_index] < (entropy_config[task][1] - entropy_jitter):
optimized = True
break
component_index += 1
return {
"combination": selected_combination,
"score": top_choice["score"],
"overall_min": top_choice["overall_min"],
"overall_max": top_choice["overall_max"],
"min_abs": top_choice["min_abs"],
"count": top_choice["count"],
"optimized": optimized,
"metadata": {
"is_single_step": is_single_step,
"components": components,
"difficulty_category": difficulty_category,
"task_enum": problem_type.name,
"entropy_jitter": entropy_jitter,
"min_element_abs": min_value_abs,
"step_size": step_size,
"samples_per_test": samples_per_test,
"target_min_value": target_min_value,
"target_max_value": target_max_value,
},
}
def rank_entropy_combinations(
statistics: dict[tuple, dict[str, Any]],
*,
target_min: float,
target_max: float,
weights: dict[str, float] | None = None,
) -> list[dict[str, Any]]:
"""Rank entropy combinations by distribution quality.
Returns a list of dicts sorted by descending score, each with keys:
- combination: tuple of entropy values
- score: float in [0, 1]
- metrics: dict as computed by _compute_distribution_metrics
"""
if weights is None:
weights = {"compliance": 0.4, "center": 0.2, "coverage": 0.2, "balance": 0.1, "zero": 0.1}
ranked: list[dict[str, Any]] = []
for combination, stats in statistics.items():
# Hard gate 1: overall range must be fully within targets if required
overall_min = stats["overall_min"]
overall_max = stats["overall_max"]
if overall_min is None or overall_max is None:
continue
if not (target_min <= float(overall_min) <= float(target_max)):
continue
if not (target_min <= float(overall_max) <= float(target_max)):
continue
# Calculate distance from targets (lower is better)
min_distance = abs(float(overall_min) - target_min)
max_distance = abs(float(overall_max) - target_max)
total_distance = min_distance + max_distance
# Get min_abs and total count from statistics
overall_min_abs = stats["overall_min_abs"]
# Calculate total count across all questions
total_count = sum(q_stats["count"] for q_stats in stats["per_question"])
ranked.append({
"combination": combination,
"score": total_distance,
"min_distance": min_distance,
"max_distance": max_distance,
"overall_min": overall_min,
"overall_max": overall_max,
"min_abs": overall_min_abs,
"count": total_count,
})
# Sort by total distance (ascending - closest first)
ranked.sort(key=lambda d: d["score"])
return ranked
|