Spaces:
Sleeping
Sleeping
File size: 22,646 Bytes
8125804 7ed7a6d 8125804 78c132b 8125804 7ed7a6d 8125804 78c132b 8125804 78c132b 8125804 78c132b 8125804 78c132b 8125804 78c132b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 | from __future__ import annotations
from dataclasses import dataclass
import math
import re
from collections.abc import Mapping
from typing import Any
import torch
import torch.nn.functional as F
_EPS = 1e-8
_LEADING_WHITESPACE_MARKERS = ("Δ ", "β", "Δ", "Δ")
@dataclass(frozen=True)
class AnchorSpanMatch:
anchor_text: str
token_start: int
token_end: int
token_count: int
char_start: int | None
char_end: int | None
match_method: str
matched_text: str
def select_representative_layers(
num_hidden_layers: int,
count: int = 4,
) -> list[int]:
if num_hidden_layers <= 0:
raise ValueError("num_hidden_layers must be positive")
if count <= 0:
raise ValueError("count must be positive")
fractions = (0.25, 0.50, 0.75, 0.96)
raw_layers = [
max(1, min(num_hidden_layers, int(round(num_hidden_layers * fraction))))
for fraction in fractions[:count]
]
layers = sorted(set(raw_layers))
if len(layers) >= count:
return layers[:count]
fallback = torch.linspace(1, num_hidden_layers, steps=min(num_hidden_layers, count)).round().int().tolist()
for layer in fallback:
layer_id = int(layer)
if layer_id not in layers:
layers.append(layer_id)
return sorted(set(layers))
def list_model_layers(num_hidden_layers: int) -> list[int]:
if num_hidden_layers <= 0:
raise ValueError("num_hidden_layers must be positive")
return list(range(int(num_hidden_layers)))
def select_tail_probe_layers(
num_hidden_layers: int,
count: int = 10,
) -> list[int]:
if num_hidden_layers <= 0:
raise ValueError("num_hidden_layers must be positive")
if count <= 0:
raise ValueError("count must be positive")
width = min(int(count), int(num_hidden_layers))
start = int(num_hidden_layers) - width
return list(range(start, int(num_hidden_layers)))
def build_tail_reference_layers(
probe_layers: list[int],
) -> dict[str, int]:
if not probe_layers:
raise ValueError("probe_layers must not be empty")
layers = sorted(int(layer) for layer in probe_layers)
mature_index = max(0, len(layers) - 4)
template_prev_index = max(0, len(layers) - 2)
template_curr_index = len(layers) - 1
mature_layer = layers[mature_index]
return {
"slope_start_layer": layers[0],
"slope_end_layer": mature_layer,
"mature_layer": mature_layer,
"template_prev_layer": layers[template_prev_index],
"template_curr_layer": layers[template_curr_index],
}
def _normalize_text(text: str) -> str:
return " ".join(text.lower().split())
def _find_unique_substring(text: str, anchor_text: str) -> tuple[int, int] | None:
matches = list(re.finditer(re.escape(anchor_text), text))
if not matches:
return None
match = matches[0]
return match.start(), match.end()
def _decode_anchor_tokens(
tokenizer: Any,
token_ids: list[int],
) -> str:
try:
return str(tokenizer.decode(token_ids, skip_special_tokens=False))
except Exception:
return ""
def decode_token_surfaces(
tokenizer: Any,
token_ids: list[int],
) -> list[str]:
convert = getattr(tokenizer, "convert_ids_to_tokens", None)
if callable(convert):
try:
tokens = convert(token_ids)
if isinstance(tokens, list) and len(tokens) == len(token_ids):
return [str(token) for token in tokens]
except Exception:
pass
return [_decode_anchor_tokens(tokenizer, [int(token_id)]) for token_id in token_ids]
def decode_token_pieces(
tokenizer: Any,
token_ids: list[int],
) -> list[str]:
return [_decode_anchor_tokens(tokenizer, [int(token_id)]) for token_id in token_ids]
def token_has_leading_whitespace(
raw_surface: str,
decoded_piece: str,
) -> bool:
if decoded_piece.startswith(" "):
return True
return raw_surface.startswith(_LEADING_WHITESPACE_MARKERS)
def _search_subsequence(
full_ids: list[int],
sub_ids: list[int],
) -> list[tuple[int, int]]:
if not sub_ids or len(sub_ids) > len(full_ids):
return []
matches: list[tuple[int, int]] = []
width = len(sub_ids)
for start in range(len(full_ids) - width + 1):
if full_ids[start : start + width] == sub_ids:
matches.append((start, start + width - 1))
return matches
def _match_from_offsets(
*,
text: str,
anchor_text: str,
offsets: list[tuple[int, int]],
) -> AnchorSpanMatch | None:
char_span = _find_unique_substring(text, anchor_text)
if char_span is None:
return None
char_start, char_end = char_span
active_tokens = [
idx
for idx, (offset_start, offset_end) in enumerate(offsets)
if offset_end > offset_start and offset_start < char_end and offset_end > char_start
]
if not active_tokens:
return None
token_start = active_tokens[0]
token_end = active_tokens[-1]
matched_text = text[offsets[token_start][0] : offsets[token_end][1]]
if _normalize_text(anchor_text) not in _normalize_text(matched_text):
return None
return AnchorSpanMatch(
anchor_text=anchor_text,
token_start=token_start,
token_end=token_end,
token_count=token_end - token_start + 1,
char_start=char_start,
char_end=char_end,
match_method="offset_mapping",
matched_text=matched_text,
)
def _match_from_token_ids(
*,
anchor_text: str,
input_ids: list[int],
tokenizer: Any,
) -> AnchorSpanMatch | None:
candidate_matches: dict[tuple[int, int], str] = {}
for variant, label in ((anchor_text, "token_ids_exact"), (f" {anchor_text}", "token_ids_prefixed_space")):
try:
encoded = tokenizer(variant, add_special_tokens=False)
except Exception:
continue
phrase_ids: Any
if isinstance(encoded, Mapping):
phrase_ids = encoded.get("input_ids")
elif hasattr(encoded, "input_ids"):
phrase_ids = getattr(encoded, "input_ids")
elif hasattr(encoded, "__getitem__"):
try:
phrase_ids = encoded["input_ids"]
except Exception:
phrase_ids = encoded
else:
phrase_ids = encoded
if phrase_ids is None:
continue
if isinstance(phrase_ids, torch.Tensor):
phrase_seq = [int(token) for token in phrase_ids.reshape(-1).tolist()]
else:
phrase_seq = [int(token) for token in phrase_ids]
for match in _search_subsequence(input_ids, phrase_seq):
candidate_matches[match] = label
if not candidate_matches:
return None
(token_start, token_end), label = next(iter(candidate_matches.items()))
matched_text = _decode_anchor_tokens(tokenizer, input_ids[token_start : token_end + 1])
return AnchorSpanMatch(
anchor_text=anchor_text,
token_start=token_start,
token_end=token_end,
token_count=token_end - token_start + 1,
char_start=None,
char_end=None,
match_method=label,
matched_text=matched_text,
)
def _match_from_decoded_pieces(
*,
anchor_text: str,
input_ids: list[int],
tokenizer: Any,
) -> AnchorSpanMatch | None:
"""Fallback: decode each token individually, build charβtoken map, search."""
pieces = decode_token_pieces(tokenizer, input_ids)
if not pieces:
return None
# Build cumulative decoded string with token boundary tracking
char_to_token: list[int] = []
decoded_full = ""
for tok_idx, piece in enumerate(pieces):
for _ in piece:
char_to_token.append(tok_idx)
decoded_full += piece
if not decoded_full:
return None
# Search for anchor_text (case-insensitive to handle tokenizer quirks)
needle = anchor_text.lower()
haystack = decoded_full.lower()
matches: list[int] = []
start_pos = 0
while True:
pos = haystack.find(needle, start_pos)
if pos == -1:
break
matches.append(pos)
start_pos = pos + 1
if not matches:
return None
char_start = matches[0]
char_end = char_start + len(needle)
# Map character positions to token indices
if char_start >= len(char_to_token) or char_end - 1 >= len(char_to_token):
return None
token_start = char_to_token[char_start]
token_end = char_to_token[min(char_end - 1, len(char_to_token) - 1)]
matched_text = decoded_full[char_start:char_end]
return AnchorSpanMatch(
anchor_text=anchor_text,
token_start=token_start,
token_end=token_end,
token_count=token_end - token_start + 1,
char_start=None,
char_end=None,
match_method="decoded_pieces",
matched_text=matched_text,
)
def match_anchor_span(
*,
text: str,
anchor_text: str,
input_ids: list[int],
tokenizer: Any,
offsets: list[tuple[int, int]] | None = None,
) -> AnchorSpanMatch | None:
if offsets is not None:
match = _match_from_offsets(
text=text,
anchor_text=anchor_text,
offsets=offsets,
)
if match is not None:
return match
match = _match_from_token_ids(
anchor_text=anchor_text,
input_ids=input_ids,
tokenizer=tokenizer,
)
if match is not None:
return match
return _match_from_decoded_pieces(
anchor_text=anchor_text,
input_ids=input_ids,
tokenizer=tokenizer,
)
def detect_anchor_span(
attentions: tuple[torch.Tensor, ...],
probe_layers: list[int] | None = None,
*,
min_width: int = 2,
max_width: int = 8,
skip_special: int = 1,
use_last_n: int = 4,
) -> AnchorSpanMatch | None:
"""Detect anchor span via attention mass from the last token.
For each available mature attention layer, sum attention the last
token pays to each preceding position (averaged across heads).
The contiguous span with highest cumulative score is the anchor.
Args:
attentions: tuple of [batch, heads, seq, seq] from model output.
May be a subset of all layers (e.g. 8 out of 32).
probe_layers: layer indices to try. If the tuple is shorter than
expected, falls back to the last ``use_last_n`` available entries.
min_width: minimum span width in tokens.
max_width: maximum span width in tokens.
skip_special: skip first N tokens (BOS / special).
use_last_n: how many of the *last* available attention entries to
aggregate when probe_layers don't map into the tuple.
Returns:
AnchorSpanMatch with detected span, or None if input is too short.
"""
if not attentions:
return None
seq_len = attentions[0].size(-1)
if seq_len < skip_special + min_width + 1:
return None
n_attn = len(attentions)
# Build list of actual attention indices to use
attn_indices: list[int] = []
if probe_layers:
for layer_idx in probe_layers:
attn_idx = layer_idx + 1
if 0 <= attn_idx < n_attn:
attn_indices.append(attn_idx)
# Fallback: use last N available attention entries
if not attn_indices:
start = max(0, n_attn - use_last_n)
attn_indices = list(range(start, n_attn))
score = torch.zeros(seq_len)
n_layers_used = 0
for attn_idx in attn_indices:
attn = attentions[attn_idx][0] # [heads, seq, seq]
# Attention from last token to all previous
last_attn = attn[:, -1, :].mean(dim=0).detach().cpu() # [seq]
score += last_attn
n_layers_used += 1
if n_layers_used == 0:
return None
score /= n_layers_used
# Zero out special tokens and the last token itself
score[:skip_special] = 0.0
score[-1] = 0.0
# Sliding window: find the contiguous span with highest total attention
best_score = -1.0
best_start = skip_special
best_end = skip_special + min_width - 1
upper_width = min(max_width, seq_len - skip_special - 1)
for width in range(min_width, upper_width + 1):
if width > seq_len:
break
windows = score[skip_special : seq_len - 1].unfold(0, width, 1)
if windows.numel() == 0:
continue
sums = windows.sum(dim=-1)
idx = int(sums.argmax().item())
val = float(sums[idx].item())
if val > best_score:
best_score = val
best_start = skip_special + idx
best_end = best_start + width - 1
return AnchorSpanMatch(
anchor_text="",
token_start=best_start,
token_end=best_end,
token_count=best_end - best_start + 1,
char_start=None,
char_end=None,
match_method="attention_mass",
matched_text="",
)
def extract_delta_vectors(
hidden_states: torch.Tensor,
token_start: int,
token_end: int,
) -> torch.Tensor:
if hidden_states.ndim != 2:
raise ValueError("hidden_states must be shaped [seq_len, hidden_dim]")
if token_start < 0 or token_end < token_start or token_end >= hidden_states.size(0):
raise ValueError("invalid token span")
span_hidden = hidden_states[token_start : token_end + 1].to(dtype=torch.float32)
if span_hidden.size(0) < 2:
return span_hidden.new_zeros((0, span_hidden.size(-1)))
return span_hidden[1:] - span_hidden[:-1]
def compute_geometry_metrics(
delta_vectors: torch.Tensor,
) -> dict[str, float | int | None]:
if delta_vectors.ndim != 2:
raise ValueError("delta_vectors must be shaped [delta_count, hidden_dim]")
delta_count = int(delta_vectors.size(0))
token_count = delta_count + 1 if delta_count > 0 else 0
if delta_count == 0:
return {
"token_count": token_count,
"delta_count": delta_count,
"mean_direction_norm": None,
"mean_step_norm": None,
"adjacent_cosine_coherence": None,
"path_tortuosity": None,
"rank1_explained_variance": None,
"curvature_proxy": None,
}
deltas = delta_vectors.to(dtype=torch.float32)
step_norms = deltas.norm(dim=-1)
mean_direction = deltas.mean(dim=0)
metrics: dict[str, float | int | None] = {
"token_count": token_count,
"delta_count": delta_count,
"mean_direction_norm": float(mean_direction.norm().item()),
"mean_step_norm": float(step_norms.mean().item()),
"adjacent_cosine_coherence": None,
"path_tortuosity": None,
"rank1_explained_variance": None,
"curvature_proxy": None,
}
if token_count < 4:
return metrics
adjacent: list[float] = []
for idx in range(delta_count - 1):
left = deltas[idx]
right = deltas[idx + 1]
left_norm = float(left.norm().item())
right_norm = float(right.norm().item())
if left_norm <= _EPS or right_norm <= _EPS:
continue
adjacent.append(float(F.cosine_similarity(left.unsqueeze(0), right.unsqueeze(0), dim=-1).item()))
if adjacent:
coherence = float(sum(adjacent) / len(adjacent))
metrics["adjacent_cosine_coherence"] = coherence
metrics["curvature_proxy"] = float(1.0 - coherence)
displacement = deltas.sum(dim=0)
displacement_norm = float(displacement.norm().item())
if displacement_norm > _EPS:
metrics["path_tortuosity"] = float(step_norms.sum().item() / displacement_norm)
singular_values = torch.linalg.svdvals(deltas)
energy = float((singular_values.square().sum()).item())
if energy > _EPS:
metrics["rank1_explained_variance"] = float((singular_values[0].item() ** 2) / energy)
elif delta_count > 0:
metrics["rank1_explained_variance"] = 1.0
return metrics
def build_computability_flags(metrics: dict[str, float | int | None]) -> dict[str, bool]:
return {
"span_mean_direction": metrics.get("delta_count") is not None and int(metrics["delta_count"] or 0) > 0,
"mean_direction_norm": metrics.get("mean_direction_norm") is not None,
"mean_step_norm": metrics.get("mean_step_norm") is not None,
"adjacent_cosine_coherence": metrics.get("adjacent_cosine_coherence") is not None,
"path_tortuosity": metrics.get("path_tortuosity") is not None,
"rank1_explained_variance": metrics.get("rank1_explained_variance") is not None,
"curvature_proxy": metrics.get("curvature_proxy") is not None,
}
def compute_mean_direction(
delta_vectors: torch.Tensor,
) -> torch.Tensor | None:
if delta_vectors.ndim != 2:
raise ValueError("delta_vectors must be shaped [delta_count, hidden_dim]")
if delta_vectors.size(0) == 0:
return None
return delta_vectors.to(dtype=torch.float32).mean(dim=0)
def compute_cross_prompt_stability(
directions: list[torch.Tensor],
) -> dict[str, float | int | None]:
valid = []
for direction in directions:
norm = float(direction.norm().item())
if norm <= _EPS:
continue
valid.append(direction / norm)
if len(valid) < 2:
return {
"pair_count": 0,
"mean_pairwise_cosine": None,
"std_pairwise_cosine": None,
"min_pairwise_cosine": None,
"max_pairwise_cosine": None,
}
values: list[float] = []
for idx in range(len(valid)):
for jdx in range(idx + 1, len(valid)):
values.append(float(F.cosine_similarity(valid[idx].unsqueeze(0), valid[jdx].unsqueeze(0), dim=-1).item()))
mean_value = sum(values) / len(values)
variance = sum((value - mean_value) ** 2 for value in values) / len(values)
return {
"pair_count": len(values),
"mean_pairwise_cosine": float(mean_value),
"std_pairwise_cosine": float(math.sqrt(max(variance, 0.0))),
"min_pairwise_cosine": float(min(values)),
"max_pairwise_cosine": float(max(values)),
}
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Auto-calibration: k-means clustering for mature/template/flat thresholds
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def auto_calibrate_thresholds(
r1_references: list[float],
delta_templates: list[float],
n_clusters: int = 3,
max_iter: int = 50,
seed: int = 42,
) -> dict[str, Any]:
"""Compute cluster thresholds from observed r1/delta values via k-means.
Returns dict with:
- mature_r1_threshold: float
- template_delta_threshold: float
- cluster_centers: list of (r1_center, delta_center)
- cluster_labels: list of "mature"/"template"/"flat" per sample
- n_samples: int
"""
n = len(r1_references)
if n < 3 or len(delta_templates) != n:
return {
"mature_r1_threshold": 0.65,
"template_delta_threshold": 0.08,
"cluster_centers": [],
"cluster_labels": [],
"n_samples": n,
"method": "fallback_default",
}
import random
rng = random.Random(seed)
# Normalize features to [0, 1] range for balanced clustering
r1_arr = list(r1_references)
dt_arr = list(delta_templates)
r1_min, r1_max = min(r1_arr), max(r1_arr)
dt_min, dt_max = min(dt_arr), max(dt_arr)
r1_range = max(r1_max - r1_min, _EPS)
dt_range = max(dt_max - dt_min, _EPS)
points = [
((r - r1_min) / r1_range, (d - dt_min) / dt_range)
for r, d in zip(r1_arr, dt_arr)
]
# k-means initialization
indices = rng.sample(range(n), min(n_clusters, n))
centers = [points[i] for i in indices]
for _ in range(max_iter):
# Assign
assignments: list[int] = []
for p in points:
dists = [(p[0] - c[0]) ** 2 + (p[1] - c[1]) ** 2 for c in centers]
assignments.append(int(min(range(len(dists)), key=lambda x: dists[x])))
# Update
new_centers: list[tuple[float, float]] = []
for k in range(n_clusters):
members = [points[i] for i in range(n) if assignments[i] == k]
if not members:
new_centers.append(centers[k])
else:
new_centers.append((
sum(m[0] for m in members) / len(members),
sum(m[1] for m in members) / len(members),
))
if new_centers == centers:
break
centers = new_centers
# Denormalize centers back to original scale
real_centers = [
(c[0] * r1_range + r1_min, c[1] * dt_range + dt_min)
for c in centers
]
# Identify clusters: highest r1 center = mature, highest delta center = template, rest = flat
sorted_by_r1 = sorted(range(len(real_centers)), key=lambda i: real_centers[i][0], reverse=True)
mature_idx = sorted_by_r1[0]
remaining = [i for i in range(len(real_centers)) if i != mature_idx]
template_idx = max(remaining, key=lambda i: real_centers[i][1])
flat_idx = [i for i in remaining if i != template_idx][0] if len(remaining) > 1 else remaining[0]
cluster_map = {mature_idx: "mature", template_idx: "template", flat_idx: "flat"}
labels = [cluster_map.get(assignments[i], "flat") for i in range(n)]
# Compute thresholds as midpoints between cluster centers
mature_r1_center = real_centers[mature_idx][0]
flat_r1_center = real_centers[flat_idx][0]
mature_r1_threshold = (mature_r1_center + flat_r1_center) / 2.0
template_delta_center = real_centers[template_idx][1]
flat_delta_center = real_centers[flat_idx][1]
template_delta_threshold = (template_delta_center + flat_delta_center) / 2.0
return {
"mature_r1_threshold": float(mature_r1_threshold),
"template_delta_threshold": float(template_delta_threshold),
"cluster_centers": [
{"label": cluster_map[i], "r1": real_centers[i][0], "delta": real_centers[i][1]}
for i in range(len(real_centers))
],
"cluster_labels": labels,
"n_samples": n,
"method": "kmeans_3",
}
|