File size: 17,022 Bytes
b0ae393
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Shared, explicitly versioned utilities for the scaled FEPoID replication.

The goal is to follow the paper where it is unambiguous and to make every
necessary reconstruction choice visible.  In particular, the released
repository computes per-layer TwoNN values but does not implement FEPoID's
peak selector.  ``select_fepoid_layer`` below implements the active rule in
arXiv:2605.26366, lines 555--558 of the extracted ``main.tex``.
"""

from __future__ import annotations

import copy
import json
import random
import re
import time
from dataclasses import dataclass
from pathlib import Path
from typing import Iterable

import numpy as np
import torch
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import train_test_split
from skdim.id import TwoNN
from torch import nn


SEED = 2024
PROBE_SEED = 42

MODEL_SPECS = {
    "mistral": {
        "model_id": "mistralai/Mistral-7B-Instruct-v0.3",
        "revision": "c170c708c41dac9275d15a8fff4eca08d52bab71",
        "layers": 32,
        "hidden_size": 4096,
    },
    "llama": {
        "model_id": "meta-llama/Llama-3.1-8B-Instruct",
        "revision": "0e9e39f249a16976918f6564b8830bc894c89659",
        "layers": 32,
        "hidden_size": 4096,
    },
}

JUDGE_SPEC = {
    "model_id": "mistralai/Ministral-8B-Instruct-2410",
    "revision": "2f494a194c5b980dfb9772cb92d26cbb671fce5a",
}

COQA_SPEC = {
    "dataset_id": "stanfordnlp/coqa",
    "revision": "0d9e9952f1ef6e5415492d3d84b5873259137e3c",
}


def seed_everything(seed: int = SEED) -> None:
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)


def coqa_prompt(context: str, question: str) -> str:
    """The active context-aware prompt in the authors' released code."""

    return (
        "Answer the question as briefly as possible, based only on the context:\n"
        f" Context:{context.strip()}\n Question:{question.strip()}\n Answer:"
    )


# Verbatim behavior of the released FST path, reorganized dependency-free.
FST_FILTERS = (
    "\n", "Q:", "A:", "question:", "answer:", "Question:", "Answer:",
    "Questions:", "questions:", "QUESTION:", "ANSWER:", "REF", ".Forms",
    "http", "php", "Question", "Answer",
)
FST_WORD_ABBREVIATIONS = {
    "Mr", "Mrs", "Ms", "Dr", "Prof", "Sr", "Jr", "Gen", "Brig", "Adm",
    "Rear", "Lt", "Col", "Maj", "Capt", "St", "vs", "etc", "Fig", "Eq", "No",
}
FST_MULTI_DOT_ABBREVIATION = re.compile(r"(?:[A-Za-z]\.){2,}$")
FST_SINGLE_INITIAL = re.compile(r"^[A-Za-z]$")


def extract_first_sentence(text: str) -> str:
    text = text.strip()
    length = len(text)
    cursor = 0
    while cursor < length:
        char = text[cursor]
        if char not in ".!?":
            cursor += 1
            continue
        if char == "." and text[cursor : cursor + 3] == "...":
            cursor += 3
            continue
        if (
            char == "."
            and 0 < cursor < length - 1
            and text[cursor - 1].isdigit()
            and text[cursor + 1].isdigit()
        ):
            cursor += 1
            continue
        left = cursor - 1
        while left >= 0 and (text[left].isalpha() or text[left] == "."):
            left -= 1
        token = text[left + 1 : cursor].strip()
        if char == ".":
            right_is_letter_dot = (
                cursor + 2 < length
                and text[cursor + 1].isalpha()
                and text[cursor + 2] == "."
            )
            if cursor > 0 and text[cursor - 1].isalpha() and right_is_letter_dot:
                cursor += 1
                continue
        if "." in token and FST_MULTI_DOT_ABBREVIATION.match(token + "."):
            cursor += 1
            continue
        if token in FST_WORD_ABBREVIATIONS:
            if token == "No":
                right = cursor + 1
                while right < length and text[right].isspace():
                    right += 1
                if right < length and text[right].isdigit():
                    cursor += 1
                    continue
            else:
                cursor += 1
                continue
        if FST_SINGLE_INITIAL.match(token):
            right = cursor + 1
            while right < length and text[right].isspace():
                right += 1
            if right < length and text[right].isupper():
                cursor += 1
                continue
        return text[: cursor + 1].strip()
    return text


def first_sentence_truncation(answer: str) -> str:
    original = answer.strip()
    cut_position = len(answer)
    for marker in FST_FILTERS:
        marker_position = answer.find(marker)
        if 0 <= marker_position < cut_position:
            cut_position = marker_position
    filtered = answer[:cut_position].strip() or original
    return extract_first_sentence(filtered)


def judge_prompt(context: str, question: str, references: list[str], model_answer: str) -> str:
    """The context-aware LLM-judge prompt from the released repository."""

    reference_answer = "; ".join(references)
    return f"""
        Evaluate the following answers to questions. For each question you are given a model answer and the correct answer.
        You must determine if the model answer is correct or not. If the model answer is correct, write '1' and if it is not correct, write '0'.
        For example:

        Question: who is the young guitarist who played with buddy guy?
        Ground Truth: Quinn Sullivan
        Model Answer: Ronnie Earl Explanation: Ronnie Earl is an American blues guitarist and singer who has played with many famous blues musicians, including Buddy Guy. He is known for his soulful and melodic playing style, and has released many albums that blend blues, jazz, and rock music. Earl has also been a member of the Buddy Guy Blues Band and has played with other notable blues musicians such as B.B. King, Eric Clapton, and Stevie Ray Vaughan. He is considered one of the most
        Correctness: 0

        Question: name of the first episode of stranger things
        Ground Truth: Chapter One : The Vanishing of Will Byers
        Model Answer: The disappearance of Will Byers. Explanation: The first episode of the first season of Stranger Things is titled "The Vanishing of Will Byers". The episode introduces the main characters and sets the tone for the rest of the series. It follows the story of Will Byers, a young boy who goes missing in the fictional town of Hawkins, Indiana, and the subsequent search for him by his mother Joyce and his friends Mike, Dustin, and Lucas. The episode sets the stage for the supernatural
        Correctness: 1

        Context: {context}
        Question: {question}
        Ground Truth: {reference_answer}
        Model Answer: {model_answer}
        Correctness:
        """.strip()


def released_substring_match(references: Iterable[str], model_answer: str) -> bool:
    """Match the release, which uses containment despite the paper saying exact match."""

    normalized = model_answer.strip().lower()
    return any(str(reference).strip().lower() in normalized for reference in references)


def parse_judge_output(text: str) -> int:
    text = text.replace(".</s>", "").replace("</s>", "").split("\n")[0].strip().strip(".")
    index_one = text.find("1")
    index_zero = text.find("0")
    if index_one != -1 and (index_zero == -1 or index_one < index_zero):
        return 1
    if index_zero != -1 and (index_one == -1 or index_zero < index_one):
        return 0
    return 0


def select_fepoid_layer(ids: Iterable[float], horizon: int = 7) -> tuple[int, list[int], list[int]]:
    """Implement the paper's active FEPoID rule.

    Reconstruction choices missing from the paper/release:
    * layers are zero based;
    * local maxima are strict on the left and non-increasing on the right;
    * endpoints are not candidate peaks;
    * if no peak survives, layer 0 is selected (the paper's "shallowest").
    """

    curve = np.asarray(list(ids), dtype=np.float64)
    if curve.ndim != 1 or len(curve) < 3 or not np.isfinite(curve).all():
        raise ValueError("FEPoID requires at least three finite per-layer ID values")
    candidates = [
        layer
        for layer in range(1, len(curve) - 1)
        if curve[layer] > curve[layer - 1] and curve[layer] >= curve[layer + 1]
    ]
    survivors: list[int] = []
    for layer in candidates:
        endpoint = min(layer + horizon, len(curve) - 1)
        tail_strictly_increases = all(
            curve[index] < curve[index + 1]
            for index in range(layer + 1, endpoint)
        )
        discard = curve[layer] < curve[endpoint] and tail_strictly_increases
        if not discard:
            survivors.append(layer)
    return (survivors[0] if survivors else 0), candidates, survivors


def twonn_curve(hidden_states: np.ndarray, device: torch.device) -> tuple[np.ndarray, float]:
    """Compute paper-compatible TwoNN estimates using exact pairwise distances."""

    if hidden_states.ndim != 3:
        raise ValueError(f"expected [N,L,D], received {hidden_states.shape}")
    started = time.perf_counter()
    values = []
    for layer in range(hidden_states.shape[1]):
        features = torch.from_numpy(hidden_states[:, layer]).to(device=device, dtype=torch.float32)
        distances = torch.cdist(features, features)
        distances.fill_diagonal_(float("inf"))
        nearest_two = torch.topk(distances, k=2, dim=1, largest=False).values.cpu().numpy()
        estimator = TwoNN(dist=True)
        estimator.fit(nearest_two.astype(np.float64, copy=False))
        values.append(float(estimator.dimension_))
        del features, distances
    if device.type == "cuda":
        torch.cuda.synchronize(device)
    return np.asarray(values), time.perf_counter() - started


class Probe(nn.Module):
    def __init__(self, dimension: int):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(dimension, 512),
            nn.ReLU(),
            nn.Linear(512, 256),
            nn.ReLU(),
            nn.Linear(256, 128),
            nn.ReLU(),
            nn.Linear(128, 64),
            nn.ReLU(),
            nn.Linear(64, 2),
        )

    def forward(self, values: torch.Tensor) -> torch.Tensor:
        return self.net(values.float())


@dataclass
class ProbeResult:
    auroc: np.ndarray
    validation_loss: np.ndarray
    elapsed_seconds: float
    train_indices: np.ndarray
    validation_indices: np.ndarray


def probe_curve(
    train_hidden: np.ndarray,
    train_labels: np.ndarray,
    test_hidden: np.ndarray,
    test_labels: np.ndarray,
    device: torch.device,
    epochs: int = 15,
    learning_rate: float = 1e-2,
    weight_decay: float = 1e-4,
) -> ProbeResult:
    """Train the released MLP at every layer with one fixed seeded split.

    The fixed split follows the paper's stated setup.  It intentionally avoids
    the release's implementation accident of drawing a fresh validation split
    for each layer, which would confound layer comparisons.
    """

    labels = np.asarray(train_labels, dtype=np.int64)
    indices = np.arange(len(labels))
    train_indices, validation_indices = train_test_split(
        indices,
        test_size=0.1,
        random_state=PROBE_SEED,
        stratify=labels,
    )
    criterion = nn.CrossEntropyLoss()
    aurocs: list[float] = []
    losses: list[float] = []
    seed_everything(PROBE_SEED)
    started = time.perf_counter()

    for layer in range(train_hidden.shape[1]):
        train_x = torch.from_numpy(train_hidden[train_indices, layer]).to(device=device, dtype=torch.float32)
        train_y = torch.from_numpy(labels[train_indices]).to(device=device, dtype=torch.long)
        validation_x = torch.from_numpy(train_hidden[validation_indices, layer]).to(device=device, dtype=torch.float32)
        validation_y = torch.from_numpy(labels[validation_indices]).to(device=device, dtype=torch.long)
        test_x = torch.from_numpy(test_hidden[:, layer]).to(device=device, dtype=torch.float32)

        model = Probe(train_x.shape[1]).to(device)
        optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
        scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.1)
        best_loss = float("inf")
        best_state = None
        for _epoch in range(epochs):
            model.train()
            permutation = torch.randperm(len(train_x), device=device)
            for start in range(0, len(train_x), 2048):
                batch = permutation[start : start + 2048]
                optimizer.zero_grad(set_to_none=True)
                loss = criterion(model(train_x[batch]), train_y[batch])
                loss.backward()
                optimizer.step()
            model.eval()
            with torch.no_grad():
                validation_loss = float(criterion(model(validation_x), validation_y).item())
            if validation_loss < best_loss:
                best_loss = validation_loss
                best_state = copy.deepcopy(model.state_dict())
            scheduler.step()
        if best_state is not None:
            model.load_state_dict(best_state)
        model.eval()
        with torch.no_grad():
            probabilities = torch.softmax(model(test_x), dim=-1)[:, 1].cpu().numpy()
        aurocs.append(float(roc_auc_score(test_labels, probabilities)))
        losses.append(best_loss)
        del model, optimizer, train_x, train_y, validation_x, validation_y, test_x

    if device.type == "cuda":
        torch.cuda.synchronize(device)
    return ProbeResult(
        auroc=np.asarray(aurocs),
        validation_loss=np.asarray(losses),
        elapsed_seconds=time.perf_counter() - started,
        train_indices=np.asarray(train_indices),
        validation_indices=np.asarray(validation_indices),
    )


def analyze_hidden_states(
    train_hidden: np.ndarray,
    train_labels: np.ndarray,
    test_hidden: np.ndarray,
    test_labels: np.ndarray,
    horizon: int,
    device: torch.device,
) -> tuple[list[dict], dict]:
    if len(np.unique(train_labels)) != 2 or len(np.unique(test_labels)) != 2:
        raise ValueError("both train and test labels must contain two classes")
    ids, id_seconds = twonn_curve(train_hidden, device)
    probes = probe_curve(train_hidden, train_labels, test_hidden, test_labels, device)
    selected, candidates, survivors = select_fepoid_layer(ids, horizon)
    max_id_layer = int(np.argmax(ids))
    oracle_layer = int(np.argmax(probes.auroc))
    rows = [
        {
            "layer_zero_based": layer,
            "layer_one_based": layer + 1,
            "twonn_id": float(ids[layer]),
            "probe_auroc": float(probes.auroc[layer]),
            "validation_loss": float(probes.validation_loss[layer]),
            "fepoid_selected": layer == selected,
            "local_peak": layer in candidates,
            "surviving_peak": layer in survivors,
        }
        for layer in range(len(ids))
    ]
    summary = {
        "fepoid_layer_zero_based": selected,
        "fepoid_layer_one_based": selected + 1,
        "fepoid_auroc": float(probes.auroc[selected]),
        "last_layer_auroc": float(probes.auroc[-1]),
        "oracle_layer_zero_based": oracle_layer,
        "oracle_layer_one_based": oracle_layer + 1,
        "oracle_auroc": float(probes.auroc[oracle_layer]),
        "max_id_layer_zero_based": max_id_layer,
        "max_id_layer_one_based": max_id_layer + 1,
        "max_id_layer_auroc": float(probes.auroc[max_id_layer]),
        "fepoid_gap_to_oracle": float(probes.auroc[oracle_layer] - probes.auroc[selected]),
        "fepoid_gain_over_last": float(probes.auroc[selected] - probes.auroc[-1]),
        "id_seconds": id_seconds,
        "probe_seconds": probes.elapsed_seconds,
        "local_peak_layers_zero_based": candidates,
        "surviving_peak_layers_zero_based": survivors,
        "train_label_counts": {
            str(value): int((train_labels == value).sum()) for value in np.unique(train_labels)
        },
        "test_label_counts": {
            str(value): int((test_labels == value).sum()) for value in np.unique(test_labels)
        },
        "probe_train_indices_sha256": array_sha256(probes.train_indices),
        "probe_validation_indices_sha256": array_sha256(probes.validation_indices),
    }
    return rows, summary


def array_sha256(array: np.ndarray) -> str:
    import hashlib

    return hashlib.sha256(np.ascontiguousarray(array).tobytes()).hexdigest()


def write_json(path: Path, value: object) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text(json.dumps(value, indent=2, sort_keys=True) + "\n", encoding="utf-8")


def write_jsonl(path: Path, rows: Iterable[dict]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open("w", encoding="utf-8") as handle:
        for row in rows:
            handle.write(json.dumps(row, ensure_ascii=False, separators=(",", ":")) + "\n")