interp-engine 0.0.24__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- interp_engine-0.0.24.dist-info/METADATA +21 -0
- interp_engine-0.0.24.dist-info/RECORD +28 -0
- interp_engine-0.0.24.dist-info/WHEEL +4 -0
- interp_engine-0.0.24.dist-info/licenses/LICENSE +19 -0
- neuron_explainer/__init__.py +0 -0
- neuron_explainer/activations/__init__.py +0 -0
- neuron_explainer/activations/activation_records.py +130 -0
- neuron_explainer/activations/activations.py +311 -0
- neuron_explainer/activations/attention_utils.py +121 -0
- neuron_explainer/activations/token_connections.py +59 -0
- neuron_explainer/api_client.py +190 -0
- neuron_explainer/azure.py +5 -0
- neuron_explainer/explanations/__init__.py +0 -0
- neuron_explainer/explanations/calibrated_simulator.py +194 -0
- neuron_explainer/explanations/explainer.py +2585 -0
- neuron_explainer/explanations/explanations.py +230 -0
- neuron_explainer/explanations/few_shot_examples.py +3125 -0
- neuron_explainer/explanations/prompt_builder.py +118 -0
- neuron_explainer/explanations/puzzles.json +399 -0
- neuron_explainer/explanations/puzzles.py +50 -0
- neuron_explainer/explanations/scoring.py +155 -0
- neuron_explainer/explanations/simulator.py +1121 -0
- neuron_explainer/explanations/test_explainer.py +227 -0
- neuron_explainer/explanations/test_simulator.py +269 -0
- neuron_explainer/explanations/token_space_few_shot_examples.py +212 -0
- neuron_explainer/fast_dataclasses/__init__.py +3 -0
- neuron_explainer/fast_dataclasses/fast_dataclasses.py +85 -0
- neuron_explainer/fast_dataclasses/test_fast_dataclasses.py +83 -0
|
@@ -0,0 +1,3125 @@
|
|
|
1
|
+
# Edited examples - adds top_positive_logits to the examples
|
|
2
|
+
# Few-shot examples for generating and simulating neuron explanations.
|
|
3
|
+
|
|
4
|
+
from __future__ import annotations
|
|
5
|
+
|
|
6
|
+
from dataclasses import dataclass
|
|
7
|
+
from enum import Enum
|
|
8
|
+
from typing import List, Optional
|
|
9
|
+
|
|
10
|
+
from neuron_explainer.activations.activations import ActivationRecord
|
|
11
|
+
from neuron_explainer.fast_dataclasses import FastDataclass
|
|
12
|
+
|
|
13
|
+
|
|
14
|
+
@dataclass
|
|
15
|
+
class Example(FastDataclass):
|
|
16
|
+
activation_records: List[ActivationRecord]
|
|
17
|
+
explanation: str
|
|
18
|
+
first_revealed_activation_indices: List[int]
|
|
19
|
+
"""
|
|
20
|
+
For each activation record, the index of the first token for which the activation value in the
|
|
21
|
+
prompt should be an actual number rather than "unknown".
|
|
22
|
+
|
|
23
|
+
Examples all start with the activations rendered as "unknown", then transition to revealing
|
|
24
|
+
specific normalized activation values. The goal is to lead the model to predict that activation
|
|
25
|
+
sequences will eventually transition to predicting specific activation values instead of just
|
|
26
|
+
"unknown". This lets us cheat and get predictions of activation values for every token in a
|
|
27
|
+
single round of inference by having the activations in the sequence we're predicting always be
|
|
28
|
+
"unknown" in the prompt: the model will always think that maybe the next token will be a real
|
|
29
|
+
activation.
|
|
30
|
+
"""
|
|
31
|
+
token_index_to_score: Optional[int] = None
|
|
32
|
+
"""
|
|
33
|
+
If the prompt is used as an example for one-token-at-a-time scoring, this is the index of the
|
|
34
|
+
token to score.
|
|
35
|
+
"""
|
|
36
|
+
top_positive_logits: Optional[List[str]] = None
|
|
37
|
+
|
|
38
|
+
|
|
39
|
+
class FewShotExampleSet(Enum):
|
|
40
|
+
"""Determines which few-shot examples to use when sampling explanations."""
|
|
41
|
+
|
|
42
|
+
ORIGINAL = "original"
|
|
43
|
+
NEWER = "newer"
|
|
44
|
+
TEST = "test"
|
|
45
|
+
LOGITS = "logits"
|
|
46
|
+
ACTIVATIONS = "activations"
|
|
47
|
+
JL_FINE_TUNED = "jl_fine_tuned"
|
|
48
|
+
|
|
49
|
+
@classmethod
|
|
50
|
+
def from_string(cls, string: str) -> FewShotExampleSet:
|
|
51
|
+
for example_set in FewShotExampleSet:
|
|
52
|
+
if example_set.value == string:
|
|
53
|
+
return example_set
|
|
54
|
+
raise ValueError(f"Unrecognized example set: {string}")
|
|
55
|
+
|
|
56
|
+
def get_examples(self) -> list[Example]:
|
|
57
|
+
"""Returns regular examples for use in a few-shot prompt."""
|
|
58
|
+
if self is FewShotExampleSet.ORIGINAL:
|
|
59
|
+
return ORIGINAL_EXAMPLES
|
|
60
|
+
elif self is FewShotExampleSet.LOGITS:
|
|
61
|
+
return LOGITS_EXAMPLES
|
|
62
|
+
elif self is FewShotExampleSet.ACTIVATIONS:
|
|
63
|
+
return ACTIVATIONS_EXAMPLES
|
|
64
|
+
elif self is FewShotExampleSet.NEWER:
|
|
65
|
+
return NEWER_EXAMPLES
|
|
66
|
+
elif self is FewShotExampleSet.TEST:
|
|
67
|
+
return TEST_EXAMPLES
|
|
68
|
+
elif self is FewShotExampleSet.JL_FINE_TUNED:
|
|
69
|
+
return JL_FINE_TUNED_EXAMPLES
|
|
70
|
+
else:
|
|
71
|
+
raise ValueError(f"Unhandled example set: {self}")
|
|
72
|
+
|
|
73
|
+
def get_single_token_prediction_example(self) -> Example:
|
|
74
|
+
"""
|
|
75
|
+
Returns an example suitable for use in a subprompt for predicting a single token's
|
|
76
|
+
normalized activation, for use with the "one token at a time" scoring approach.
|
|
77
|
+
"""
|
|
78
|
+
if self is FewShotExampleSet.NEWER:
|
|
79
|
+
return NEWER_SINGLE_TOKEN_EXAMPLE
|
|
80
|
+
elif self is FewShotExampleSet.TEST:
|
|
81
|
+
return TEST_SINGLE_TOKEN_EXAMPLE
|
|
82
|
+
else:
|
|
83
|
+
raise ValueError(f"Unhandled example set: {self}")
|
|
84
|
+
|
|
85
|
+
|
|
86
|
+
TEST_EXAMPLES = [
|
|
87
|
+
Example(
|
|
88
|
+
activation_records=[
|
|
89
|
+
ActivationRecord(
|
|
90
|
+
tokens=["a", "b", "c"],
|
|
91
|
+
activations=[1.0, 0.0, 0.0],
|
|
92
|
+
),
|
|
93
|
+
ActivationRecord(
|
|
94
|
+
tokens=["d", "e", "f"],
|
|
95
|
+
activations=[0.0, 1.0, 0.0],
|
|
96
|
+
),
|
|
97
|
+
],
|
|
98
|
+
explanation="vowels",
|
|
99
|
+
first_revealed_activation_indices=[0, 1],
|
|
100
|
+
),
|
|
101
|
+
]
|
|
102
|
+
|
|
103
|
+
TEST_SINGLE_TOKEN_EXAMPLE = Example(
|
|
104
|
+
activation_records=[
|
|
105
|
+
ActivationRecord(
|
|
106
|
+
activations=[0.0, 0.0, 1.0],
|
|
107
|
+
tokens=["g", "h", "i"],
|
|
108
|
+
),
|
|
109
|
+
],
|
|
110
|
+
first_revealed_activation_indices=[],
|
|
111
|
+
token_index_to_score=2,
|
|
112
|
+
explanation="test explanation",
|
|
113
|
+
)
|
|
114
|
+
|
|
115
|
+
|
|
116
|
+
ORIGINAL_EXAMPLES = [
|
|
117
|
+
Example(
|
|
118
|
+
activation_records=[
|
|
119
|
+
ActivationRecord(
|
|
120
|
+
tokens=[
|
|
121
|
+
"t",
|
|
122
|
+
"urt",
|
|
123
|
+
"ur",
|
|
124
|
+
"ro",
|
|
125
|
+
" is",
|
|
126
|
+
" fab",
|
|
127
|
+
"ulously",
|
|
128
|
+
" funny",
|
|
129
|
+
" and",
|
|
130
|
+
" over",
|
|
131
|
+
" the",
|
|
132
|
+
" top",
|
|
133
|
+
" as",
|
|
134
|
+
" a",
|
|
135
|
+
" '",
|
|
136
|
+
"very",
|
|
137
|
+
" sneaky",
|
|
138
|
+
"'",
|
|
139
|
+
" but",
|
|
140
|
+
"ler",
|
|
141
|
+
" who",
|
|
142
|
+
" excel",
|
|
143
|
+
"s",
|
|
144
|
+
" in",
|
|
145
|
+
" the",
|
|
146
|
+
" art",
|
|
147
|
+
" of",
|
|
148
|
+
" impossible",
|
|
149
|
+
" disappearing",
|
|
150
|
+
"/",
|
|
151
|
+
"re",
|
|
152
|
+
"app",
|
|
153
|
+
"earing",
|
|
154
|
+
" acts",
|
|
155
|
+
],
|
|
156
|
+
activations=[
|
|
157
|
+
-0.71,
|
|
158
|
+
-1.85,
|
|
159
|
+
-2.39,
|
|
160
|
+
-2.58,
|
|
161
|
+
-1.34,
|
|
162
|
+
-1.92,
|
|
163
|
+
-1.69,
|
|
164
|
+
-0.84,
|
|
165
|
+
-1.25,
|
|
166
|
+
-1.75,
|
|
167
|
+
-1.42,
|
|
168
|
+
-1.47,
|
|
169
|
+
-1.51,
|
|
170
|
+
-0.8,
|
|
171
|
+
-1.89,
|
|
172
|
+
-1.56,
|
|
173
|
+
-1.63,
|
|
174
|
+
0.44,
|
|
175
|
+
-1.87,
|
|
176
|
+
-2.55,
|
|
177
|
+
-2.09,
|
|
178
|
+
-1.76,
|
|
179
|
+
-1.33,
|
|
180
|
+
-0.88,
|
|
181
|
+
-1.63,
|
|
182
|
+
-2.39,
|
|
183
|
+
-2.63,
|
|
184
|
+
-0.99,
|
|
185
|
+
2.83,
|
|
186
|
+
-1.11,
|
|
187
|
+
-1.19,
|
|
188
|
+
-1.33,
|
|
189
|
+
4.24,
|
|
190
|
+
-1.51,
|
|
191
|
+
],
|
|
192
|
+
),
|
|
193
|
+
ActivationRecord(
|
|
194
|
+
tokens=[
|
|
195
|
+
"esc",
|
|
196
|
+
"aping",
|
|
197
|
+
" the",
|
|
198
|
+
" studio",
|
|
199
|
+
" ,",
|
|
200
|
+
" pic",
|
|
201
|
+
"col",
|
|
202
|
+
"i",
|
|
203
|
+
" is",
|
|
204
|
+
" warm",
|
|
205
|
+
"ly",
|
|
206
|
+
" affecting",
|
|
207
|
+
" and",
|
|
208
|
+
" so",
|
|
209
|
+
" is",
|
|
210
|
+
" this",
|
|
211
|
+
" ad",
|
|
212
|
+
"roit",
|
|
213
|
+
"ly",
|
|
214
|
+
" minimalist",
|
|
215
|
+
" movie",
|
|
216
|
+
" .",
|
|
217
|
+
],
|
|
218
|
+
activations=[
|
|
219
|
+
-0.69,
|
|
220
|
+
4.12,
|
|
221
|
+
1.83,
|
|
222
|
+
-2.28,
|
|
223
|
+
-0.28,
|
|
224
|
+
-0.79,
|
|
225
|
+
-2.2,
|
|
226
|
+
-2.03,
|
|
227
|
+
-1.77,
|
|
228
|
+
-1.71,
|
|
229
|
+
-2.44,
|
|
230
|
+
1.6,
|
|
231
|
+
-1,
|
|
232
|
+
-0.38,
|
|
233
|
+
-1.93,
|
|
234
|
+
-2.09,
|
|
235
|
+
-1.63,
|
|
236
|
+
-1.94,
|
|
237
|
+
-1.82,
|
|
238
|
+
-1.64,
|
|
239
|
+
-1.32,
|
|
240
|
+
-1.92,
|
|
241
|
+
],
|
|
242
|
+
),
|
|
243
|
+
],
|
|
244
|
+
first_revealed_activation_indices=[10, 3],
|
|
245
|
+
explanation="present tense verbs ending in 'ing'",
|
|
246
|
+
top_positive_logits=[
|
|
247
|
+
" calling",
|
|
248
|
+
" sleeping",
|
|
249
|
+
" laughing",
|
|
250
|
+
" dancing",
|
|
251
|
+
" singing",
|
|
252
|
+
" frying",
|
|
253
|
+
" grounding",
|
|
254
|
+
" laying",
|
|
255
|
+
" eating",
|
|
256
|
+
" meeting",
|
|
257
|
+
],
|
|
258
|
+
),
|
|
259
|
+
Example(
|
|
260
|
+
activation_records=[
|
|
261
|
+
ActivationRecord(
|
|
262
|
+
tokens=[
|
|
263
|
+
"as",
|
|
264
|
+
" sac",
|
|
265
|
+
"char",
|
|
266
|
+
"ine",
|
|
267
|
+
" movies",
|
|
268
|
+
" go",
|
|
269
|
+
" ,",
|
|
270
|
+
" this",
|
|
271
|
+
" is",
|
|
272
|
+
" likely",
|
|
273
|
+
" to",
|
|
274
|
+
" cause",
|
|
275
|
+
" massive",
|
|
276
|
+
" cardiac",
|
|
277
|
+
" arrest",
|
|
278
|
+
" if",
|
|
279
|
+
" taken",
|
|
280
|
+
" in",
|
|
281
|
+
" large",
|
|
282
|
+
" doses",
|
|
283
|
+
" .",
|
|
284
|
+
],
|
|
285
|
+
activations=[
|
|
286
|
+
-0.14,
|
|
287
|
+
-1.37,
|
|
288
|
+
-0.68,
|
|
289
|
+
-2.27,
|
|
290
|
+
-1.46,
|
|
291
|
+
-1.11,
|
|
292
|
+
-0.9,
|
|
293
|
+
-2.48,
|
|
294
|
+
-2.07,
|
|
295
|
+
-3.49,
|
|
296
|
+
-2.16,
|
|
297
|
+
-1.79,
|
|
298
|
+
-0.23,
|
|
299
|
+
-0.04,
|
|
300
|
+
4.46,
|
|
301
|
+
-1.02,
|
|
302
|
+
-2.26,
|
|
303
|
+
-2.95,
|
|
304
|
+
-1.49,
|
|
305
|
+
-1.46,
|
|
306
|
+
-0.6,
|
|
307
|
+
],
|
|
308
|
+
),
|
|
309
|
+
ActivationRecord(
|
|
310
|
+
tokens=[
|
|
311
|
+
"shot",
|
|
312
|
+
" perhaps",
|
|
313
|
+
" '",
|
|
314
|
+
"art",
|
|
315
|
+
"istically",
|
|
316
|
+
"'",
|
|
317
|
+
" with",
|
|
318
|
+
" handheld",
|
|
319
|
+
" cameras",
|
|
320
|
+
" and",
|
|
321
|
+
" apparently",
|
|
322
|
+
" no",
|
|
323
|
+
" movie",
|
|
324
|
+
" lights",
|
|
325
|
+
" by",
|
|
326
|
+
" jo",
|
|
327
|
+
"aquin",
|
|
328
|
+
" b",
|
|
329
|
+
"aca",
|
|
330
|
+
"-",
|
|
331
|
+
"as",
|
|
332
|
+
"ay",
|
|
333
|
+
" ,",
|
|
334
|
+
" the",
|
|
335
|
+
" low",
|
|
336
|
+
"-",
|
|
337
|
+
"budget",
|
|
338
|
+
" production",
|
|
339
|
+
" swings",
|
|
340
|
+
" annoy",
|
|
341
|
+
"ingly",
|
|
342
|
+
" between",
|
|
343
|
+
" vert",
|
|
344
|
+
"igo",
|
|
345
|
+
" and",
|
|
346
|
+
" opacity",
|
|
347
|
+
" .",
|
|
348
|
+
],
|
|
349
|
+
activations=[
|
|
350
|
+
-0.09,
|
|
351
|
+
-3.53,
|
|
352
|
+
-0.72,
|
|
353
|
+
-2.36,
|
|
354
|
+
-1.05,
|
|
355
|
+
-1.12,
|
|
356
|
+
-2.49,
|
|
357
|
+
-2.14,
|
|
358
|
+
-1.98,
|
|
359
|
+
-1.59,
|
|
360
|
+
-2.62,
|
|
361
|
+
-2,
|
|
362
|
+
-2.73,
|
|
363
|
+
-2.87,
|
|
364
|
+
-3.23,
|
|
365
|
+
-1.11,
|
|
366
|
+
-2.23,
|
|
367
|
+
-0.97,
|
|
368
|
+
-2.28,
|
|
369
|
+
-2.37,
|
|
370
|
+
-1.5,
|
|
371
|
+
-2.81,
|
|
372
|
+
-1.73,
|
|
373
|
+
-3.14,
|
|
374
|
+
-2.61,
|
|
375
|
+
-1.7,
|
|
376
|
+
-3.08,
|
|
377
|
+
-4,
|
|
378
|
+
-0.71,
|
|
379
|
+
-2.48,
|
|
380
|
+
-1.39,
|
|
381
|
+
-1.96,
|
|
382
|
+
-1.09,
|
|
383
|
+
4.37,
|
|
384
|
+
-0.74,
|
|
385
|
+
-0.5,
|
|
386
|
+
-0.62,
|
|
387
|
+
],
|
|
388
|
+
),
|
|
389
|
+
],
|
|
390
|
+
first_revealed_activation_indices=[5, 20],
|
|
391
|
+
explanation="words related to physical medical conditions",
|
|
392
|
+
top_positive_logits=[
|
|
393
|
+
" fainting",
|
|
394
|
+
" bleed",
|
|
395
|
+
" stroke",
|
|
396
|
+
" hernia",
|
|
397
|
+
" disease",
|
|
398
|
+
" broken",
|
|
399
|
+
" cast",
|
|
400
|
+
" emergency",
|
|
401
|
+
" sprain",
|
|
402
|
+
" pain",
|
|
403
|
+
],
|
|
404
|
+
),
|
|
405
|
+
Example(
|
|
406
|
+
activation_records=[
|
|
407
|
+
ActivationRecord(
|
|
408
|
+
tokens=[
|
|
409
|
+
"the",
|
|
410
|
+
" sense",
|
|
411
|
+
" of",
|
|
412
|
+
" together",
|
|
413
|
+
"ness",
|
|
414
|
+
" in",
|
|
415
|
+
" our",
|
|
416
|
+
" town",
|
|
417
|
+
" is",
|
|
418
|
+
" strong",
|
|
419
|
+
" .",
|
|
420
|
+
],
|
|
421
|
+
activations=[
|
|
422
|
+
0,
|
|
423
|
+
0,
|
|
424
|
+
0,
|
|
425
|
+
1,
|
|
426
|
+
2,
|
|
427
|
+
0,
|
|
428
|
+
0.23,
|
|
429
|
+
0.5,
|
|
430
|
+
0,
|
|
431
|
+
0,
|
|
432
|
+
0,
|
|
433
|
+
],
|
|
434
|
+
),
|
|
435
|
+
ActivationRecord(
|
|
436
|
+
tokens=[
|
|
437
|
+
"a",
|
|
438
|
+
" buoy",
|
|
439
|
+
"ant",
|
|
440
|
+
" romantic",
|
|
441
|
+
" comedy",
|
|
442
|
+
" about",
|
|
443
|
+
" friendship",
|
|
444
|
+
" ,",
|
|
445
|
+
" love",
|
|
446
|
+
" ,",
|
|
447
|
+
" and",
|
|
448
|
+
" the",
|
|
449
|
+
" truth",
|
|
450
|
+
" that",
|
|
451
|
+
" we",
|
|
452
|
+
"'re",
|
|
453
|
+
" all",
|
|
454
|
+
" in",
|
|
455
|
+
" this",
|
|
456
|
+
" together",
|
|
457
|
+
" .",
|
|
458
|
+
],
|
|
459
|
+
activations=[
|
|
460
|
+
-0.15,
|
|
461
|
+
-2.33,
|
|
462
|
+
-1.4,
|
|
463
|
+
-2.17,
|
|
464
|
+
-2.53,
|
|
465
|
+
-0.85,
|
|
466
|
+
0.23,
|
|
467
|
+
-1.89,
|
|
468
|
+
0.09,
|
|
469
|
+
-0.47,
|
|
470
|
+
-0.5,
|
|
471
|
+
-0.58,
|
|
472
|
+
-0.87,
|
|
473
|
+
0.22,
|
|
474
|
+
0.58,
|
|
475
|
+
1.34,
|
|
476
|
+
0.98,
|
|
477
|
+
2.21,
|
|
478
|
+
2.84,
|
|
479
|
+
1.7,
|
|
480
|
+
-0.89,
|
|
481
|
+
],
|
|
482
|
+
),
|
|
483
|
+
],
|
|
484
|
+
first_revealed_activation_indices=[0, 10],
|
|
485
|
+
explanation="phrases related to community",
|
|
486
|
+
top_positive_logits=[
|
|
487
|
+
" community",
|
|
488
|
+
" together",
|
|
489
|
+
" love",
|
|
490
|
+
" friendship",
|
|
491
|
+
" friendly",
|
|
492
|
+
" joint",
|
|
493
|
+
" peace",
|
|
494
|
+
" companion",
|
|
495
|
+
" team",
|
|
496
|
+
" sympathy",
|
|
497
|
+
],
|
|
498
|
+
),
|
|
499
|
+
]
|
|
500
|
+
|
|
501
|
+
|
|
502
|
+
LOGITS_EXAMPLES = [
|
|
503
|
+
Example(
|
|
504
|
+
activation_records=[
|
|
505
|
+
ActivationRecord(
|
|
506
|
+
tokens=[
|
|
507
|
+
"She",
|
|
508
|
+
" was",
|
|
509
|
+
" taking",
|
|
510
|
+
" a",
|
|
511
|
+
" nap",
|
|
512
|
+
" when",
|
|
513
|
+
" her",
|
|
514
|
+
" phone",
|
|
515
|
+
" started",
|
|
516
|
+
" ringing",
|
|
517
|
+
".",
|
|
518
|
+
],
|
|
519
|
+
activations=[8, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0],
|
|
520
|
+
),
|
|
521
|
+
ActivationRecord(
|
|
522
|
+
tokens=[
|
|
523
|
+
"I",
|
|
524
|
+
" enjoy",
|
|
525
|
+
" watching",
|
|
526
|
+
" movies",
|
|
527
|
+
" with",
|
|
528
|
+
" my",
|
|
529
|
+
" family",
|
|
530
|
+
".",
|
|
531
|
+
],
|
|
532
|
+
activations=[0, 9, 0, 7.5, 0, 0, 0, 0],
|
|
533
|
+
),
|
|
534
|
+
],
|
|
535
|
+
first_revealed_activation_indices=[0, 1],
|
|
536
|
+
explanation='Method 1 fails: MAX_ACTIVATING_TOKENS (She, enjoy) are not similar tokens.\nMethod 2 succeeds: All TOKENS_AFTER_MAX_ACTIVATING_TOKEN have a pattern in common: they all start with "w".\nExplanation: say "w" words',
|
|
537
|
+
top_positive_logits=[
|
|
538
|
+
" walking",
|
|
539
|
+
" WA",
|
|
540
|
+
" waiting",
|
|
541
|
+
" was",
|
|
542
|
+
" we",
|
|
543
|
+
" WHAM",
|
|
544
|
+
" wish",
|
|
545
|
+
" win",
|
|
546
|
+
" wake",
|
|
547
|
+
" whisper",
|
|
548
|
+
],
|
|
549
|
+
),
|
|
550
|
+
Example(
|
|
551
|
+
activation_records=[
|
|
552
|
+
ActivationRecord(
|
|
553
|
+
tokens=[
|
|
554
|
+
"The",
|
|
555
|
+
" apple",
|
|
556
|
+
" and",
|
|
557
|
+
" banana",
|
|
558
|
+
" are",
|
|
559
|
+
" delicious",
|
|
560
|
+
" foods",
|
|
561
|
+
" that",
|
|
562
|
+
" provide",
|
|
563
|
+
" essential",
|
|
564
|
+
" vitamins",
|
|
565
|
+
" and",
|
|
566
|
+
" nutrients",
|
|
567
|
+
".",
|
|
568
|
+
],
|
|
569
|
+
activations=[0, 20, 0, 30, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0],
|
|
570
|
+
),
|
|
571
|
+
ActivationRecord(
|
|
572
|
+
tokens=[
|
|
573
|
+
"I",
|
|
574
|
+
" enjoy",
|
|
575
|
+
" eating",
|
|
576
|
+
" fresh",
|
|
577
|
+
" strawberries",
|
|
578
|
+
",",
|
|
579
|
+
" blueberries",
|
|
580
|
+
",",
|
|
581
|
+
" and",
|
|
582
|
+
" mangoes",
|
|
583
|
+
" during",
|
|
584
|
+
" the",
|
|
585
|
+
" summer",
|
|
586
|
+
" months",
|
|
587
|
+
".",
|
|
588
|
+
],
|
|
589
|
+
activations=[0, 0, 0, 0, 25, 0, 35, 0, 0, 15, 0, 0, 0, 0, 0],
|
|
590
|
+
),
|
|
591
|
+
],
|
|
592
|
+
first_revealed_activation_indices=[1, 4],
|
|
593
|
+
explanation="Method 1 succeeds: All MAX_ACTIVATING_TOKENS (banana, blueberries) are fruits.\nExplanation: fruits",
|
|
594
|
+
top_positive_logits=[
|
|
595
|
+
" apple",
|
|
596
|
+
" orange",
|
|
597
|
+
" pineapple",
|
|
598
|
+
" watermelon",
|
|
599
|
+
" kiwi",
|
|
600
|
+
" peach",
|
|
601
|
+
" pear",
|
|
602
|
+
" grape",
|
|
603
|
+
" cherry",
|
|
604
|
+
" plum",
|
|
605
|
+
],
|
|
606
|
+
),
|
|
607
|
+
Example(
|
|
608
|
+
activation_records=[
|
|
609
|
+
ActivationRecord(
|
|
610
|
+
tokens=[
|
|
611
|
+
"It",
|
|
612
|
+
" was",
|
|
613
|
+
" a",
|
|
614
|
+
" beautiful",
|
|
615
|
+
" day",
|
|
616
|
+
" outside",
|
|
617
|
+
" with",
|
|
618
|
+
" clear",
|
|
619
|
+
" skies",
|
|
620
|
+
" and",
|
|
621
|
+
" warm",
|
|
622
|
+
" sunshine",
|
|
623
|
+
".",
|
|
624
|
+
],
|
|
625
|
+
activations=[0, 0, 0, 0, 0, 0, 0, 0, 0, 8, 0, 0, 0],
|
|
626
|
+
),
|
|
627
|
+
ActivationRecord(
|
|
628
|
+
tokens=[
|
|
629
|
+
"And",
|
|
630
|
+
" the",
|
|
631
|
+
" garden",
|
|
632
|
+
" has",
|
|
633
|
+
" roses",
|
|
634
|
+
" and",
|
|
635
|
+
" tulips",
|
|
636
|
+
" and",
|
|
637
|
+
" daisies",
|
|
638
|
+
" and",
|
|
639
|
+
" sunflowers",
|
|
640
|
+
" blooming",
|
|
641
|
+
" together",
|
|
642
|
+
".",
|
|
643
|
+
],
|
|
644
|
+
activations=[9, 0, 0, 0, 0, 6, 0, 5, 0, 5, 0, 0, 0, 0],
|
|
645
|
+
),
|
|
646
|
+
],
|
|
647
|
+
first_revealed_activation_indices=[1, 4],
|
|
648
|
+
explanation='Method 1 succeeds: All MAX_ACTIVATING_TOKENS are the word "and".\nExplanation: and',
|
|
649
|
+
top_positive_logits=[
|
|
650
|
+
" elephant",
|
|
651
|
+
" guitar",
|
|
652
|
+
" mountain",
|
|
653
|
+
" bicycle",
|
|
654
|
+
" ocean",
|
|
655
|
+
" telescope",
|
|
656
|
+
" candle",
|
|
657
|
+
" umbrella",
|
|
658
|
+
" tornado",
|
|
659
|
+
" butterfly",
|
|
660
|
+
],
|
|
661
|
+
),
|
|
662
|
+
Example(
|
|
663
|
+
activation_records=[
|
|
664
|
+
ActivationRecord(
|
|
665
|
+
tokens=[
|
|
666
|
+
"the",
|
|
667
|
+
" civil",
|
|
668
|
+
" war",
|
|
669
|
+
" was",
|
|
670
|
+
" a",
|
|
671
|
+
" major",
|
|
672
|
+
" topic",
|
|
673
|
+
" in",
|
|
674
|
+
" history",
|
|
675
|
+
" class",
|
|
676
|
+
" .",
|
|
677
|
+
],
|
|
678
|
+
activations=[
|
|
679
|
+
0,
|
|
680
|
+
0,
|
|
681
|
+
15,
|
|
682
|
+
0,
|
|
683
|
+
0,
|
|
684
|
+
0,
|
|
685
|
+
0,
|
|
686
|
+
0,
|
|
687
|
+
0,
|
|
688
|
+
0,
|
|
689
|
+
0,
|
|
690
|
+
],
|
|
691
|
+
),
|
|
692
|
+
ActivationRecord(
|
|
693
|
+
tokens=[
|
|
694
|
+
" seasons",
|
|
695
|
+
" of",
|
|
696
|
+
" the",
|
|
697
|
+
" year",
|
|
698
|
+
" are",
|
|
699
|
+
" winter",
|
|
700
|
+
" ,",
|
|
701
|
+
" spring",
|
|
702
|
+
" ,",
|
|
703
|
+
" summer",
|
|
704
|
+
" ,",
|
|
705
|
+
" and",
|
|
706
|
+
" fall",
|
|
707
|
+
" or",
|
|
708
|
+
" autumn",
|
|
709
|
+
" in",
|
|
710
|
+
" some",
|
|
711
|
+
" places",
|
|
712
|
+
" .",
|
|
713
|
+
],
|
|
714
|
+
activations=[
|
|
715
|
+
0,
|
|
716
|
+
0,
|
|
717
|
+
0,
|
|
718
|
+
0,
|
|
719
|
+
0,
|
|
720
|
+
0,
|
|
721
|
+
0,
|
|
722
|
+
0,
|
|
723
|
+
0,
|
|
724
|
+
0,
|
|
725
|
+
0,
|
|
726
|
+
0,
|
|
727
|
+
0,
|
|
728
|
+
0,
|
|
729
|
+
0,
|
|
730
|
+
0,
|
|
731
|
+
10,
|
|
732
|
+
0,
|
|
733
|
+
0,
|
|
734
|
+
],
|
|
735
|
+
),
|
|
736
|
+
],
|
|
737
|
+
first_revealed_activation_indices=[2, 1],
|
|
738
|
+
explanation="Method 1 fails: MAX_ACTIVATING_TOKENS (war, some) are not all the same token.\nMethod 2 fails: TOKENS_AFTER_MAX_ACTIVATING_TOKEN (was, places) are not all similar tokens and don't have a text pattern in common.\nMethod 3 succeeds: All TOP_POSITIVE_LOGITS are the number 4.\nExplanation: 4",
|
|
739
|
+
top_positive_logits=[
|
|
740
|
+
" 4",
|
|
741
|
+
" four",
|
|
742
|
+
" fourth",
|
|
743
|
+
" 4th",
|
|
744
|
+
" IV",
|
|
745
|
+
" Four",
|
|
746
|
+
" FOUR",
|
|
747
|
+
" ~4",
|
|
748
|
+
" 4.0",
|
|
749
|
+
" quartet",
|
|
750
|
+
],
|
|
751
|
+
),
|
|
752
|
+
]
|
|
753
|
+
|
|
754
|
+
|
|
755
|
+
ACTIVATIONS_EXAMPLES = [
|
|
756
|
+
Example(
|
|
757
|
+
activation_records=[
|
|
758
|
+
ActivationRecord(
|
|
759
|
+
tokens=[
|
|
760
|
+
"The",
|
|
761
|
+
" apple",
|
|
762
|
+
" and",
|
|
763
|
+
" banana",
|
|
764
|
+
" are",
|
|
765
|
+
" delicious",
|
|
766
|
+
" foods",
|
|
767
|
+
" that",
|
|
768
|
+
" provide",
|
|
769
|
+
" essential",
|
|
770
|
+
" vitamins",
|
|
771
|
+
" and",
|
|
772
|
+
" nutrients",
|
|
773
|
+
".",
|
|
774
|
+
],
|
|
775
|
+
activations=[0, 20, 0, 30, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0],
|
|
776
|
+
),
|
|
777
|
+
ActivationRecord(
|
|
778
|
+
tokens=[
|
|
779
|
+
"I",
|
|
780
|
+
" enjoy",
|
|
781
|
+
" eating",
|
|
782
|
+
" fresh",
|
|
783
|
+
" strawberries",
|
|
784
|
+
",",
|
|
785
|
+
" blueberries",
|
|
786
|
+
",",
|
|
787
|
+
" and",
|
|
788
|
+
" mangoes",
|
|
789
|
+
" during",
|
|
790
|
+
" the",
|
|
791
|
+
" summer",
|
|
792
|
+
" months",
|
|
793
|
+
".",
|
|
794
|
+
],
|
|
795
|
+
activations=[0, 0, 0, 0, 25, 0, 35, 0, 0, 15, 0, 0, 0, 0, 0],
|
|
796
|
+
),
|
|
797
|
+
],
|
|
798
|
+
first_revealed_activation_indices=[1, 4],
|
|
799
|
+
explanation="Method 1 succeeds: All MAX_ACTIVATING_TOKENS (banana, blueberries) are fruits.\nExplanation: fruits",
|
|
800
|
+
),
|
|
801
|
+
Example(
|
|
802
|
+
activation_records=[
|
|
803
|
+
ActivationRecord(
|
|
804
|
+
tokens=[
|
|
805
|
+
"It",
|
|
806
|
+
" was",
|
|
807
|
+
" a",
|
|
808
|
+
" beautiful",
|
|
809
|
+
" day",
|
|
810
|
+
" outside",
|
|
811
|
+
" with",
|
|
812
|
+
" clear",
|
|
813
|
+
" skies",
|
|
814
|
+
" and",
|
|
815
|
+
" warm",
|
|
816
|
+
" sunshine",
|
|
817
|
+
".",
|
|
818
|
+
],
|
|
819
|
+
activations=[0, 0, 0, 0, 0, 0, 0, 0, 0, 8, 0, 0, 0],
|
|
820
|
+
),
|
|
821
|
+
ActivationRecord(
|
|
822
|
+
tokens=[
|
|
823
|
+
"And",
|
|
824
|
+
" the",
|
|
825
|
+
" garden",
|
|
826
|
+
" has",
|
|
827
|
+
" roses",
|
|
828
|
+
" and",
|
|
829
|
+
" tulips",
|
|
830
|
+
" and",
|
|
831
|
+
" daisies",
|
|
832
|
+
" and",
|
|
833
|
+
" sunflowers",
|
|
834
|
+
" blooming",
|
|
835
|
+
" together",
|
|
836
|
+
".",
|
|
837
|
+
],
|
|
838
|
+
activations=[9, 0, 0, 0, 0, 6, 0, 5, 0, 5, 0, 0, 0, 0],
|
|
839
|
+
),
|
|
840
|
+
],
|
|
841
|
+
first_revealed_activation_indices=[1, 4],
|
|
842
|
+
explanation='Method 1 succeeds: All MAX_ACTIVATING_TOKENS are the word "and".\nExplanation: and',
|
|
843
|
+
),
|
|
844
|
+
Example(
|
|
845
|
+
activation_records=[
|
|
846
|
+
ActivationRecord(
|
|
847
|
+
tokens=[
|
|
848
|
+
"the",
|
|
849
|
+
" civil",
|
|
850
|
+
" war",
|
|
851
|
+
" was",
|
|
852
|
+
" a",
|
|
853
|
+
" major",
|
|
854
|
+
" topic",
|
|
855
|
+
" in",
|
|
856
|
+
" history",
|
|
857
|
+
" class",
|
|
858
|
+
" .",
|
|
859
|
+
],
|
|
860
|
+
activations=[
|
|
861
|
+
0,
|
|
862
|
+
5,
|
|
863
|
+
10,
|
|
864
|
+
3,
|
|
865
|
+
2,
|
|
866
|
+
0,
|
|
867
|
+
3,
|
|
868
|
+
4,
|
|
869
|
+
3,
|
|
870
|
+
20,
|
|
871
|
+
2,
|
|
872
|
+
],
|
|
873
|
+
),
|
|
874
|
+
ActivationRecord(
|
|
875
|
+
tokens=[
|
|
876
|
+
" her",
|
|
877
|
+
" professor",
|
|
878
|
+
" teaches",
|
|
879
|
+
" psychology",
|
|
880
|
+
" courses",
|
|
881
|
+
" and",
|
|
882
|
+
" is",
|
|
883
|
+
" a",
|
|
884
|
+
" tough",
|
|
885
|
+
" grader",
|
|
886
|
+
" .",
|
|
887
|
+
],
|
|
888
|
+
activations=[0, 8, 15, 5, 10, 0, 0, 0, 2, 10, 2],
|
|
889
|
+
),
|
|
890
|
+
],
|
|
891
|
+
first_revealed_activation_indices=[2, 1],
|
|
892
|
+
explanation="Method 1 fails: MAX_ACTIVATING_TOKENS (class, teaches) are not all the same token.\nMethod 2 succeeds: The TOP_ACTIVATING_TEXTS are broadly about education.\nExplanation: education",
|
|
893
|
+
),
|
|
894
|
+
# # It doesn't seem to work as well when you give a 'fallback' example like this one - the model is too tempted to always use the fallback. So we comment it out.
|
|
895
|
+
# Example(
|
|
896
|
+
# activation_records=[
|
|
897
|
+
# ActivationRecord(
|
|
898
|
+
# tokens=[
|
|
899
|
+
# "the",
|
|
900
|
+
# " bright",
|
|
901
|
+
# " red",
|
|
902
|
+
# " apple",
|
|
903
|
+
# " fell",
|
|
904
|
+
# " from",
|
|
905
|
+
# " the",
|
|
906
|
+
# " tree",
|
|
907
|
+
# " and",
|
|
908
|
+
# " rolled",
|
|
909
|
+
# " down",
|
|
910
|
+
# " the",
|
|
911
|
+
# " hill",
|
|
912
|
+
# ".",
|
|
913
|
+
# ],
|
|
914
|
+
# activations=[
|
|
915
|
+
# 0,
|
|
916
|
+
# 2,
|
|
917
|
+
# 1,
|
|
918
|
+
# 15,
|
|
919
|
+
# 3,
|
|
920
|
+
# 0,
|
|
921
|
+
# 0,
|
|
922
|
+
# 5,
|
|
923
|
+
# 2,
|
|
924
|
+
# 8,
|
|
925
|
+
# 4,
|
|
926
|
+
# 0,
|
|
927
|
+
# 1,
|
|
928
|
+
# 0,
|
|
929
|
+
# ],
|
|
930
|
+
# ),
|
|
931
|
+
# ActivationRecord(
|
|
932
|
+
# tokens=[
|
|
933
|
+
# " the",
|
|
934
|
+
# " rocket",
|
|
935
|
+
# " launch",
|
|
936
|
+
# " was",
|
|
937
|
+
# " scheduled",
|
|
938
|
+
# " for",
|
|
939
|
+
# " midnight",
|
|
940
|
+
# " but",
|
|
941
|
+
# " weather",
|
|
942
|
+
# " conditions",
|
|
943
|
+
# " caused",
|
|
944
|
+
# " a",
|
|
945
|
+
# " delay",
|
|
946
|
+
# ".",
|
|
947
|
+
# ],
|
|
948
|
+
# activations=[0, 3, 12, 1, 2, 0, 8, 1, 4, 6, 2, 0, 5, 0],
|
|
949
|
+
# ),
|
|
950
|
+
# ],
|
|
951
|
+
# first_revealed_activation_indices=[3, 2],
|
|
952
|
+
# explanation="Method 1 fails: MAX_ACTIVATING_TOKENS (apple, launch) are not all the same token.\nMethod 2 fails: The TOP_ACTIVATING_TEXTS are completely unrelated with no overlap whatsoever (fruit falling vs space launch).\nSince both fail, I'll return the first token in MAX_ACTIVATING_TOKENS.\nExplanation: apple",
|
|
953
|
+
# ),
|
|
954
|
+
]
|
|
955
|
+
|
|
956
|
+
NEWER_EXAMPLES = [
|
|
957
|
+
Example(
|
|
958
|
+
activation_records=[
|
|
959
|
+
ActivationRecord(
|
|
960
|
+
tokens=[
|
|
961
|
+
"The",
|
|
962
|
+
" editors",
|
|
963
|
+
" of",
|
|
964
|
+
" Bi",
|
|
965
|
+
"opol",
|
|
966
|
+
"ym",
|
|
967
|
+
"ers",
|
|
968
|
+
" are",
|
|
969
|
+
" delighted",
|
|
970
|
+
" to",
|
|
971
|
+
" present",
|
|
972
|
+
" the",
|
|
973
|
+
" ",
|
|
974
|
+
"201",
|
|
975
|
+
"8",
|
|
976
|
+
" Murray",
|
|
977
|
+
" Goodman",
|
|
978
|
+
" Memorial",
|
|
979
|
+
" Prize",
|
|
980
|
+
" to",
|
|
981
|
+
" Professor",
|
|
982
|
+
" David",
|
|
983
|
+
" N",
|
|
984
|
+
".",
|
|
985
|
+
" Ber",
|
|
986
|
+
"atan",
|
|
987
|
+
" in",
|
|
988
|
+
" recognition",
|
|
989
|
+
" of",
|
|
990
|
+
" his",
|
|
991
|
+
" seminal",
|
|
992
|
+
" contributions",
|
|
993
|
+
" to",
|
|
994
|
+
" bi",
|
|
995
|
+
"oph",
|
|
996
|
+
"ysics",
|
|
997
|
+
" and",
|
|
998
|
+
" their",
|
|
999
|
+
" impact",
|
|
1000
|
+
" on",
|
|
1001
|
+
" our",
|
|
1002
|
+
" understanding",
|
|
1003
|
+
" of",
|
|
1004
|
+
" charge",
|
|
1005
|
+
" transport",
|
|
1006
|
+
" in",
|
|
1007
|
+
" biom",
|
|
1008
|
+
"olecules",
|
|
1009
|
+
".\n\n",
|
|
1010
|
+
"In",
|
|
1011
|
+
"aug",
|
|
1012
|
+
"ur",
|
|
1013
|
+
"ated",
|
|
1014
|
+
" in",
|
|
1015
|
+
" ",
|
|
1016
|
+
"200",
|
|
1017
|
+
"7",
|
|
1018
|
+
" in",
|
|
1019
|
+
" honor",
|
|
1020
|
+
" of",
|
|
1021
|
+
" the",
|
|
1022
|
+
" Bi",
|
|
1023
|
+
"opol",
|
|
1024
|
+
"ym",
|
|
1025
|
+
"ers",
|
|
1026
|
+
" Found",
|
|
1027
|
+
"ing",
|
|
1028
|
+
" Editor",
|
|
1029
|
+
",",
|
|
1030
|
+
" the",
|
|
1031
|
+
" prize",
|
|
1032
|
+
" is",
|
|
1033
|
+
" awarded",
|
|
1034
|
+
" for",
|
|
1035
|
+
" outstanding",
|
|
1036
|
+
" accomplishments",
|
|
1037
|
+
],
|
|
1038
|
+
activations=[
|
|
1039
|
+
0,
|
|
1040
|
+
0.01,
|
|
1041
|
+
0.01,
|
|
1042
|
+
0,
|
|
1043
|
+
0,
|
|
1044
|
+
0,
|
|
1045
|
+
-0.01,
|
|
1046
|
+
0,
|
|
1047
|
+
-0.01,
|
|
1048
|
+
0,
|
|
1049
|
+
0,
|
|
1050
|
+
0,
|
|
1051
|
+
0,
|
|
1052
|
+
0,
|
|
1053
|
+
0.04,
|
|
1054
|
+
0,
|
|
1055
|
+
0,
|
|
1056
|
+
0,
|
|
1057
|
+
0,
|
|
1058
|
+
0,
|
|
1059
|
+
0,
|
|
1060
|
+
0,
|
|
1061
|
+
0,
|
|
1062
|
+
0,
|
|
1063
|
+
0,
|
|
1064
|
+
0,
|
|
1065
|
+
0,
|
|
1066
|
+
0,
|
|
1067
|
+
0,
|
|
1068
|
+
0,
|
|
1069
|
+
3.39,
|
|
1070
|
+
0.12,
|
|
1071
|
+
0,
|
|
1072
|
+
-0.01,
|
|
1073
|
+
0,
|
|
1074
|
+
0,
|
|
1075
|
+
0,
|
|
1076
|
+
0,
|
|
1077
|
+
-0,
|
|
1078
|
+
0,
|
|
1079
|
+
-0,
|
|
1080
|
+
0,
|
|
1081
|
+
0,
|
|
1082
|
+
-0,
|
|
1083
|
+
0,
|
|
1084
|
+
0,
|
|
1085
|
+
0,
|
|
1086
|
+
0,
|
|
1087
|
+
0,
|
|
1088
|
+
0,
|
|
1089
|
+
0,
|
|
1090
|
+
0,
|
|
1091
|
+
0,
|
|
1092
|
+
0,
|
|
1093
|
+
0,
|
|
1094
|
+
0,
|
|
1095
|
+
0,
|
|
1096
|
+
0,
|
|
1097
|
+
0,
|
|
1098
|
+
0,
|
|
1099
|
+
0,
|
|
1100
|
+
-0,
|
|
1101
|
+
0,
|
|
1102
|
+
0,
|
|
1103
|
+
-0.01,
|
|
1104
|
+
0,
|
|
1105
|
+
0.41,
|
|
1106
|
+
0,
|
|
1107
|
+
0,
|
|
1108
|
+
0,
|
|
1109
|
+
-0.01,
|
|
1110
|
+
0,
|
|
1111
|
+
0,
|
|
1112
|
+
0,
|
|
1113
|
+
0,
|
|
1114
|
+
0,
|
|
1115
|
+
],
|
|
1116
|
+
),
|
|
1117
|
+
# We sometimes exceed the max context size when this is included :(
|
|
1118
|
+
# ActivationRecord(
|
|
1119
|
+
# tokens=[
|
|
1120
|
+
# " We",
|
|
1121
|
+
# " are",
|
|
1122
|
+
# " proud",
|
|
1123
|
+
# " of",
|
|
1124
|
+
# " our",
|
|
1125
|
+
# " national",
|
|
1126
|
+
# " achievements",
|
|
1127
|
+
# " in",
|
|
1128
|
+
# " mastering",
|
|
1129
|
+
# " all",
|
|
1130
|
+
# " aspects",
|
|
1131
|
+
# " of",
|
|
1132
|
+
# " the",
|
|
1133
|
+
# " fuel",
|
|
1134
|
+
# " cycle",
|
|
1135
|
+
# ".",
|
|
1136
|
+
# " The",
|
|
1137
|
+
# " current",
|
|
1138
|
+
# " international",
|
|
1139
|
+
# " interest",
|
|
1140
|
+
# " in",
|
|
1141
|
+
# " closing",
|
|
1142
|
+
# " the",
|
|
1143
|
+
# " fuel",
|
|
1144
|
+
# " cycle",
|
|
1145
|
+
# " is",
|
|
1146
|
+
# " a",
|
|
1147
|
+
# " vind",
|
|
1148
|
+
# "ication",
|
|
1149
|
+
# " of",
|
|
1150
|
+
# " Dr",
|
|
1151
|
+
# ".",
|
|
1152
|
+
# " B",
|
|
1153
|
+
# "hab",
|
|
1154
|
+
# "ha",
|
|
1155
|
+
# "’s",
|
|
1156
|
+
# " pioneering",
|
|
1157
|
+
# " vision",
|
|
1158
|
+
# " and",
|
|
1159
|
+
# " genius",
|
|
1160
|
+
# ],
|
|
1161
|
+
# activations=[
|
|
1162
|
+
# -0,
|
|
1163
|
+
# -0,
|
|
1164
|
+
# 0,
|
|
1165
|
+
# -0,
|
|
1166
|
+
# -0,
|
|
1167
|
+
# 0,
|
|
1168
|
+
# 0,
|
|
1169
|
+
# 0,
|
|
1170
|
+
# -0,
|
|
1171
|
+
# 0,
|
|
1172
|
+
# 0,
|
|
1173
|
+
# -0,
|
|
1174
|
+
# 0,
|
|
1175
|
+
# -0.01,
|
|
1176
|
+
# 0,
|
|
1177
|
+
# 0,
|
|
1178
|
+
# -0,
|
|
1179
|
+
# -0,
|
|
1180
|
+
# 0,
|
|
1181
|
+
# 0,
|
|
1182
|
+
# 0,
|
|
1183
|
+
# -0,
|
|
1184
|
+
# -0,
|
|
1185
|
+
# -0.01,
|
|
1186
|
+
# 0,
|
|
1187
|
+
# 0,
|
|
1188
|
+
# -0,
|
|
1189
|
+
# 0,
|
|
1190
|
+
# 0,
|
|
1191
|
+
# 0,
|
|
1192
|
+
# 0,
|
|
1193
|
+
# 0,
|
|
1194
|
+
# -0,
|
|
1195
|
+
# 0,
|
|
1196
|
+
# 0,
|
|
1197
|
+
# 0,
|
|
1198
|
+
# 2.15,
|
|
1199
|
+
# 0,
|
|
1200
|
+
# 0,
|
|
1201
|
+
# 0.03,
|
|
1202
|
+
# ],
|
|
1203
|
+
# ),
|
|
1204
|
+
],
|
|
1205
|
+
first_revealed_activation_indices=[7], # , 19],
|
|
1206
|
+
explanation="language related to something being groundbreaking",
|
|
1207
|
+
),
|
|
1208
|
+
Example(
|
|
1209
|
+
activation_records=[
|
|
1210
|
+
ActivationRecord(
|
|
1211
|
+
tokens=[
|
|
1212
|
+
'{"',
|
|
1213
|
+
"widget",
|
|
1214
|
+
"Class",
|
|
1215
|
+
'":"',
|
|
1216
|
+
"Variant",
|
|
1217
|
+
"Matrix",
|
|
1218
|
+
"Widget",
|
|
1219
|
+
'","',
|
|
1220
|
+
"back",
|
|
1221
|
+
"order",
|
|
1222
|
+
"Message",
|
|
1223
|
+
'":"',
|
|
1224
|
+
"Back",
|
|
1225
|
+
"ordered",
|
|
1226
|
+
'","',
|
|
1227
|
+
"back",
|
|
1228
|
+
"order",
|
|
1229
|
+
"Message",
|
|
1230
|
+
"Single",
|
|
1231
|
+
"Variant",
|
|
1232
|
+
'":"',
|
|
1233
|
+
"This",
|
|
1234
|
+
" item",
|
|
1235
|
+
" is",
|
|
1236
|
+
" back",
|
|
1237
|
+
"ordered",
|
|
1238
|
+
'.","',
|
|
1239
|
+
"ordered",
|
|
1240
|
+
"Selection",
|
|
1241
|
+
'":',
|
|
1242
|
+
"true",
|
|
1243
|
+
',"',
|
|
1244
|
+
"product",
|
|
1245
|
+
"Variant",
|
|
1246
|
+
"Id",
|
|
1247
|
+
'":',
|
|
1248
|
+
"0",
|
|
1249
|
+
',"',
|
|
1250
|
+
"variant",
|
|
1251
|
+
"Id",
|
|
1252
|
+
"Field",
|
|
1253
|
+
'":"',
|
|
1254
|
+
"product",
|
|
1255
|
+
"196",
|
|
1256
|
+
"39",
|
|
1257
|
+
"_V",
|
|
1258
|
+
"ariant",
|
|
1259
|
+
"Id",
|
|
1260
|
+
'","',
|
|
1261
|
+
"back",
|
|
1262
|
+
"order",
|
|
1263
|
+
"To",
|
|
1264
|
+
"Message",
|
|
1265
|
+
"Single",
|
|
1266
|
+
"Variant",
|
|
1267
|
+
'":"',
|
|
1268
|
+
"This",
|
|
1269
|
+
" item",
|
|
1270
|
+
" is",
|
|
1271
|
+
" back",
|
|
1272
|
+
"ordered",
|
|
1273
|
+
" and",
|
|
1274
|
+
" is",
|
|
1275
|
+
" expected",
|
|
1276
|
+
" by",
|
|
1277
|
+
" {",
|
|
1278
|
+
"0",
|
|
1279
|
+
"}.",
|
|
1280
|
+
'","',
|
|
1281
|
+
"low",
|
|
1282
|
+
"Price",
|
|
1283
|
+
'":',
|
|
1284
|
+
"999",
|
|
1285
|
+
"9",
|
|
1286
|
+
".",
|
|
1287
|
+
"0",
|
|
1288
|
+
',"',
|
|
1289
|
+
"attribute",
|
|
1290
|
+
"Indexes",
|
|
1291
|
+
'":[',
|
|
1292
|
+
'],"',
|
|
1293
|
+
"productId",
|
|
1294
|
+
'":',
|
|
1295
|
+
"196",
|
|
1296
|
+
"39",
|
|
1297
|
+
',"',
|
|
1298
|
+
"price",
|
|
1299
|
+
"V",
|
|
1300
|
+
"ariance",
|
|
1301
|
+
'":',
|
|
1302
|
+
"true",
|
|
1303
|
+
',"',
|
|
1304
|
+
],
|
|
1305
|
+
activations=[
|
|
1306
|
+
0,
|
|
1307
|
+
0,
|
|
1308
|
+
0,
|
|
1309
|
+
0,
|
|
1310
|
+
4.2,
|
|
1311
|
+
0,
|
|
1312
|
+
0,
|
|
1313
|
+
0,
|
|
1314
|
+
0,
|
|
1315
|
+
0,
|
|
1316
|
+
0,
|
|
1317
|
+
0,
|
|
1318
|
+
0,
|
|
1319
|
+
0,
|
|
1320
|
+
0,
|
|
1321
|
+
0,
|
|
1322
|
+
0,
|
|
1323
|
+
0,
|
|
1324
|
+
0,
|
|
1325
|
+
3.6,
|
|
1326
|
+
0,
|
|
1327
|
+
0,
|
|
1328
|
+
0,
|
|
1329
|
+
0,
|
|
1330
|
+
0,
|
|
1331
|
+
0,
|
|
1332
|
+
0,
|
|
1333
|
+
0,
|
|
1334
|
+
0,
|
|
1335
|
+
0,
|
|
1336
|
+
0,
|
|
1337
|
+
0,
|
|
1338
|
+
0,
|
|
1339
|
+
3.7,
|
|
1340
|
+
0,
|
|
1341
|
+
0,
|
|
1342
|
+
0,
|
|
1343
|
+
0,
|
|
1344
|
+
4.02,
|
|
1345
|
+
0,
|
|
1346
|
+
0,
|
|
1347
|
+
0,
|
|
1348
|
+
0,
|
|
1349
|
+
0,
|
|
1350
|
+
0,
|
|
1351
|
+
3.5,
|
|
1352
|
+
3.7,
|
|
1353
|
+
0,
|
|
1354
|
+
0,
|
|
1355
|
+
0,
|
|
1356
|
+
0,
|
|
1357
|
+
0,
|
|
1358
|
+
0,
|
|
1359
|
+
0,
|
|
1360
|
+
2.9,
|
|
1361
|
+
0,
|
|
1362
|
+
0,
|
|
1363
|
+
0,
|
|
1364
|
+
0,
|
|
1365
|
+
0,
|
|
1366
|
+
0,
|
|
1367
|
+
0,
|
|
1368
|
+
0,
|
|
1369
|
+
0,
|
|
1370
|
+
0,
|
|
1371
|
+
0,
|
|
1372
|
+
0,
|
|
1373
|
+
0,
|
|
1374
|
+
0,
|
|
1375
|
+
0,
|
|
1376
|
+
0,
|
|
1377
|
+
0,
|
|
1378
|
+
0,
|
|
1379
|
+
0,
|
|
1380
|
+
0,
|
|
1381
|
+
0,
|
|
1382
|
+
0,
|
|
1383
|
+
0,
|
|
1384
|
+
0,
|
|
1385
|
+
0,
|
|
1386
|
+
0,
|
|
1387
|
+
0,
|
|
1388
|
+
0,
|
|
1389
|
+
0,
|
|
1390
|
+
0,
|
|
1391
|
+
0,
|
|
1392
|
+
0,
|
|
1393
|
+
2.3,
|
|
1394
|
+
2.24,
|
|
1395
|
+
0,
|
|
1396
|
+
0,
|
|
1397
|
+
0,
|
|
1398
|
+
],
|
|
1399
|
+
),
|
|
1400
|
+
ActivationRecord(
|
|
1401
|
+
tokens=[
|
|
1402
|
+
"A",
|
|
1403
|
+
" regular",
|
|
1404
|
+
" look",
|
|
1405
|
+
" at",
|
|
1406
|
+
" the",
|
|
1407
|
+
" ups",
|
|
1408
|
+
" and",
|
|
1409
|
+
" downs",
|
|
1410
|
+
" of",
|
|
1411
|
+
" variant",
|
|
1412
|
+
" covers",
|
|
1413
|
+
" in",
|
|
1414
|
+
" the",
|
|
1415
|
+
" comics",
|
|
1416
|
+
" industry",
|
|
1417
|
+
"…\n\n",
|
|
1418
|
+
"Here",
|
|
1419
|
+
" are",
|
|
1420
|
+
" the",
|
|
1421
|
+
" Lego",
|
|
1422
|
+
" variant",
|
|
1423
|
+
" sketch",
|
|
1424
|
+
" covers",
|
|
1425
|
+
" by",
|
|
1426
|
+
" Leon",
|
|
1427
|
+
"el",
|
|
1428
|
+
" Cast",
|
|
1429
|
+
"ell",
|
|
1430
|
+
"ani",
|
|
1431
|
+
" for",
|
|
1432
|
+
" a",
|
|
1433
|
+
" variety",
|
|
1434
|
+
" of",
|
|
1435
|
+
" Marvel",
|
|
1436
|
+
" titles",
|
|
1437
|
+
",",
|
|
1438
|
+
],
|
|
1439
|
+
activations=[
|
|
1440
|
+
0,
|
|
1441
|
+
0,
|
|
1442
|
+
0,
|
|
1443
|
+
0,
|
|
1444
|
+
0,
|
|
1445
|
+
0,
|
|
1446
|
+
0,
|
|
1447
|
+
0,
|
|
1448
|
+
0,
|
|
1449
|
+
6.52,
|
|
1450
|
+
0,
|
|
1451
|
+
0,
|
|
1452
|
+
0,
|
|
1453
|
+
0,
|
|
1454
|
+
0,
|
|
1455
|
+
0,
|
|
1456
|
+
0,
|
|
1457
|
+
0,
|
|
1458
|
+
0,
|
|
1459
|
+
0,
|
|
1460
|
+
1.62,
|
|
1461
|
+
0,
|
|
1462
|
+
0,
|
|
1463
|
+
0,
|
|
1464
|
+
0,
|
|
1465
|
+
0,
|
|
1466
|
+
0,
|
|
1467
|
+
0,
|
|
1468
|
+
0,
|
|
1469
|
+
0,
|
|
1470
|
+
0,
|
|
1471
|
+
3.23,
|
|
1472
|
+
0,
|
|
1473
|
+
0,
|
|
1474
|
+
0,
|
|
1475
|
+
0,
|
|
1476
|
+
],
|
|
1477
|
+
),
|
|
1478
|
+
],
|
|
1479
|
+
first_revealed_activation_indices=[2, 8],
|
|
1480
|
+
explanation="the word “variant” and other words with the same ”vari” root",
|
|
1481
|
+
),
|
|
1482
|
+
]
|
|
1483
|
+
|
|
1484
|
+
|
|
1485
|
+
NEWER_SINGLE_TOKEN_EXAMPLE = Example(
|
|
1486
|
+
activation_records=[
|
|
1487
|
+
ActivationRecord(
|
|
1488
|
+
tokens=[
|
|
1489
|
+
"B",
|
|
1490
|
+
"10",
|
|
1491
|
+
" ",
|
|
1492
|
+
"111",
|
|
1493
|
+
" MON",
|
|
1494
|
+
"DAY",
|
|
1495
|
+
",",
|
|
1496
|
+
" F",
|
|
1497
|
+
"EB",
|
|
1498
|
+
"RU",
|
|
1499
|
+
"ARY",
|
|
1500
|
+
" ",
|
|
1501
|
+
"11",
|
|
1502
|
+
",",
|
|
1503
|
+
" ",
|
|
1504
|
+
"201",
|
|
1505
|
+
"9",
|
|
1506
|
+
" DON",
|
|
1507
|
+
"ATE",
|
|
1508
|
+
"fake higher scoring token", # See below.
|
|
1509
|
+
],
|
|
1510
|
+
activations=[
|
|
1511
|
+
0,
|
|
1512
|
+
0,
|
|
1513
|
+
0,
|
|
1514
|
+
0,
|
|
1515
|
+
0,
|
|
1516
|
+
0,
|
|
1517
|
+
0,
|
|
1518
|
+
0,
|
|
1519
|
+
0,
|
|
1520
|
+
0,
|
|
1521
|
+
0,
|
|
1522
|
+
0,
|
|
1523
|
+
0,
|
|
1524
|
+
0,
|
|
1525
|
+
0,
|
|
1526
|
+
0,
|
|
1527
|
+
0,
|
|
1528
|
+
0,
|
|
1529
|
+
0.37,
|
|
1530
|
+
# This fake activation makes the previous token's activation normalize to 8, which
|
|
1531
|
+
# might help address overconfidence in "10" activations for the one-token-at-a-time
|
|
1532
|
+
# scoring prompt. This value and the associated token don't actually appear anywhere
|
|
1533
|
+
# in the prompt.
|
|
1534
|
+
0.45,
|
|
1535
|
+
],
|
|
1536
|
+
),
|
|
1537
|
+
],
|
|
1538
|
+
first_revealed_activation_indices=[],
|
|
1539
|
+
token_index_to_score=18,
|
|
1540
|
+
explanation="instances of the token 'ate' as part of another word",
|
|
1541
|
+
)
|
|
1542
|
+
|
|
1543
|
+
|
|
1544
|
+
JL_FINE_TUNED_EXAMPLES = [
|
|
1545
|
+
Example(
|
|
1546
|
+
activation_records=[
|
|
1547
|
+
ActivationRecord(
|
|
1548
|
+
tokens=[
|
|
1549
|
+
"The",
|
|
1550
|
+
" cat",
|
|
1551
|
+
" jumped",
|
|
1552
|
+
" on",
|
|
1553
|
+
" my",
|
|
1554
|
+
" laptop",
|
|
1555
|
+
".",
|
|
1556
|
+
],
|
|
1557
|
+
activations=[0, 0, 0, 0, 0, 0, 0],
|
|
1558
|
+
),
|
|
1559
|
+
],
|
|
1560
|
+
first_revealed_activation_indices=[],
|
|
1561
|
+
explanation='the word "laptop" before the word "cat"',
|
|
1562
|
+
),
|
|
1563
|
+
Example(
|
|
1564
|
+
activation_records=[
|
|
1565
|
+
ActivationRecord(
|
|
1566
|
+
tokens=[
|
|
1567
|
+
"The",
|
|
1568
|
+
" cat",
|
|
1569
|
+
" jumped",
|
|
1570
|
+
" on",
|
|
1571
|
+
" my",
|
|
1572
|
+
" laptop",
|
|
1573
|
+
".",
|
|
1574
|
+
],
|
|
1575
|
+
activations=[0, 10, 0, 0, 0, 0, 0],
|
|
1576
|
+
),
|
|
1577
|
+
],
|
|
1578
|
+
first_revealed_activation_indices=[],
|
|
1579
|
+
explanation='the word "cat" before the word "laptop"',
|
|
1580
|
+
),
|
|
1581
|
+
Example(
|
|
1582
|
+
activation_records=[
|
|
1583
|
+
ActivationRecord(
|
|
1584
|
+
tokens=[
|
|
1585
|
+
"I",
|
|
1586
|
+
" am",
|
|
1587
|
+
" using",
|
|
1588
|
+
" a",
|
|
1589
|
+
" keyboard",
|
|
1590
|
+
".",
|
|
1591
|
+
],
|
|
1592
|
+
activations=[0, 0, 0, 0, 10, 0],
|
|
1593
|
+
),
|
|
1594
|
+
],
|
|
1595
|
+
first_revealed_activation_indices=[],
|
|
1596
|
+
explanation="the word before a period",
|
|
1597
|
+
),
|
|
1598
|
+
Example(
|
|
1599
|
+
activation_records=[
|
|
1600
|
+
ActivationRecord(
|
|
1601
|
+
tokens=[
|
|
1602
|
+
"The",
|
|
1603
|
+
" sun",
|
|
1604
|
+
" is",
|
|
1605
|
+
" shining",
|
|
1606
|
+
".",
|
|
1607
|
+
" The",
|
|
1608
|
+
" clouds",
|
|
1609
|
+
" are",
|
|
1610
|
+
" gone",
|
|
1611
|
+
".",
|
|
1612
|
+
" Great",
|
|
1613
|
+
" weather",
|
|
1614
|
+
"!",
|
|
1615
|
+
],
|
|
1616
|
+
activations=[0, 0, 0, 10, 0, 0, 0, 0, 10, 0, 0, 0, 0],
|
|
1617
|
+
),
|
|
1618
|
+
],
|
|
1619
|
+
first_revealed_activation_indices=[],
|
|
1620
|
+
explanation="the word before period",
|
|
1621
|
+
),
|
|
1622
|
+
]
|
|
1623
|
+
|
|
1624
|
+
NEWER_SINGLE_TOKEN_EXAMPLE = Example(
|
|
1625
|
+
activation_records=[
|
|
1626
|
+
ActivationRecord(
|
|
1627
|
+
tokens=[
|
|
1628
|
+
"B",
|
|
1629
|
+
"10",
|
|
1630
|
+
" ",
|
|
1631
|
+
"111",
|
|
1632
|
+
" MON",
|
|
1633
|
+
"DAY",
|
|
1634
|
+
",",
|
|
1635
|
+
" F",
|
|
1636
|
+
"EB",
|
|
1637
|
+
"RU",
|
|
1638
|
+
"ARY",
|
|
1639
|
+
" ",
|
|
1640
|
+
"11",
|
|
1641
|
+
",",
|
|
1642
|
+
" ",
|
|
1643
|
+
"201",
|
|
1644
|
+
"9",
|
|
1645
|
+
" DON",
|
|
1646
|
+
"ATE",
|
|
1647
|
+
"fake higher scoring token", # See below.
|
|
1648
|
+
],
|
|
1649
|
+
activations=[
|
|
1650
|
+
0,
|
|
1651
|
+
0,
|
|
1652
|
+
0,
|
|
1653
|
+
0,
|
|
1654
|
+
0,
|
|
1655
|
+
0,
|
|
1656
|
+
0,
|
|
1657
|
+
0,
|
|
1658
|
+
0,
|
|
1659
|
+
0,
|
|
1660
|
+
0,
|
|
1661
|
+
0,
|
|
1662
|
+
0,
|
|
1663
|
+
0,
|
|
1664
|
+
0,
|
|
1665
|
+
0,
|
|
1666
|
+
0,
|
|
1667
|
+
0,
|
|
1668
|
+
0.37,
|
|
1669
|
+
# This fake activation makes the previous token's activation normalize to 8, which
|
|
1670
|
+
# might help address overconfidence in "10" activations for the one-token-at-a-time
|
|
1671
|
+
# scoring prompt. This value and the associated token don't actually appear anywhere
|
|
1672
|
+
# in the prompt.
|
|
1673
|
+
0.45,
|
|
1674
|
+
],
|
|
1675
|
+
),
|
|
1676
|
+
],
|
|
1677
|
+
first_revealed_activation_indices=[],
|
|
1678
|
+
token_index_to_score=18,
|
|
1679
|
+
explanation="instances of the token 'ate' as part of another word",
|
|
1680
|
+
)
|
|
1681
|
+
|
|
1682
|
+
|
|
1683
|
+
@dataclass
|
|
1684
|
+
class AttentionSimulationExample(FastDataclass):
|
|
1685
|
+
token_pair_example_index: int
|
|
1686
|
+
token_pair_coordinates: tuple[int, int]
|
|
1687
|
+
label: int
|
|
1688
|
+
|
|
1689
|
+
|
|
1690
|
+
@dataclass
|
|
1691
|
+
class AttentionTokenPairExample(FastDataclass):
|
|
1692
|
+
tokens: list[str]
|
|
1693
|
+
token_pair_coordinates: list[tuple[int, int]]
|
|
1694
|
+
|
|
1695
|
+
|
|
1696
|
+
@dataclass
|
|
1697
|
+
class AttentionHeadFewShotExample(FastDataclass):
|
|
1698
|
+
token_pair_examples: list[AttentionTokenPairExample]
|
|
1699
|
+
explanation: str
|
|
1700
|
+
simulation_examples: list[AttentionSimulationExample] | None = None
|
|
1701
|
+
|
|
1702
|
+
|
|
1703
|
+
ATTENTION_HEAD_FEW_SHOT_EXAMPLES: list[AttentionHeadFewShotExample] = [
|
|
1704
|
+
# gpt2-xl, layer 1, head 1
|
|
1705
|
+
AttentionHeadFewShotExample(
|
|
1706
|
+
token_pair_examples=[
|
|
1707
|
+
AttentionTokenPairExample(
|
|
1708
|
+
tokens=[
|
|
1709
|
+
" dreams",
|
|
1710
|
+
" of",
|
|
1711
|
+
" a",
|
|
1712
|
+
" future",
|
|
1713
|
+
" like",
|
|
1714
|
+
" her",
|
|
1715
|
+
" biggest",
|
|
1716
|
+
" idol",
|
|
1717
|
+
",",
|
|
1718
|
+
" who",
|
|
1719
|
+
" was",
|
|
1720
|
+
" also",
|
|
1721
|
+
" born",
|
|
1722
|
+
" visually",
|
|
1723
|
+
" impaired",
|
|
1724
|
+
".",
|
|
1725
|
+
"\n",
|
|
1726
|
+
"\n",
|
|
1727
|
+
'"',
|
|
1728
|
+
"My",
|
|
1729
|
+
" ultimate",
|
|
1730
|
+
" dream",
|
|
1731
|
+
" would",
|
|
1732
|
+
" be",
|
|
1733
|
+
" to",
|
|
1734
|
+
" sing",
|
|
1735
|
+
" at",
|
|
1736
|
+
" Carol",
|
|
1737
|
+
"s",
|
|
1738
|
+
" [",
|
|
1739
|
+
"by",
|
|
1740
|
+
" Candle",
|
|
1741
|
+
"light",
|
|
1742
|
+
"]",
|
|
1743
|
+
" and",
|
|
1744
|
+
" to",
|
|
1745
|
+
" become",
|
|
1746
|
+
" a",
|
|
1747
|
+
" famous",
|
|
1748
|
+
" musician",
|
|
1749
|
+
" like",
|
|
1750
|
+
" Andrea",
|
|
1751
|
+
" Bo",
|
|
1752
|
+
"cell",
|
|
1753
|
+
"i",
|
|
1754
|
+
" ...",
|
|
1755
|
+
" and",
|
|
1756
|
+
" to",
|
|
1757
|
+
" show",
|
|
1758
|
+
" people",
|
|
1759
|
+
" that",
|
|
1760
|
+
" if",
|
|
1761
|
+
" you",
|
|
1762
|
+
" have",
|
|
1763
|
+
" a",
|
|
1764
|
+
" disability",
|
|
1765
|
+
" it",
|
|
1766
|
+
" doesn",
|
|
1767
|
+
"'t",
|
|
1768
|
+
" matter",
|
|
1769
|
+
',"',
|
|
1770
|
+
" she",
|
|
1771
|
+
" said",
|
|
1772
|
+
".",
|
|
1773
|
+
],
|
|
1774
|
+
# 45 = "attended from", 33 = "attended to"
|
|
1775
|
+
token_pair_coordinates=[(45, 33)],
|
|
1776
|
+
),
|
|
1777
|
+
AttentionTokenPairExample(
|
|
1778
|
+
tokens=[
|
|
1779
|
+
"omes",
|
|
1780
|
+
" Ever",
|
|
1781
|
+
" Sequ",
|
|
1782
|
+
"enced",
|
|
1783
|
+
"]",
|
|
1784
|
+
"\n",
|
|
1785
|
+
"\n",
|
|
1786
|
+
"One",
|
|
1787
|
+
" mystery",
|
|
1788
|
+
" of",
|
|
1789
|
+
" cat",
|
|
1790
|
+
" development",
|
|
1791
|
+
" is",
|
|
1792
|
+
" how",
|
|
1793
|
+
" cats",
|
|
1794
|
+
" have",
|
|
1795
|
+
" come",
|
|
1796
|
+
" to",
|
|
1797
|
+
" have",
|
|
1798
|
+
" such",
|
|
1799
|
+
" varied",
|
|
1800
|
+
" coats",
|
|
1801
|
+
",",
|
|
1802
|
+
" from",
|
|
1803
|
+
" solid",
|
|
1804
|
+
" colours",
|
|
1805
|
+
" to",
|
|
1806
|
+
' "',
|
|
1807
|
+
"mac",
|
|
1808
|
+
"ke",
|
|
1809
|
+
"rel",
|
|
1810
|
+
'"',
|
|
1811
|
+
" tab",
|
|
1812
|
+
"by",
|
|
1813
|
+
" patterns",
|
|
1814
|
+
" of",
|
|
1815
|
+
" thin",
|
|
1816
|
+
" vertical",
|
|
1817
|
+
" stripes",
|
|
1818
|
+
".",
|
|
1819
|
+
" The",
|
|
1820
|
+
" researchers",
|
|
1821
|
+
" were",
|
|
1822
|
+
" particularly",
|
|
1823
|
+
" interested",
|
|
1824
|
+
" in",
|
|
1825
|
+
" what",
|
|
1826
|
+
" turns",
|
|
1827
|
+
" the",
|
|
1828
|
+
" mac",
|
|
1829
|
+
"ke",
|
|
1830
|
+
"rel",
|
|
1831
|
+
" pattern",
|
|
1832
|
+
" into",
|
|
1833
|
+
" a",
|
|
1834
|
+
' "',
|
|
1835
|
+
"bl",
|
|
1836
|
+
"ot",
|
|
1837
|
+
"ched",
|
|
1838
|
+
'"',
|
|
1839
|
+
" tab",
|
|
1840
|
+
"by",
|
|
1841
|
+
" pattern",
|
|
1842
|
+
",",
|
|
1843
|
+
],
|
|
1844
|
+
token_pair_coordinates=[(5, 4)],
|
|
1845
|
+
),
|
|
1846
|
+
AttentionTokenPairExample(
|
|
1847
|
+
tokens=[
|
|
1848
|
+
",",
|
|
1849
|
+
" 6",
|
|
1850
|
+
",",
|
|
1851
|
+
" 8",
|
|
1852
|
+
",",
|
|
1853
|
+
" 4",
|
|
1854
|
+
"]",
|
|
1855
|
+
"':",
|
|
1856
|
+
"rb",
|
|
1857
|
+
".",
|
|
1858
|
+
"sort",
|
|
1859
|
+
".",
|
|
1860
|
+
"slice",
|
|
1861
|
+
"(",
|
|
1862
|
+
"1",
|
|
1863
|
+
",",
|
|
1864
|
+
"2",
|
|
1865
|
+
");",
|
|
1866
|
+
" #",
|
|
1867
|
+
" More",
|
|
1868
|
+
" advanced",
|
|
1869
|
+
",",
|
|
1870
|
+
" this",
|
|
1871
|
+
" is",
|
|
1872
|
+
" Ruby",
|
|
1873
|
+
"'s",
|
|
1874
|
+
" map",
|
|
1875
|
+
" and",
|
|
1876
|
+
" each",
|
|
1877
|
+
"_",
|
|
1878
|
+
"with",
|
|
1879
|
+
"_",
|
|
1880
|
+
"index",
|
|
1881
|
+
" #",
|
|
1882
|
+
" This",
|
|
1883
|
+
" shows",
|
|
1884
|
+
" the",
|
|
1885
|
+
" :",
|
|
1886
|
+
"rb",
|
|
1887
|
+
" post",
|
|
1888
|
+
"fix",
|
|
1889
|
+
"-",
|
|
1890
|
+
"operator",
|
|
1891
|
+
" sugar",
|
|
1892
|
+
" instead",
|
|
1893
|
+
" of",
|
|
1894
|
+
" EV",
|
|
1895
|
+
"AL",
|
|
1896
|
+
' "[',
|
|
1897
|
+
"1",
|
|
1898
|
+
",",
|
|
1899
|
+
"2",
|
|
1900
|
+
",",
|
|
1901
|
+
"3",
|
|
1902
|
+
",",
|
|
1903
|
+
"4",
|
|
1904
|
+
"]",
|
|
1905
|
+
'":',
|
|
1906
|
+
"rb",
|
|
1907
|
+
" .",
|
|
1908
|
+
"map",
|
|
1909
|
+
"(",
|
|
1910
|
+
"->",
|
|
1911
|
+
" $",
|
|
1912
|
+
],
|
|
1913
|
+
token_pair_coordinates=[(7, 6)],
|
|
1914
|
+
),
|
|
1915
|
+
AttentionTokenPairExample(
|
|
1916
|
+
tokens=[
|
|
1917
|
+
" him",
|
|
1918
|
+
" a",
|
|
1919
|
+
" W",
|
|
1920
|
+
"N",
|
|
1921
|
+
" [",
|
|
1922
|
+
"white",
|
|
1923
|
+
" nationalist",
|
|
1924
|
+
"]",
|
|
1925
|
+
" until",
|
|
1926
|
+
" there",
|
|
1927
|
+
" is",
|
|
1928
|
+
" an",
|
|
1929
|
+
" indication",
|
|
1930
|
+
" as",
|
|
1931
|
+
" such",
|
|
1932
|
+
"...",
|
|
1933
|
+
" The",
|
|
1934
|
+
" fact",
|
|
1935
|
+
" that",
|
|
1936
|
+
" he",
|
|
1937
|
+
" targeted",
|
|
1938
|
+
" a",
|
|
1939
|
+
" church",
|
|
1940
|
+
" gives",
|
|
1941
|
+
" me",
|
|
1942
|
+
" an",
|
|
1943
|
+
" ink",
|
|
1944
|
+
"ling",
|
|
1945
|
+
" that",
|
|
1946
|
+
" it",
|
|
1947
|
+
" was",
|
|
1948
|
+
" religion",
|
|
1949
|
+
"-",
|
|
1950
|
+
"related",
|
|
1951
|
+
',"',
|
|
1952
|
+
" wrote",
|
|
1953
|
+
" White",
|
|
1954
|
+
"Virgin",
|
|
1955
|
+
"ian",
|
|
1956
|
+
".",
|
|
1957
|
+
"\n",
|
|
1958
|
+
"\n",
|
|
1959
|
+
'"',
|
|
1960
|
+
"Yep",
|
|
1961
|
+
",",
|
|
1962
|
+
" bad",
|
|
1963
|
+
" news",
|
|
1964
|
+
" for",
|
|
1965
|
+
" gun",
|
|
1966
|
+
" rights",
|
|
1967
|
+
" advocates",
|
|
1968
|
+
" as",
|
|
1969
|
+
" well",
|
|
1970
|
+
',"',
|
|
1971
|
+
" wrote",
|
|
1972
|
+
" math",
|
|
1973
|
+
"the",
|
|
1974
|
+
"ory",
|
|
1975
|
+
"l",
|
|
1976
|
+
"over",
|
|
1977
|
+
"2008",
|
|
1978
|
+
".",
|
|
1979
|
+
' "',
|
|
1980
|
+
"Another",
|
|
1981
|
+
],
|
|
1982
|
+
token_pair_coordinates=[(15, 7)],
|
|
1983
|
+
),
|
|
1984
|
+
AttentionTokenPairExample(
|
|
1985
|
+
tokens=[
|
|
1986
|
+
"23",
|
|
1987
|
+
"]",
|
|
1988
|
+
"\n",
|
|
1989
|
+
"\n",
|
|
1990
|
+
"While",
|
|
1991
|
+
" preparing",
|
|
1992
|
+
" to",
|
|
1993
|
+
" take",
|
|
1994
|
+
" the",
|
|
1995
|
+
" fight",
|
|
1996
|
+
" to",
|
|
1997
|
+
" Prim",
|
|
1998
|
+
"ord",
|
|
1999
|
+
"us",
|
|
2000
|
+
",",
|
|
2001
|
+
" B",
|
|
2002
|
+
"alth",
|
|
2003
|
+
"azar",
|
|
2004
|
+
" learned",
|
|
2005
|
+
" about",
|
|
2006
|
+
" T",
|
|
2007
|
+
"aim",
|
|
2008
|
+
"i",
|
|
2009
|
+
"'s",
|
|
2010
|
+
" machine",
|
|
2011
|
+
" and",
|
|
2012
|
+
" how",
|
|
2013
|
+
" it",
|
|
2014
|
+
" could",
|
|
2015
|
+
" supposedly",
|
|
2016
|
+
" kill",
|
|
2017
|
+
" two",
|
|
2018
|
+
" Elder",
|
|
2019
|
+
" Dragons",
|
|
2020
|
+
" with",
|
|
2021
|
+
" a",
|
|
2022
|
+
" single",
|
|
2023
|
+
" blow",
|
|
2024
|
+
",",
|
|
2025
|
+
" which",
|
|
2026
|
+
" p",
|
|
2027
|
+
"iqu",
|
|
2028
|
+
"ed",
|
|
2029
|
+
" his",
|
|
2030
|
+
" interest",
|
|
2031
|
+
".",
|
|
2032
|
+
" This",
|
|
2033
|
+
" piece",
|
|
2034
|
+
" of",
|
|
2035
|
+
" news",
|
|
2036
|
+
",",
|
|
2037
|
+
" as",
|
|
2038
|
+
" well",
|
|
2039
|
+
" as",
|
|
2040
|
+
" Mar",
|
|
2041
|
+
"j",
|
|
2042
|
+
"ory",
|
|
2043
|
+
"'s",
|
|
2044
|
+
" sudden",
|
|
2045
|
+
" departure",
|
|
2046
|
+
" from",
|
|
2047
|
+
" his",
|
|
2048
|
+
" side",
|
|
2049
|
+
" which",
|
|
2050
|
+
],
|
|
2051
|
+
token_pair_coordinates=[(3, 1)],
|
|
2052
|
+
),
|
|
2053
|
+
],
|
|
2054
|
+
explanation="attends to the latest closing square bracket from arbitrary subsequent tokens",
|
|
2055
|
+
),
|
|
2056
|
+
# gpt2-xl, layer 2, head 8
|
|
2057
|
+
AttentionHeadFewShotExample(
|
|
2058
|
+
simulation_examples=[
|
|
2059
|
+
AttentionSimulationExample(
|
|
2060
|
+
token_pair_example_index=0,
|
|
2061
|
+
token_pair_coordinates=(63, 15),
|
|
2062
|
+
label=0,
|
|
2063
|
+
),
|
|
2064
|
+
AttentionSimulationExample(
|
|
2065
|
+
token_pair_example_index=0,
|
|
2066
|
+
token_pair_coordinates=(50, 15),
|
|
2067
|
+
label=1,
|
|
2068
|
+
),
|
|
2069
|
+
],
|
|
2070
|
+
token_pair_examples=[
|
|
2071
|
+
AttentionTokenPairExample(
|
|
2072
|
+
tokens=[
|
|
2073
|
+
" he",
|
|
2074
|
+
" said",
|
|
2075
|
+
".",
|
|
2076
|
+
' "',
|
|
2077
|
+
"Coming",
|
|
2078
|
+
" off",
|
|
2079
|
+
" winning",
|
|
2080
|
+
" the",
|
|
2081
|
+
" year",
|
|
2082
|
+
" before",
|
|
2083
|
+
",",
|
|
2084
|
+
" I",
|
|
2085
|
+
" love",
|
|
2086
|
+
" playing",
|
|
2087
|
+
" links",
|
|
2088
|
+
" golf",
|
|
2089
|
+
",",
|
|
2090
|
+
" and",
|
|
2091
|
+
" I",
|
|
2092
|
+
" love",
|
|
2093
|
+
" playing",
|
|
2094
|
+
" the",
|
|
2095
|
+
" week",
|
|
2096
|
+
" before",
|
|
2097
|
+
" a",
|
|
2098
|
+
" major",
|
|
2099
|
+
".",
|
|
2100
|
+
" It",
|
|
2101
|
+
" was",
|
|
2102
|
+
" tough",
|
|
2103
|
+
" to",
|
|
2104
|
+
" miss",
|
|
2105
|
+
" it",
|
|
2106
|
+
".",
|
|
2107
|
+
" I",
|
|
2108
|
+
"'m",
|
|
2109
|
+
" just",
|
|
2110
|
+
" glad",
|
|
2111
|
+
" to",
|
|
2112
|
+
" be",
|
|
2113
|
+
" back",
|
|
2114
|
+
'."',
|
|
2115
|
+
"\n",
|
|
2116
|
+
"\n",
|
|
2117
|
+
"F",
|
|
2118
|
+
"owler",
|
|
2119
|
+
" out",
|
|
2120
|
+
"played",
|
|
2121
|
+
" his",
|
|
2122
|
+
" partners",
|
|
2123
|
+
" Rory",
|
|
2124
|
+
" Mc",
|
|
2125
|
+
"Il",
|
|
2126
|
+
"roy",
|
|
2127
|
+
" (",
|
|
2128
|
+
"74",
|
|
2129
|
+
")",
|
|
2130
|
+
" and",
|
|
2131
|
+
" Hen",
|
|
2132
|
+
"rik",
|
|
2133
|
+
" St",
|
|
2134
|
+
"enson",
|
|
2135
|
+
" (",
|
|
2136
|
+
"72",
|
|
2137
|
+
],
|
|
2138
|
+
token_pair_coordinates=[(50, 15)],
|
|
2139
|
+
),
|
|
2140
|
+
AttentionTokenPairExample(
|
|
2141
|
+
tokens=[
|
|
2142
|
+
" Club",
|
|
2143
|
+
":",
|
|
2144
|
+
"\n",
|
|
2145
|
+
"\n",
|
|
2146
|
+
"1",
|
|
2147
|
+
".",
|
|
2148
|
+
" World",
|
|
2149
|
+
" renowned",
|
|
2150
|
+
" golf",
|
|
2151
|
+
" course",
|
|
2152
|
+
"\n",
|
|
2153
|
+
"\n",
|
|
2154
|
+
"2",
|
|
2155
|
+
".",
|
|
2156
|
+
" Vern",
|
|
2157
|
+
" Mor",
|
|
2158
|
+
"com",
|
|
2159
|
+
" designed",
|
|
2160
|
+
" golf",
|
|
2161
|
+
" course",
|
|
2162
|
+
"\n",
|
|
2163
|
+
"\n",
|
|
2164
|
+
"3",
|
|
2165
|
+
".",
|
|
2166
|
+
" Great",
|
|
2167
|
+
" family",
|
|
2168
|
+
" holiday",
|
|
2169
|
+
" destination",
|
|
2170
|
+
"\n",
|
|
2171
|
+
"\n",
|
|
2172
|
+
"4",
|
|
2173
|
+
".",
|
|
2174
|
+
" Play",
|
|
2175
|
+
" amid",
|
|
2176
|
+
" our",
|
|
2177
|
+
" resident",
|
|
2178
|
+
" Eastern",
|
|
2179
|
+
" Grey",
|
|
2180
|
+
" k",
|
|
2181
|
+
"ang",
|
|
2182
|
+
"aroo",
|
|
2183
|
+
" population",
|
|
2184
|
+
"\n",
|
|
2185
|
+
"\n",
|
|
2186
|
+
"5",
|
|
2187
|
+
".",
|
|
2188
|
+
" Terr",
|
|
2189
|
+
"ific",
|
|
2190
|
+
" friendly",
|
|
2191
|
+
" staff",
|
|
2192
|
+
"\n",
|
|
2193
|
+
"\n",
|
|
2194
|
+
"6",
|
|
2195
|
+
".",
|
|
2196
|
+
" Natural",
|
|
2197
|
+
" pictures",
|
|
2198
|
+
"que",
|
|
2199
|
+
" bush",
|
|
2200
|
+
" setting",
|
|
2201
|
+
"\n",
|
|
2202
|
+
"\n",
|
|
2203
|
+
"7",
|
|
2204
|
+
".",
|
|
2205
|
+
" L",
|
|
2206
|
+
],
|
|
2207
|
+
token_pair_coordinates=[(9, 8)],
|
|
2208
|
+
),
|
|
2209
|
+
AttentionTokenPairExample(
|
|
2210
|
+
tokens=[
|
|
2211
|
+
"615",
|
|
2212
|
+
" rpm",
|
|
2213
|
+
" on",
|
|
2214
|
+
" average",
|
|
2215
|
+
").",
|
|
2216
|
+
" As",
|
|
2217
|
+
" a",
|
|
2218
|
+
" result",
|
|
2219
|
+
",",
|
|
2220
|
+
" each",
|
|
2221
|
+
" player",
|
|
2222
|
+
" was",
|
|
2223
|
+
" hitting",
|
|
2224
|
+
" longer",
|
|
2225
|
+
" drives",
|
|
2226
|
+
" on",
|
|
2227
|
+
" their",
|
|
2228
|
+
" best",
|
|
2229
|
+
" shots",
|
|
2230
|
+
",",
|
|
2231
|
+
" while",
|
|
2232
|
+
" achieving",
|
|
2233
|
+
" a",
|
|
2234
|
+
" stra",
|
|
2235
|
+
"ighter",
|
|
2236
|
+
" ball",
|
|
2237
|
+
" flight",
|
|
2238
|
+
" that",
|
|
2239
|
+
" was",
|
|
2240
|
+
" less",
|
|
2241
|
+
" affected",
|
|
2242
|
+
" by",
|
|
2243
|
+
" wind",
|
|
2244
|
+
".",
|
|
2245
|
+
"\n",
|
|
2246
|
+
"\n",
|
|
2247
|
+
"Every",
|
|
2248
|
+
" Golf",
|
|
2249
|
+
"WR",
|
|
2250
|
+
"X",
|
|
2251
|
+
" Member",
|
|
2252
|
+
" gained",
|
|
2253
|
+
" yard",
|
|
2254
|
+
"age",
|
|
2255
|
+
" with",
|
|
2256
|
+
" a",
|
|
2257
|
+
" new",
|
|
2258
|
+
" Taylor",
|
|
2259
|
+
"Made",
|
|
2260
|
+
" driver",
|
|
2261
|
+
";",
|
|
2262
|
+
" the",
|
|
2263
|
+
" largest",
|
|
2264
|
+
" distance",
|
|
2265
|
+
" gain",
|
|
2266
|
+
" was",
|
|
2267
|
+
" an",
|
|
2268
|
+
" impressive",
|
|
2269
|
+
" +",
|
|
2270
|
+
"10",
|
|
2271
|
+
".",
|
|
2272
|
+
"1",
|
|
2273
|
+
" yards",
|
|
2274
|
+
",",
|
|
2275
|
+
],
|
|
2276
|
+
token_pair_coordinates=[(47, 37)],
|
|
2277
|
+
),
|
|
2278
|
+
AttentionTokenPairExample(
|
|
2279
|
+
tokens=[
|
|
2280
|
+
" of",
|
|
2281
|
+
" being",
|
|
2282
|
+
"?",
|
|
2283
|
+
" Well",
|
|
2284
|
+
",",
|
|
2285
|
+
" having",
|
|
2286
|
+
" perfected",
|
|
2287
|
+
" the",
|
|
2288
|
+
" art",
|
|
2289
|
+
" of",
|
|
2290
|
+
" swimming",
|
|
2291
|
+
",",
|
|
2292
|
+
" Phelps",
|
|
2293
|
+
" has",
|
|
2294
|
+
" moved",
|
|
2295
|
+
" on",
|
|
2296
|
+
" to",
|
|
2297
|
+
" another",
|
|
2298
|
+
" cherished",
|
|
2299
|
+
" summer",
|
|
2300
|
+
" past",
|
|
2301
|
+
"ime",
|
|
2302
|
+
" –",
|
|
2303
|
+
" golf",
|
|
2304
|
+
".",
|
|
2305
|
+
" Here",
|
|
2306
|
+
" he",
|
|
2307
|
+
" is",
|
|
2308
|
+
" participating",
|
|
2309
|
+
" in",
|
|
2310
|
+
" the",
|
|
2311
|
+
" Dun",
|
|
2312
|
+
"hill",
|
|
2313
|
+
" Links",
|
|
2314
|
+
" Championship",
|
|
2315
|
+
" at",
|
|
2316
|
+
" Kings",
|
|
2317
|
+
"b",
|
|
2318
|
+
"arn",
|
|
2319
|
+
"s",
|
|
2320
|
+
" in",
|
|
2321
|
+
" Scotland",
|
|
2322
|
+
" today",
|
|
2323
|
+
".",
|
|
2324
|
+
" The",
|
|
2325
|
+
" greens",
|
|
2326
|
+
" over",
|
|
2327
|
+
" there",
|
|
2328
|
+
" are",
|
|
2329
|
+
" really",
|
|
2330
|
+
" big",
|
|
2331
|
+
",",
|
|
2332
|
+
" so",
|
|
2333
|
+
" the",
|
|
2334
|
+
" opportunity",
|
|
2335
|
+
" for",
|
|
2336
|
+
" 50",
|
|
2337
|
+
"-",
|
|
2338
|
+
"yard",
|
|
2339
|
+
" put",
|
|
2340
|
+
"ts",
|
|
2341
|
+
" exist",
|
|
2342
|
+
".",
|
|
2343
|
+
" Of",
|
|
2344
|
+
],
|
|
2345
|
+
token_pair_coordinates=[(45, 23)],
|
|
2346
|
+
),
|
|
2347
|
+
AttentionTokenPairExample(
|
|
2348
|
+
tokens=[
|
|
2349
|
+
"OTUS",
|
|
2350
|
+
" is",
|
|
2351
|
+
" getting",
|
|
2352
|
+
" to",
|
|
2353
|
+
" see",
|
|
2354
|
+
" aliens",
|
|
2355
|
+
".",
|
|
2356
|
+
"\n",
|
|
2357
|
+
"\n",
|
|
2358
|
+
"RELATED",
|
|
2359
|
+
":",
|
|
2360
|
+
" Barack",
|
|
2361
|
+
" Obama",
|
|
2362
|
+
" joins",
|
|
2363
|
+
" second",
|
|
2364
|
+
" D",
|
|
2365
|
+
".",
|
|
2366
|
+
"C",
|
|
2367
|
+
".-",
|
|
2368
|
+
"area",
|
|
2369
|
+
" golf",
|
|
2370
|
+
" club",
|
|
2371
|
+
"\n",
|
|
2372
|
+
"\n",
|
|
2373
|
+
'"',
|
|
2374
|
+
"He",
|
|
2375
|
+
" goes",
|
|
2376
|
+
",",
|
|
2377
|
+
" '",
|
|
2378
|
+
"they",
|
|
2379
|
+
"'re",
|
|
2380
|
+
" freaking",
|
|
2381
|
+
" crazy",
|
|
2382
|
+
" looking",
|
|
2383
|
+
".'",
|
|
2384
|
+
" And",
|
|
2385
|
+
" then",
|
|
2386
|
+
" he",
|
|
2387
|
+
" walks",
|
|
2388
|
+
" up",
|
|
2389
|
+
",",
|
|
2390
|
+
" makes",
|
|
2391
|
+
" his",
|
|
2392
|
+
" put",
|
|
2393
|
+
"t",
|
|
2394
|
+
",",
|
|
2395
|
+
" turns",
|
|
2396
|
+
" back",
|
|
2397
|
+
",",
|
|
2398
|
+
" walks",
|
|
2399
|
+
" off",
|
|
2400
|
+
" the",
|
|
2401
|
+
" green",
|
|
2402
|
+
",",
|
|
2403
|
+
" leaves",
|
|
2404
|
+
" it",
|
|
2405
|
+
" at",
|
|
2406
|
+
" that",
|
|
2407
|
+
" and",
|
|
2408
|
+
" gives",
|
|
2409
|
+
" me",
|
|
2410
|
+
" a",
|
|
2411
|
+
" wink",
|
|
2412
|
+
',"',
|
|
2413
|
+
],
|
|
2414
|
+
token_pair_coordinates=[(52, 20)],
|
|
2415
|
+
),
|
|
2416
|
+
],
|
|
2417
|
+
explanation='attends to the token "golf" from golf-related tokens',
|
|
2418
|
+
),
|
|
2419
|
+
# gpt2-xl, layer 1, head 10
|
|
2420
|
+
AttentionHeadFewShotExample(
|
|
2421
|
+
simulation_examples=[
|
|
2422
|
+
AttentionSimulationExample(
|
|
2423
|
+
token_pair_example_index=0,
|
|
2424
|
+
token_pair_coordinates=(37, 36),
|
|
2425
|
+
label=0,
|
|
2426
|
+
),
|
|
2427
|
+
AttentionSimulationExample(
|
|
2428
|
+
token_pair_example_index=0,
|
|
2429
|
+
token_pair_coordinates=(14, 12),
|
|
2430
|
+
label=1,
|
|
2431
|
+
),
|
|
2432
|
+
],
|
|
2433
|
+
token_pair_examples=[
|
|
2434
|
+
AttentionTokenPairExample(
|
|
2435
|
+
tokens=[
|
|
2436
|
+
" security",
|
|
2437
|
+
" by",
|
|
2438
|
+
" requiring",
|
|
2439
|
+
" the",
|
|
2440
|
+
" user",
|
|
2441
|
+
" to",
|
|
2442
|
+
" enter",
|
|
2443
|
+
" a",
|
|
2444
|
+
" numeric",
|
|
2445
|
+
" code",
|
|
2446
|
+
" sent",
|
|
2447
|
+
" to",
|
|
2448
|
+
" his",
|
|
2449
|
+
" or",
|
|
2450
|
+
" her",
|
|
2451
|
+
" cellphone",
|
|
2452
|
+
" in",
|
|
2453
|
+
" addition",
|
|
2454
|
+
" to",
|
|
2455
|
+
" a",
|
|
2456
|
+
" password",
|
|
2457
|
+
".",
|
|
2458
|
+
" A",
|
|
2459
|
+
" lot",
|
|
2460
|
+
" of",
|
|
2461
|
+
" websites",
|
|
2462
|
+
" have",
|
|
2463
|
+
" offered",
|
|
2464
|
+
" this",
|
|
2465
|
+
" feature",
|
|
2466
|
+
" for",
|
|
2467
|
+
" years",
|
|
2468
|
+
",",
|
|
2469
|
+
" but",
|
|
2470
|
+
" Int",
|
|
2471
|
+
"uit",
|
|
2472
|
+
" just",
|
|
2473
|
+
" made",
|
|
2474
|
+
" it",
|
|
2475
|
+
" widely",
|
|
2476
|
+
" available",
|
|
2477
|
+
" earlier",
|
|
2478
|
+
" this",
|
|
2479
|
+
" year",
|
|
2480
|
+
".",
|
|
2481
|
+
"\n",
|
|
2482
|
+
"\n",
|
|
2483
|
+
'"',
|
|
2484
|
+
"When",
|
|
2485
|
+
" you",
|
|
2486
|
+
" give",
|
|
2487
|
+
" your",
|
|
2488
|
+
" most",
|
|
2489
|
+
" sensitive",
|
|
2490
|
+
" data",
|
|
2491
|
+
" and",
|
|
2492
|
+
" that",
|
|
2493
|
+
" of",
|
|
2494
|
+
" your",
|
|
2495
|
+
" family",
|
|
2496
|
+
" to",
|
|
2497
|
+
" a",
|
|
2498
|
+
" company",
|
|
2499
|
+
",",
|
|
2500
|
+
],
|
|
2501
|
+
token_pair_coordinates=[(14, 12)],
|
|
2502
|
+
),
|
|
2503
|
+
AttentionTokenPairExample(
|
|
2504
|
+
tokens=[
|
|
2505
|
+
" 3",
|
|
2506
|
+
" months",
|
|
2507
|
+
",",
|
|
2508
|
+
" they",
|
|
2509
|
+
" separated",
|
|
2510
|
+
" the",
|
|
2511
|
+
" men",
|
|
2512
|
+
" and",
|
|
2513
|
+
" women",
|
|
2514
|
+
" here",
|
|
2515
|
+
".",
|
|
2516
|
+
" I",
|
|
2517
|
+
" don",
|
|
2518
|
+
"'t",
|
|
2519
|
+
" know",
|
|
2520
|
+
" where",
|
|
2521
|
+
" they",
|
|
2522
|
+
" took",
|
|
2523
|
+
" the",
|
|
2524
|
+
" men",
|
|
2525
|
+
" and",
|
|
2526
|
+
" the",
|
|
2527
|
+
" children",
|
|
2528
|
+
",",
|
|
2529
|
+
" but",
|
|
2530
|
+
" they",
|
|
2531
|
+
" took",
|
|
2532
|
+
" us",
|
|
2533
|
+
" women",
|
|
2534
|
+
" to",
|
|
2535
|
+
" Syria",
|
|
2536
|
+
".",
|
|
2537
|
+
" They",
|
|
2538
|
+
" kept",
|
|
2539
|
+
" us",
|
|
2540
|
+
" in",
|
|
2541
|
+
" an",
|
|
2542
|
+
" underground",
|
|
2543
|
+
" prison",
|
|
2544
|
+
".",
|
|
2545
|
+
" My",
|
|
2546
|
+
" only",
|
|
2547
|
+
" wish",
|
|
2548
|
+
" is",
|
|
2549
|
+
" that",
|
|
2550
|
+
" my",
|
|
2551
|
+
" children",
|
|
2552
|
+
" and",
|
|
2553
|
+
" husband",
|
|
2554
|
+
" escape",
|
|
2555
|
+
" ISIS",
|
|
2556
|
+
".",
|
|
2557
|
+
" They",
|
|
2558
|
+
" brought",
|
|
2559
|
+
" us",
|
|
2560
|
+
" here",
|
|
2561
|
+
" from",
|
|
2562
|
+
" Raqqa",
|
|
2563
|
+
",",
|
|
2564
|
+
" my",
|
|
2565
|
+
" sisters",
|
|
2566
|
+
" from",
|
|
2567
|
+
" the",
|
|
2568
|
+
" PKK",
|
|
2569
|
+
],
|
|
2570
|
+
token_pair_coordinates=[(8, 6)],
|
|
2571
|
+
),
|
|
2572
|
+
AttentionTokenPairExample(
|
|
2573
|
+
tokens=[
|
|
2574
|
+
" an",
|
|
2575
|
+
" emphasis",
|
|
2576
|
+
" on",
|
|
2577
|
+
" the",
|
|
2578
|
+
" pursuit",
|
|
2579
|
+
" of",
|
|
2580
|
+
" power",
|
|
2581
|
+
" despite",
|
|
2582
|
+
" interpersonal",
|
|
2583
|
+
" costs",
|
|
2584
|
+
'."',
|
|
2585
|
+
"\n",
|
|
2586
|
+
"\n",
|
|
2587
|
+
"The",
|
|
2588
|
+
" study",
|
|
2589
|
+
",",
|
|
2590
|
+
" which",
|
|
2591
|
+
" involved",
|
|
2592
|
+
" over",
|
|
2593
|
+
" 600",
|
|
2594
|
+
" young",
|
|
2595
|
+
" men",
|
|
2596
|
+
" and",
|
|
2597
|
+
" women",
|
|
2598
|
+
",",
|
|
2599
|
+
" makes",
|
|
2600
|
+
" a",
|
|
2601
|
+
" strong",
|
|
2602
|
+
" case",
|
|
2603
|
+
" for",
|
|
2604
|
+
" assessing",
|
|
2605
|
+
" such",
|
|
2606
|
+
" traits",
|
|
2607
|
+
" as",
|
|
2608
|
+
' "',
|
|
2609
|
+
"r",
|
|
2610
|
+
"uth",
|
|
2611
|
+
"less",
|
|
2612
|
+
" ambition",
|
|
2613
|
+
',"',
|
|
2614
|
+
' "',
|
|
2615
|
+
"dis",
|
|
2616
|
+
"comfort",
|
|
2617
|
+
" with",
|
|
2618
|
+
" leadership",
|
|
2619
|
+
'"',
|
|
2620
|
+
" and",
|
|
2621
|
+
' "',
|
|
2622
|
+
"hub",
|
|
2623
|
+
"rist",
|
|
2624
|
+
"ic",
|
|
2625
|
+
" pride",
|
|
2626
|
+
'"',
|
|
2627
|
+
" to",
|
|
2628
|
+
" understand",
|
|
2629
|
+
" psychopath",
|
|
2630
|
+
"ologies",
|
|
2631
|
+
".",
|
|
2632
|
+
"\n",
|
|
2633
|
+
"\n",
|
|
2634
|
+
"The",
|
|
2635
|
+
" researchers",
|
|
2636
|
+
" looked",
|
|
2637
|
+
" at",
|
|
2638
|
+
],
|
|
2639
|
+
token_pair_coordinates=[(23, 21)],
|
|
2640
|
+
),
|
|
2641
|
+
AttentionTokenPairExample(
|
|
2642
|
+
tokens=[
|
|
2643
|
+
" 4",
|
|
2644
|
+
" hours",
|
|
2645
|
+
".",
|
|
2646
|
+
" These",
|
|
2647
|
+
" results",
|
|
2648
|
+
",",
|
|
2649
|
+
" differ",
|
|
2650
|
+
" between",
|
|
2651
|
+
" men",
|
|
2652
|
+
" and",
|
|
2653
|
+
" women",
|
|
2654
|
+
",",
|
|
2655
|
+
" however",
|
|
2656
|
+
".",
|
|
2657
|
+
" We",
|
|
2658
|
+
" can",
|
|
2659
|
+
" see",
|
|
2660
|
+
" that",
|
|
2661
|
+
" although",
|
|
2662
|
+
" both",
|
|
2663
|
+
" groups",
|
|
2664
|
+
" have",
|
|
2665
|
+
" a",
|
|
2666
|
+
" large",
|
|
2667
|
+
" cluster",
|
|
2668
|
+
" of",
|
|
2669
|
+
" people",
|
|
2670
|
+
" at",
|
|
2671
|
+
" exactly",
|
|
2672
|
+
" 40",
|
|
2673
|
+
" hours",
|
|
2674
|
+
" per",
|
|
2675
|
+
" week",
|
|
2676
|
+
",",
|
|
2677
|
+
" there",
|
|
2678
|
+
" are",
|
|
2679
|
+
" more",
|
|
2680
|
+
" men",
|
|
2681
|
+
" reporting",
|
|
2682
|
+
" hours",
|
|
2683
|
+
" above",
|
|
2684
|
+
" 40",
|
|
2685
|
+
",",
|
|
2686
|
+
" whereas",
|
|
2687
|
+
" there",
|
|
2688
|
+
" are",
|
|
2689
|
+
" more",
|
|
2690
|
+
" women",
|
|
2691
|
+
" reporting",
|
|
2692
|
+
" hours",
|
|
2693
|
+
" below",
|
|
2694
|
+
" 40",
|
|
2695
|
+
".",
|
|
2696
|
+
" Result",
|
|
2697
|
+
" 3",
|
|
2698
|
+
":",
|
|
2699
|
+
" Male",
|
|
2700
|
+
" Hours",
|
|
2701
|
+
" Work",
|
|
2702
|
+
"ed",
|
|
2703
|
+
" [",
|
|
2704
|
+
"Info",
|
|
2705
|
+
"]",
|
|
2706
|
+
" Owner",
|
|
2707
|
+
],
|
|
2708
|
+
token_pair_coordinates=[(10, 8)],
|
|
2709
|
+
),
|
|
2710
|
+
AttentionTokenPairExample(
|
|
2711
|
+
tokens=[
|
|
2712
|
+
" they",
|
|
2713
|
+
" were",
|
|
2714
|
+
" perceived",
|
|
2715
|
+
" as",
|
|
2716
|
+
" more",
|
|
2717
|
+
" emotional",
|
|
2718
|
+
",",
|
|
2719
|
+
" which",
|
|
2720
|
+
" made",
|
|
2721
|
+
" participants",
|
|
2722
|
+
" more",
|
|
2723
|
+
" confident",
|
|
2724
|
+
" in",
|
|
2725
|
+
" their",
|
|
2726
|
+
" own",
|
|
2727
|
+
" opinion",
|
|
2728
|
+
'."',
|
|
2729
|
+
"\n",
|
|
2730
|
+
"\n",
|
|
2731
|
+
"Ms",
|
|
2732
|
+
" Sal",
|
|
2733
|
+
"erno",
|
|
2734
|
+
" said",
|
|
2735
|
+
" both",
|
|
2736
|
+
" men",
|
|
2737
|
+
" and",
|
|
2738
|
+
" women",
|
|
2739
|
+
" reacted",
|
|
2740
|
+
" in",
|
|
2741
|
+
" the",
|
|
2742
|
+
" same",
|
|
2743
|
+
" way",
|
|
2744
|
+
" to",
|
|
2745
|
+
" women",
|
|
2746
|
+
" expressing",
|
|
2747
|
+
" themselves",
|
|
2748
|
+
" angrily",
|
|
2749
|
+
".",
|
|
2750
|
+
"\n",
|
|
2751
|
+
"\n",
|
|
2752
|
+
'"',
|
|
2753
|
+
"Particip",
|
|
2754
|
+
"ants",
|
|
2755
|
+
" confidence",
|
|
2756
|
+
" in",
|
|
2757
|
+
" their",
|
|
2758
|
+
" own",
|
|
2759
|
+
" verdict",
|
|
2760
|
+
" dropped",
|
|
2761
|
+
" significantly",
|
|
2762
|
+
" after",
|
|
2763
|
+
" male",
|
|
2764
|
+
" hold",
|
|
2765
|
+
"outs",
|
|
2766
|
+
" expressed",
|
|
2767
|
+
" anger",
|
|
2768
|
+
',"',
|
|
2769
|
+
" the",
|
|
2770
|
+
" paper",
|
|
2771
|
+
"'s",
|
|
2772
|
+
" findings",
|
|
2773
|
+
" stated",
|
|
2774
|
+
".",
|
|
2775
|
+
"\n",
|
|
2776
|
+
],
|
|
2777
|
+
token_pair_coordinates=[(26, 24)],
|
|
2778
|
+
),
|
|
2779
|
+
],
|
|
2780
|
+
explanation="attends to male-related tokens from paired female-related tokens",
|
|
2781
|
+
),
|
|
2782
|
+
# gpt2-xl, layer 1, head 3
|
|
2783
|
+
AttentionHeadFewShotExample(
|
|
2784
|
+
token_pair_examples=[
|
|
2785
|
+
AttentionTokenPairExample(
|
|
2786
|
+
tokens=[
|
|
2787
|
+
"Viet" "namese",
|
|
2788
|
+
" Ministry",
|
|
2789
|
+
" of",
|
|
2790
|
+
" Foreign",
|
|
2791
|
+
" Affairs",
|
|
2792
|
+
" spokesperson",
|
|
2793
|
+
" Le",
|
|
2794
|
+
" Hai",
|
|
2795
|
+
" Bin",
|
|
2796
|
+
"h",
|
|
2797
|
+
" is",
|
|
2798
|
+
" seen",
|
|
2799
|
+
" in",
|
|
2800
|
+
" this",
|
|
2801
|
+
" file",
|
|
2802
|
+
" photo",
|
|
2803
|
+
".",
|
|
2804
|
+
" .",
|
|
2805
|
+
" Tu",
|
|
2806
|
+
"oi",
|
|
2807
|
+
" Tre",
|
|
2808
|
+
"\n",
|
|
2809
|
+
"\n",
|
|
2810
|
+
"The",
|
|
2811
|
+
" Ministry",
|
|
2812
|
+
" of",
|
|
2813
|
+
" Foreign",
|
|
2814
|
+
" Affairs",
|
|
2815
|
+
" has",
|
|
2816
|
+
" ordered",
|
|
2817
|
+
" a",
|
|
2818
|
+
" thorough",
|
|
2819
|
+
" investigation",
|
|
2820
|
+
" into",
|
|
2821
|
+
" a",
|
|
2822
|
+
" case",
|
|
2823
|
+
" in",
|
|
2824
|
+
" which",
|
|
2825
|
+
" a",
|
|
2826
|
+
" Vietnamese",
|
|
2827
|
+
" fisherman",
|
|
2828
|
+
" was",
|
|
2829
|
+
" shot",
|
|
2830
|
+
" dead",
|
|
2831
|
+
" on",
|
|
2832
|
+
" his",
|
|
2833
|
+
" boat",
|
|
2834
|
+
" in",
|
|
2835
|
+
" Vietnam",
|
|
2836
|
+
"'s",
|
|
2837
|
+
" Tru",
|
|
2838
|
+
"ong",
|
|
2839
|
+
" Sa",
|
|
2840
|
+
" (",
|
|
2841
|
+
"Spr",
|
|
2842
|
+
"at",
|
|
2843
|
+
],
|
|
2844
|
+
token_pair_coordinates=[(1, 1)],
|
|
2845
|
+
),
|
|
2846
|
+
AttentionTokenPairExample(
|
|
2847
|
+
tokens=[
|
|
2848
|
+
" J",
|
|
2849
|
+
"okin",
|
|
2850
|
+
"en",
|
|
2851
|
+
" tells",
|
|
2852
|
+
" a",
|
|
2853
|
+
" much",
|
|
2854
|
+
" different",
|
|
2855
|
+
" story",
|
|
2856
|
+
".",
|
|
2857
|
+
" He",
|
|
2858
|
+
" almost",
|
|
2859
|
+
" sounded",
|
|
2860
|
+
" like",
|
|
2861
|
+
" a",
|
|
2862
|
+
" pitch",
|
|
2863
|
+
"man",
|
|
2864
|
+
".",
|
|
2865
|
+
"\n",
|
|
2866
|
+
"\n",
|
|
2867
|
+
'"',
|
|
2868
|
+
"All",
|
|
2869
|
+
" the",
|
|
2870
|
+
" staff",
|
|
2871
|
+
",",
|
|
2872
|
+
" team",
|
|
2873
|
+
" service",
|
|
2874
|
+
" guys",
|
|
2875
|
+
",",
|
|
2876
|
+
" all",
|
|
2877
|
+
" the",
|
|
2878
|
+
" trainers",
|
|
2879
|
+
",",
|
|
2880
|
+
" they",
|
|
2881
|
+
"'re",
|
|
2882
|
+
" unbelievable",
|
|
2883
|
+
" guys",
|
|
2884
|
+
',"',
|
|
2885
|
+
" said",
|
|
2886
|
+
" J",
|
|
2887
|
+
"okin",
|
|
2888
|
+
"en",
|
|
2889
|
+
".",
|
|
2890
|
+
' "',
|
|
2891
|
+
"It",
|
|
2892
|
+
"'s",
|
|
2893
|
+
" not",
|
|
2894
|
+
" just",
|
|
2895
|
+
" the",
|
|
2896
|
+
" players",
|
|
2897
|
+
",",
|
|
2898
|
+
" it",
|
|
2899
|
+
"'s",
|
|
2900
|
+
" the",
|
|
2901
|
+
" staff",
|
|
2902
|
+
" around",
|
|
2903
|
+
" the",
|
|
2904
|
+
" team",
|
|
2905
|
+
".",
|
|
2906
|
+
" I",
|
|
2907
|
+
" feel",
|
|
2908
|
+
" really",
|
|
2909
|
+
" bad",
|
|
2910
|
+
" for",
|
|
2911
|
+
" them",
|
|
2912
|
+
],
|
|
2913
|
+
token_pair_coordinates=[(1, 1)],
|
|
2914
|
+
),
|
|
2915
|
+
AttentionTokenPairExample(
|
|
2916
|
+
tokens=[
|
|
2917
|
+
" a",
|
|
2918
|
+
" Pv",
|
|
2919
|
+
"E",
|
|
2920
|
+
" game",
|
|
2921
|
+
",",
|
|
2922
|
+
" we",
|
|
2923
|
+
" probably",
|
|
2924
|
+
" would",
|
|
2925
|
+
" use",
|
|
2926
|
+
" it",
|
|
2927
|
+
" but",
|
|
2928
|
+
" based",
|
|
2929
|
+
" on",
|
|
2930
|
+
" the",
|
|
2931
|
+
" tests",
|
|
2932
|
+
" we",
|
|
2933
|
+
"'ve",
|
|
2934
|
+
" run",
|
|
2935
|
+
" on",
|
|
2936
|
+
" it",
|
|
2937
|
+
",",
|
|
2938
|
+
" that",
|
|
2939
|
+
" wouldn",
|
|
2940
|
+
"'t",
|
|
2941
|
+
" be",
|
|
2942
|
+
" our",
|
|
2943
|
+
" first",
|
|
2944
|
+
" choice",
|
|
2945
|
+
" for",
|
|
2946
|
+
" a",
|
|
2947
|
+
" live",
|
|
2948
|
+
" R",
|
|
2949
|
+
"v",
|
|
2950
|
+
"R",
|
|
2951
|
+
" game",
|
|
2952
|
+
".",
|
|
2953
|
+
" Now",
|
|
2954
|
+
",",
|
|
2955
|
+
" could",
|
|
2956
|
+
" we",
|
|
2957
|
+
" use",
|
|
2958
|
+
" it",
|
|
2959
|
+
" for",
|
|
2960
|
+
" prototyp",
|
|
2961
|
+
"ing",
|
|
2962
|
+
"?",
|
|
2963
|
+
" Yep",
|
|
2964
|
+
",",
|
|
2965
|
+
" we",
|
|
2966
|
+
" are",
|
|
2967
|
+
" already",
|
|
2968
|
+
" doing",
|
|
2969
|
+
" that",
|
|
2970
|
+
".",
|
|
2971
|
+
" Second",
|
|
2972
|
+
",",
|
|
2973
|
+
" as",
|
|
2974
|
+
" to",
|
|
2975
|
+
" other",
|
|
2976
|
+
" engines",
|
|
2977
|
+
" there",
|
|
2978
|
+
" are",
|
|
2979
|
+
" both",
|
|
2980
|
+
" financial",
|
|
2981
|
+
],
|
|
2982
|
+
token_pair_coordinates=[(1, 1)],
|
|
2983
|
+
),
|
|
2984
|
+
AttentionTokenPairExample(
|
|
2985
|
+
tokens=[
|
|
2986
|
+
"-",
|
|
2987
|
+
"tun",
|
|
2988
|
+
"er",
|
|
2989
|
+
" is",
|
|
2990
|
+
" also",
|
|
2991
|
+
" custom",
|
|
2992
|
+
"isable",
|
|
2993
|
+
" for",
|
|
2994
|
+
" hassle",
|
|
2995
|
+
"-",
|
|
2996
|
+
"free",
|
|
2997
|
+
" experimentation",
|
|
2998
|
+
".",
|
|
2999
|
+
" The",
|
|
3000
|
+
" St",
|
|
3001
|
+
"rix",
|
|
3002
|
+
" X",
|
|
3003
|
+
"399",
|
|
3004
|
+
"-",
|
|
3005
|
+
"E",
|
|
3006
|
+
" Gaming",
|
|
3007
|
+
" takes",
|
|
3008
|
+
" up",
|
|
3009
|
+
" to",
|
|
3010
|
+
" three",
|
|
3011
|
+
" double",
|
|
3012
|
+
"-",
|
|
3013
|
+
"wide",
|
|
3014
|
+
" cards",
|
|
3015
|
+
" in",
|
|
3016
|
+
" SLI",
|
|
3017
|
+
" or",
|
|
3018
|
+
" Cross",
|
|
3019
|
+
"Fire",
|
|
3020
|
+
"X",
|
|
3021
|
+
".",
|
|
3022
|
+
" Primary",
|
|
3023
|
+
" graphics",
|
|
3024
|
+
" slots",
|
|
3025
|
+
" are",
|
|
3026
|
+
" protected",
|
|
3027
|
+
" by",
|
|
3028
|
+
" Safe",
|
|
3029
|
+
"Slot",
|
|
3030
|
+
" from",
|
|
3031
|
+
" damages",
|
|
3032
|
+
" that",
|
|
3033
|
+
" heavy",
|
|
3034
|
+
" GPU",
|
|
3035
|
+
" cool",
|
|
3036
|
+
"ers",
|
|
3037
|
+
" can",
|
|
3038
|
+
" potentially",
|
|
3039
|
+
" cause",
|
|
3040
|
+
".",
|
|
3041
|
+
"\n",
|
|
3042
|
+
"\n",
|
|
3043
|
+
"Personal",
|
|
3044
|
+
"ised",
|
|
3045
|
+
" RGB",
|
|
3046
|
+
" lighting",
|
|
3047
|
+
" is",
|
|
3048
|
+
" made",
|
|
3049
|
+
" possible",
|
|
3050
|
+
],
|
|
3051
|
+
token_pair_coordinates=[(1, 1)],
|
|
3052
|
+
),
|
|
3053
|
+
AttentionTokenPairExample(
|
|
3054
|
+
tokens=[
|
|
3055
|
+
" to",
|
|
3056
|
+
" abolish",
|
|
3057
|
+
" such",
|
|
3058
|
+
" a",
|
|
3059
|
+
" complex",
|
|
3060
|
+
"?",
|
|
3061
|
+
" Are",
|
|
3062
|
+
" there",
|
|
3063
|
+
" ways",
|
|
3064
|
+
" ve",
|
|
3065
|
+
"gans",
|
|
3066
|
+
" can",
|
|
3067
|
+
" eat",
|
|
3068
|
+
" more",
|
|
3069
|
+
" sustain",
|
|
3070
|
+
"ably",
|
|
3071
|
+
"?",
|
|
3072
|
+
" What",
|
|
3073
|
+
" are",
|
|
3074
|
+
" some",
|
|
3075
|
+
" of",
|
|
3076
|
+
" the",
|
|
3077
|
+
" health",
|
|
3078
|
+
" challenges",
|
|
3079
|
+
" for",
|
|
3080
|
+
" new",
|
|
3081
|
+
" ve",
|
|
3082
|
+
"gans",
|
|
3083
|
+
",",
|
|
3084
|
+
" and",
|
|
3085
|
+
" how",
|
|
3086
|
+
" can",
|
|
3087
|
+
" we",
|
|
3088
|
+
" raise",
|
|
3089
|
+
" awareness",
|
|
3090
|
+
" of",
|
|
3091
|
+
" these",
|
|
3092
|
+
" issues",
|
|
3093
|
+
" so",
|
|
3094
|
+
" that",
|
|
3095
|
+
",",
|
|
3096
|
+
" for",
|
|
3097
|
+
" instance",
|
|
3098
|
+
",",
|
|
3099
|
+
" medical",
|
|
3100
|
+
" professionals",
|
|
3101
|
+
" are",
|
|
3102
|
+
" more",
|
|
3103
|
+
" supportive",
|
|
3104
|
+
" of",
|
|
3105
|
+
" vegan",
|
|
3106
|
+
"ism",
|
|
3107
|
+
"?",
|
|
3108
|
+
"\n",
|
|
3109
|
+
"\n",
|
|
3110
|
+
"Moreover",
|
|
3111
|
+
",",
|
|
3112
|
+
" it",
|
|
3113
|
+
" is",
|
|
3114
|
+
" essential",
|
|
3115
|
+
" that",
|
|
3116
|
+
" ve",
|
|
3117
|
+
"gans",
|
|
3118
|
+
" differentiate",
|
|
3119
|
+
],
|
|
3120
|
+
token_pair_coordinates=[(1, 1)],
|
|
3121
|
+
),
|
|
3122
|
+
],
|
|
3123
|
+
explanation="attends from the second token in the sequence to the second token in the sequence",
|
|
3124
|
+
),
|
|
3125
|
+
]
|