Batnini commited on
Commit
2b8b800
·
verified ·
1 Parent(s): daa2716

A/B 2026-10-22: manifest.json

Browse files
Files changed (1) hide show
  1. probe/ab_2026-10-22/manifest.json +354 -0
probe/ab_2026-10-22/manifest.json ADDED
@@ -0,0 +1,354 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "ausgefuehrt",
3
+ "generated_by": "m1_m2_m3_probe.py (read-only, kein Training; Amendment A1: R3a/R3b)",
4
+ "base_model": "Qwen/Qwen2.5-3B-Instruct",
5
+ "resolved_commits": {
6
+ "requested_base_revision": "aa8e72537993ba99e69dfaafa59ed015b17504d1",
7
+ "requested_tokenizer_revision": "aa8e72537993ba99e69dfaafa59ed015b17504d1",
8
+ "base_commit": "aa8e72537993ba99e69dfaafa59ed015b17504d1",
9
+ "base_commit_source": "config._commit_hash",
10
+ "tokenizer_commit": null,
11
+ "tokenizer_commit_source": "unresolved"
12
+ },
13
+ "fingerprints": {
14
+ "tokenizer_chat_template_sha256": "cd8e9439f0570856fd70470bf8889ebd8b5d1107207f67a5efb46e342330527f",
15
+ "tokenizer_vocab_size": 151643,
16
+ "tokenizer_name_or_path": "Qwen/Qwen2.5-3B-Instruct"
17
+ },
18
+ "dataset": "/kaggle/working/ds15.jsonl",
19
+ "dataset_sha256": "18caaae7c6f7f506a0cdd132d15f23c0e7044b07d894b73eb68b302605b01c7c",
20
+ "n_probes": 15,
21
+ "adapter_b2prime": "/kaggle/working/ab_rslora1/adapter",
22
+ "adapter_b2prime_sha256": "a595d437f4fcb0c6030840d9e70e7ff6818a5e490943aabe19f00c3dee6ba9c5",
23
+ "adapter_b2prime_cfg_sha256": "3df2101372f92c100dd3c15275af0f05faeb8f36f0cce7e41d24134203abe129",
24
+ "adapter_b2": "/kaggle/working/ab_rsloRa0/adapter",
25
+ "adapter_b2_sha256": "e3be2c8518433696a6aa5b088a43823facc099b20c7beab385e9352ec3023f2e",
26
+ "adapter_b2_cfg_sha256": "56cc999b782b44a4f2f50b76100a9fa848eaecf3b333677fa243b50257a4c2cf",
27
+ "b2prime_cfg": {
28
+ "r": 64,
29
+ "alpha": 256,
30
+ "use_rslora": true,
31
+ "base_model_name_or_path": "Qwen/Qwen2.5-3B-Instruct",
32
+ "target_modules": [
33
+ "down_proj",
34
+ "k_proj",
35
+ "o_proj",
36
+ "v_proj",
37
+ "gate_proj",
38
+ "q_proj",
39
+ "up_proj"
40
+ ],
41
+ "lora_dropout": 0.05
42
+ },
43
+ "b2_cfg": {
44
+ "r": 64,
45
+ "alpha": 256,
46
+ "use_rslora": false,
47
+ "base_model_name_or_path": "Qwen/Qwen2.5-3B-Instruct",
48
+ "target_modules": [
49
+ "o_proj",
50
+ "v_proj",
51
+ "down_proj",
52
+ "up_proj",
53
+ "q_proj",
54
+ "gate_proj",
55
+ "k_proj"
56
+ ],
57
+ "lora_dropout": 0.05
58
+ },
59
+ "env": {
60
+ "python": "3.13.15",
61
+ "torch": "2.11.0+cu128",
62
+ "transformers": "5.16.1",
63
+ "peft": "0.20.0",
64
+ "torchao": "0.18.0",
65
+ "gpu": "Tesla T4",
66
+ "capability": [
67
+ 7,
68
+ 5
69
+ ],
70
+ "dtype": "float16"
71
+ },
72
+ "decoding": {
73
+ "max_new_tokens": 16,
74
+ "do_sample": false,
75
+ "seed": 42,
76
+ "reseeded_before_each_pass": true
77
+ },
78
+ "thresholds": {
79
+ "good": 11,
80
+ "bad": 4,
81
+ "diff": 3,
82
+ "collapse_min": 12,
83
+ "cross_instance_noise_budget": 2,
84
+ "hinweis": "deskriptiv; n=15 => 1 Item = 6,7 pp; keine Signifikanz"
85
+ },
86
+ "grader_crosscheck": "ok (repo acceptance_gate.py identisch, 120 Paare x 2 Modi)",
87
+ "measurement_passes": [
88
+ {
89
+ "pass": "M1_base",
90
+ "instance": "base#1",
91
+ "n": 15,
92
+ "strict": 0,
93
+ "substring": 0
94
+ },
95
+ {
96
+ "pass": "M1_repeat",
97
+ "instance": "base#1",
98
+ "n": 15,
99
+ "strict": 0,
100
+ "substring": 0
101
+ },
102
+ {
103
+ "pass": "M1b_base",
104
+ "instance": "base#2",
105
+ "n": 15,
106
+ "strict": 0,
107
+ "substring": 0
108
+ },
109
+ {
110
+ "pass": "M2_b2prime",
111
+ "instance": "base#2",
112
+ "n": 15,
113
+ "strict": 1,
114
+ "substring": 1
115
+ },
116
+ {
117
+ "pass": "canary_disable_adapter",
118
+ "instance": "base#2",
119
+ "n": 15,
120
+ "strict": 0,
121
+ "substring": 0
122
+ },
123
+ {
124
+ "pass": "M2_repeat",
125
+ "instance": "base#2",
126
+ "n": 15,
127
+ "strict": 1,
128
+ "substring": 1
129
+ },
130
+ {
131
+ "pass": "M1c_base",
132
+ "instance": "base#3",
133
+ "n": 15,
134
+ "strict": 0,
135
+ "substring": 0
136
+ },
137
+ {
138
+ "pass": "M3_b2",
139
+ "instance": "base#3",
140
+ "n": 15,
141
+ "strict": 15,
142
+ "substring": 15
143
+ }
144
+ ],
145
+ "counts": {
146
+ "M1_base": {
147
+ "n": 15,
148
+ "strict": 0,
149
+ "substring": 0,
150
+ "gold_in_raw_output": 0
151
+ },
152
+ "M1_repeat": {
153
+ "n": 15,
154
+ "strict": 0,
155
+ "substring": 0,
156
+ "gold_in_raw_output": 0
157
+ },
158
+ "M1b_base": {
159
+ "n": 15,
160
+ "strict": 0,
161
+ "substring": 0,
162
+ "gold_in_raw_output": 0
163
+ },
164
+ "M2_b2prime": {
165
+ "n": 15,
166
+ "strict": 1,
167
+ "substring": 1,
168
+ "gold_in_raw_output": 1
169
+ },
170
+ "canary_disable_adapter": {
171
+ "n": 15,
172
+ "strict": 0,
173
+ "substring": 0,
174
+ "gold_in_raw_output": 0
175
+ },
176
+ "M2_repeat": {
177
+ "n": 15,
178
+ "strict": 1,
179
+ "substring": 1,
180
+ "gold_in_raw_output": 1
181
+ },
182
+ "M1c_base": {
183
+ "n": 15,
184
+ "strict": 0,
185
+ "substring": 0,
186
+ "gold_in_raw_output": 0
187
+ },
188
+ "M3_b2": {
189
+ "n": 15,
190
+ "strict": 15,
191
+ "substring": 15,
192
+ "gold_in_raw_output": 15
193
+ }
194
+ },
195
+ "determinism_baseline": {
196
+ "matched": 15,
197
+ "n": 15,
198
+ "ratio": 1.0,
199
+ "mismatch_items": [],
200
+ "a": "M1_base",
201
+ "b": "M1_repeat",
202
+ "same_instance": true,
203
+ "threshold": 1.0,
204
+ "status": "pass"
205
+ },
206
+ "cross_instance_noise": {
207
+ "matched": 15,
208
+ "n": 15,
209
+ "ratio": 1.0,
210
+ "mismatch_items": [],
211
+ "a": "M1_base",
212
+ "b": "M1b_base",
213
+ "same_instance": false,
214
+ "budget": 2,
215
+ "within_budget": true,
216
+ "status": "informational"
217
+ },
218
+ "canary": {
219
+ "matched": 15,
220
+ "n": 15,
221
+ "ratio": 1.0,
222
+ "mismatch_items": [],
223
+ "a": "M1b_base",
224
+ "b": "canary_disable_adapter",
225
+ "same_instance": true,
226
+ "threshold": 1.0,
227
+ "status": "pass",
228
+ "hinweis": "Referenz ist M1b (dieselbe Instanz, unwrapped) — kein Cross-Instance-Rauschen"
229
+ },
230
+ "positive_control": {
231
+ "w1_generation": {
232
+ "matched": 0,
233
+ "n": 15,
234
+ "ratio": 0.0,
235
+ "mismatch_items": [
236
+ "At what airport can you find UTair Aviation?",
237
+ "On what date did 40th government of Turkey end?",
238
+ "What city was Christians & Lions formed in?",
239
+ "What is the name of the singer or writer of Odorono?",
240
+ "What person illustrated The Faraway Lurs?",
241
+ "What team is Beau Allred affiliated with?",
242
+ "What town or city does KKBB serve?",
243
+ "What was Isaac Nicola's city of birth?",
244
+ "What was the name of the city or state where KOAN Sound was formed?",
245
+ "What was the production company for Getting Acquainted?",
246
+ "What was the title that Traidenis held?",
247
+ "What year did Jyllands-Posten Muhammad cartoons controversy start?",
248
+ "What year was the end of Comdial?",
249
+ "Which was the record label for Renincarnated?",
250
+ "Who is the illustrator of Cycle of the Werewolf?"
251
+ ],
252
+ "a": "M2_b2prime",
253
+ "b": "canary_disable_adapter",
254
+ "same_instance": true,
255
+ "criterion": "differing_items >= 1",
256
+ "differing_items": 15
257
+ },
258
+ "w2_logit": {
259
+ "probe": "What was the title that Traidenis held?",
260
+ "max_abs_logit_delta": 25.0888671875,
261
+ "argmax_changed": true,
262
+ "source": "pre-wrap vs post-wrap, Instanz #2"
263
+ },
264
+ "w3_structural": {
265
+ "lora_modules_total": 252,
266
+ "lora_B_nonzero": 252,
267
+ "target_modules_expected": [
268
+ "down_proj",
269
+ "k_proj",
270
+ "o_proj",
271
+ "v_proj",
272
+ "gate_proj",
273
+ "q_proj",
274
+ "up_proj"
275
+ ]
276
+ },
277
+ "w4_adapter_path_determinism": {
278
+ "a": "M2_b2prime",
279
+ "b": "M2_repeat",
280
+ "matched": 15,
281
+ "n": 15
282
+ },
283
+ "status": "pass",
284
+ "interpretation": "M2 != canary auf >= 1 Item => Adapter war aktiv (Positivkontrolle bestanden)."
285
+ },
286
+ "format_control": {
287
+ "collapse_min_items": 12,
288
+ "per_pass": {
289
+ "M1_base": {
290
+ "collapse": false,
291
+ "top_answer": "I'm sorry, but there seems to be an error in your question. The",
292
+ "top_count": 1
293
+ },
294
+ "M1b_base": {
295
+ "collapse": false,
296
+ "top_answer": "I'm sorry, but there seems to be an error in your question. The",
297
+ "top_count": 1
298
+ },
299
+ "M1c_base": {
300
+ "collapse": false,
301
+ "top_answer": "I'm sorry, but there seems to be an error in your question. The",
302
+ "top_count": 1
303
+ },
304
+ "M2_b2prime": {
305
+ "collapse": true,
306
+ "top_answer": "Jamaica",
307
+ "top_count": 13
308
+ },
309
+ "M3_b2": {
310
+ "collapse": false,
311
+ "top_answer": "Grand Duke of Lithuania",
312
+ "top_count": 1
313
+ }
314
+ },
315
+ "b2prime_strict_vs_substring_diff_items": [],
316
+ "b2_strict_vs_substring_diff_items": [],
317
+ "note": "substring prueft gold ⊂ antwort (acceptance_gate.py:123-130); gold 'Jamaica Plain' vs antwort 'Jamaica' ist in BEIDEN Modi FAIL."
318
+ },
319
+ "regel6_reference": {
320
+ "reference": null,
321
+ "available": false
322
+ },
323
+ "decision": {
324
+ "result": "unentschieden",
325
+ "rule": 9,
326
+ "precondition_failed": [
327
+ "m2_collapse"
328
+ ],
329
+ "base_reference_used": {
330
+ "M2": "M1b_base",
331
+ "M3": "M1c_base"
332
+ },
333
+ "rules": [
334
+ {
335
+ "nr": 9,
336
+ "text": "M2 degeneriert (13/15 identische Antwort 'Jamaica') — Diagnose blockiert, kein P-1/P-2/P-3-Zweig"
337
+ },
338
+ {
339
+ "nr": 5,
340
+ "text": "P-2a: N8 invalidiert base_answer/em_base/changed/Delta (write_back_cycle.py:420-421/442-445)"
341
+ },
342
+ {
343
+ "nr": 6,
344
+ "text": "P-2b: nicht ausgefuehrt (kein Referenz-Altlauf uebergeben)"
345
+ }
346
+ ]
347
+ },
348
+ "grenzen": [
349
+ "Die 15 Zeilen sind TRAININGSzeilen (Dataset 18caaae7...) => M2/M3 messen Memorization, keine Generalisierung.",
350
+ "M3 hat ein Ceiling: der B2-Adapter beantwortete im Altlauf alle 15 Gold-Strings exakt (15/15) => M3 kann P-1 vs. P-3 nicht trennen. Diskriminativ waere ein Held-out-Probe-Set.",
351
+ "M3 ist ein historischer Code-Stand (der B2-Adapter hat use_rslora=false) => 'konsistent mit P-1', nicht bewiesen.",
352
+ "Cross-Instance-Noise (M1 vs M1b) ist Rauschen und wird NICHT als Diagnose verwendet; alle Regelvergleiche nutzen gepaarte Referenzen (M1b fuer M2, M1c fuer M3)."
353
+ ]
354
+ }