AbstractPhil commited on
Commit
4dcc048
·
verified ·
1 Parent(s): 2b5d552

btx-e003 R3 ledger

Browse files
mini-beatrix-1/reports/arms/btx_e003_R3_ledger.json ADDED
@@ -0,0 +1,338 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "rung": "btx-e003-R3",
3
+ "started": "2026-09-06 08:22",
4
+ "teacher_gate": {
5
+ "instructed_strict": true,
6
+ "uninstructed_chain": true,
7
+ "final_predicate_any": true
8
+ },
9
+ "restricted_targets": false,
10
+ "lm_clean": {
11
+ "nextbyte": 0.5679,
12
+ "bpb": 2.1786
13
+ },
14
+ "floors_from_R0": [
15
+ {
16
+ "arm": "bare@88508",
17
+ "exam150": 0.0,
18
+ "exam30": 0.0
19
+ },
20
+ {
21
+ "arm": "bare@58664",
22
+ "exam150": 0.0,
23
+ "exam30": 0.0
24
+ },
25
+ {
26
+ "arm": "d5-XL",
27
+ "step": 88508,
28
+ "params": 3185008,
29
+ "spec": {
30
+ "n_slots": 16,
31
+ "K": 64,
32
+ "D": 4,
33
+ "hidden": 178
34
+ },
35
+ "exam150": 0.4,
36
+ "exam30": 1.0,
37
+ "exam150_v0_wordform_shift": 0.20666666666666667
38
+ },
39
+ {
40
+ "arm": "d5-XL-wide",
41
+ "step": 88508,
42
+ "params": 5275664,
43
+ "spec": {
44
+ "n_slots": 32,
45
+ "K": 64,
46
+ "D": 4,
47
+ "hidden": 256
48
+ },
49
+ "exam150": 0.47333333333333333,
50
+ "exam30": 1.0,
51
+ "exam150_v0_wordform_shift": 0.23333333333333334
52
+ },
53
+ {
54
+ "arm": "loopctrl-d5-s0",
55
+ "step": 88508,
56
+ "params": 5275664,
57
+ "spec": {
58
+ "n_slots": 32,
59
+ "K": 64,
60
+ "D": 4,
61
+ "hidden": 256
62
+ },
63
+ "exam150": 0.1,
64
+ "exam30": 0.8333333333333334,
65
+ "exam150_v0_wordform_shift": 0.09333333333333334
66
+ },
67
+ {
68
+ "arm": "d5ctrl-s0",
69
+ "step": 58664,
70
+ "params": 3185008,
71
+ "spec": {
72
+ "n_slots": 16,
73
+ "K": 64,
74
+ "D": 4,
75
+ "hidden": 178
76
+ },
77
+ "exam150": 0.44,
78
+ "exam30": 0.8,
79
+ "exam150_v0_wordform_shift": 0.24666666666666667
80
+ },
81
+ {
82
+ "arm": "d5ctrl-s1",
83
+ "step": 58664,
84
+ "params": 3185008,
85
+ "spec": {
86
+ "n_slots": 16,
87
+ "K": 64,
88
+ "D": 4,
89
+ "hidden": 178
90
+ },
91
+ "exam150": 0.43333333333333335,
92
+ "exam30": 0.6333333333333333,
93
+ "exam150_v0_wordform_shift": 0.30666666666666664
94
+ }
95
+ ],
96
+ "gauge_only": [
97
+ {
98
+ "arm": "d5-XL",
99
+ "step": 88508,
100
+ "params": 3185008,
101
+ "spec": {
102
+ "n_slots": 16,
103
+ "K": 64,
104
+ "D": 4,
105
+ "hidden": 178
106
+ },
107
+ "meta": {
108
+ "step": 88508,
109
+ "name": "d5-XL",
110
+ "group": "day1",
111
+ "kind": "behavior",
112
+ "behavior": "d5-XL",
113
+ "sites": 16,
114
+ "params": 3185008,
115
+ "base_model_id": "alephllm/mini-beatrix-1@step88508",
116
+ "verified": true,
117
+ "template": {
118
+ "frame": "qa",
119
+ "stops": [
120
+ "\n\n",
121
+ "\nUser:",
122
+ "\nBeatrix:",
123
+ "User:",
124
+ "Beatrix:"
125
+ ],
126
+ "chat_ok": true,
127
+ "turn_end": "newline pair",
128
+ "source": "overnight_arms.row(): single-turn QA frame, reply + b'\\n\\n'"
129
+ },
130
+ "decode": {
131
+ "temperature": 0.0,
132
+ "top_p": 1.0,
133
+ "why": "task arms collapse to turn-end off-distribution and degenerate into punctuation under sampling (measured 2026-08-19) \u2014 greedy is their honest setting"
134
+ },
135
+ "examples": [
136
+ "If someone is glent, then they are sorb. If someone is sorb, then they are trell. Wren is glent. What follows?",
137
+ "Rules: If someone is vint, then they are kresp. If someone is kresp, then they are dorn. Fact: Nils is vint. Conclusion?"
138
+ ]
139
+ },
140
+ "exam150": 0.94,
141
+ "exam150_n_answers": 5.06,
142
+ "exam30": 1.0,
143
+ "exam30_n_answers": 5.0,
144
+ "exam150_v1_novel_aliens": 0.4,
145
+ "exam150_v1_novel_aliens_n_answers": 5.22,
146
+ "exam150_v0_wordform_shift": 0.20666666666666667,
147
+ "exam150_v0_wordform_shift_n_answers": 5.133,
148
+ "interference": {
149
+ "nextbyte_on": 0.5822,
150
+ "bpb_on": 2.5533,
151
+ "bpb_clean": 2.1786,
152
+ "bpb_delta": 0.3747,
153
+ "nextbyte_clean": 0.5679
154
+ }
155
+ },
156
+ {
157
+ "arm": "d5-XL-wide",
158
+ "step": 88508,
159
+ "params": 5275664,
160
+ "spec": {
161
+ "n_slots": 32,
162
+ "K": 64,
163
+ "D": 4,
164
+ "hidden": 256
165
+ },
166
+ "meta": {
167
+ "step": 88508,
168
+ "name": "d5-XL-wide",
169
+ "group": "day1",
170
+ "kind": "behavior",
171
+ "behavior": "d5-XL-wide",
172
+ "sites": 16,
173
+ "params": 5275664,
174
+ "base_model_id": "alephllm/mini-beatrix-1@step88508",
175
+ "verified": true,
176
+ "template": {
177
+ "frame": "qa",
178
+ "stops": [
179
+ "\n\n",
180
+ "\nUser:",
181
+ "\nBeatrix:",
182
+ "User:",
183
+ "Beatrix:"
184
+ ],
185
+ "chat_ok": true,
186
+ "turn_end": "newline pair",
187
+ "source": "overnight_arms.row(): single-turn QA frame, reply + b'\\n\\n'"
188
+ },
189
+ "decode": {
190
+ "temperature": 0.0,
191
+ "top_p": 1.0,
192
+ "why": "task arms collapse to turn-end off-distribution and degenerate into punctuation under sampling (measured 2026-08-19) \u2014 greedy is their honest setting"
193
+ },
194
+ "examples": [
195
+ "If someone is glent, then they are sorb. If someone is sorb, then they are trell. Wren is glent. What follows?",
196
+ "Rules: If someone is vint, then they are kresp. If someone is kresp, then they are dorn. Fact: Nils is vint. Conclusion?"
197
+ ]
198
+ },
199
+ "exam150": 1.0,
200
+ "exam150_n_answers": 5.0,
201
+ "exam30": 1.0,
202
+ "exam30_n_answers": 5.0,
203
+ "exam150_v1_novel_aliens": 0.47333333333333333,
204
+ "exam150_v1_novel_aliens_n_answers": 5.033,
205
+ "exam150_v0_wordform_shift": 0.23333333333333334,
206
+ "exam150_v0_wordform_shift_n_answers": 4.78,
207
+ "interference": {
208
+ "nextbyte_on": 0.5828,
209
+ "bpb_on": 2.5352,
210
+ "bpb_clean": 2.1786,
211
+ "bpb_delta": 0.3566,
212
+ "nextbyte_clean": 0.5679
213
+ }
214
+ }
215
+ ],
216
+ "arms": [
217
+ {
218
+ "arm": "mb-T1-s0",
219
+ "torch_seed": 719195110,
220
+ "restricted": false,
221
+ "descent_gate": {
222
+ "passed": true,
223
+ "first": 0.402341,
224
+ "last": 6.5e-05,
225
+ "rel_drop": 0.9998
226
+ },
227
+ "train_s": 3716,
228
+ "curve": [
229
+ [
230
+ 250,
231
+ 0.14236
232
+ ],
233
+ [
234
+ 500,
235
+ 0.03115
236
+ ],
237
+ [
238
+ 750,
239
+ 0.02579
240
+ ],
241
+ [
242
+ 1000,
243
+ 0.02036
244
+ ],
245
+ [
246
+ 1250,
247
+ 0.01346
248
+ ],
249
+ [
250
+ 1500,
251
+ 0.01316
252
+ ],
253
+ [
254
+ 1750,
255
+ 0.00972
256
+ ],
257
+ [
258
+ 2000,
259
+ 0.00956
260
+ ],
261
+ [
262
+ 2250,
263
+ 0.00994
264
+ ],
265
+ [
266
+ 2500,
267
+ 0.0082
268
+ ],
269
+ [
270
+ 2750,
271
+ 0.00711
272
+ ],
273
+ [
274
+ 3000,
275
+ 0.01392
276
+ ],
277
+ [
278
+ 3250,
279
+ 0.01091
280
+ ],
281
+ [
282
+ 3500,
283
+ 0.0075
284
+ ],
285
+ [
286
+ 3750,
287
+ 0.0075
288
+ ],
289
+ [
290
+ 4000,
291
+ 0.01151
292
+ ]
293
+ ],
294
+ "target_stats": {
295
+ "positions": 2714921,
296
+ "token_start_fraction": 0.4125,
297
+ "entropy_at_starts": 0.0903,
298
+ "entropy_mid_token": 0.1743,
299
+ "argmax_ne_gold_fraction": 0.107,
300
+ "fallback_fraction": 0.0011,
301
+ "restricted_fraction": 0.0
302
+ },
303
+ "exam150": 0.0,
304
+ "exam150_n_answers": 0.0,
305
+ "exam30": 0.0,
306
+ "exam30_n_answers": 0.0,
307
+ "exam150_v1_novel_aliens": 0.0,
308
+ "exam150_v1_novel_aliens_n_answers": 0.0,
309
+ "exam150_v0_wordform_shift": 0.0,
310
+ "exam150_v0_wordform_shift_n_answers": 0.0,
311
+ "interference": {
312
+ "nextbyte_on": 0.5789,
313
+ "bpb_on": 2.2834,
314
+ "bpb_clean": 2.1786,
315
+ "bpb_delta": 0.1048,
316
+ "nextbyte_clean": 0.5679
317
+ },
318
+ "bitexact_off": true,
319
+ "anchor": "E:\\mirel\\data\\btx_e003\\arms\\btx-e003-R3-mb-T1-s0@step88508.anchor.pt",
320
+ "alloy_ledger": {
321
+ "mb": {
322
+ "d5": {
323
+ "n": 4000,
324
+ "mean": 0.036229
325
+ }
326
+ }
327
+ },
328
+ "peak_vram_gb": 8.23
329
+ }
330
+ ],
331
+ "_env": {
332
+ "torch": "2.11.0+cu128",
333
+ "gpu": "NVIDIA GeForce RTX 4090",
334
+ "steps": 4000,
335
+ "batch": 8,
336
+ "spec": "WIDE n32 K64 D4 h256"
337
+ }
338
+ }