lamshiu commited on
Commit
2d5a15b
·
verified ·
1 Parent(s): b4f8eb0

Upload folder using huggingface_hub

Browse files
5m-s0-bpe4096-step-bytes3-v17-nofs2md-nofs2cm-nofs2dd-none-docmask/ckpt.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4f804d55d4a4c31762904cfed9a226eefcd6dc2fb9862a1b86080c3bb0bc12de
3
+ size 23171663
5m-s0-bpe4096-step-bytes3-v17-nofs2md-nofs2cm-nofs2dd-none-docmask/config.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "size": "5m",
3
+ "name": "5m-s0-bpe4096-step-bytes3-v17-nofs2md-nofs2cm-nofs2dd-none-docmask",
4
+ "seed": 0,
5
+ "tokens": 400000000.0,
6
+ "batch_size": 64,
7
+ "block_size": 512,
8
+ "lr": 0.0006,
9
+ "eval_every": 4000,
10
+ "vocab": "bpe4096",
11
+ "pos": "none",
12
+ "reason": "step",
13
+ "reason_codec": "bytes",
14
+ "per_chunk": 3,
15
+ "docmask": true,
16
+ "compile": true,
17
+ "obs_loss": false,
18
+ "weight": [],
19
+ "holdout": [
20
+ "fs2md",
21
+ "fs2cm",
22
+ "fs2dd"
23
+ ],
24
+ "dataset": "v17",
25
+ "vocab_size": 4096,
26
+ "n_layer": 6,
27
+ "n_head": 4,
28
+ "d_model": 256,
29
+ "loops": 1,
30
+ "params": 5787648,
31
+ "flops_per_token": 44163072,
32
+ "steps": 12207,
33
+ "n_docs": 1171434,
34
+ "stream_tokens": 193933137
35
+ }
5m-s0-bpe4096-step-bytes3-v17-nofs2md-nofs2cm-nofs2dd-none-docmask/log.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
5m-s0-bpe4096-step-bytes3-v17-nofs2md-nofs2cm-nofs2dd-none-docmask/summary.json ADDED
@@ -0,0 +1,985 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "params": 5787648,
3
+ "tokens": 399998976,
4
+ "train_flops": 17665183577014272,
5
+ "wall_s": 1080.1574218273163,
6
+ "peak_mem_gb": 2.520964096,
7
+ "final_val_loss": 1.7903438568115235,
8
+ "final_val_bpb": 0.6453898739968539,
9
+ "bytes_seen": 1600839241.8284984,
10
+ "add8": {
11
+ "acc": 0.996,
12
+ "reasoning_bytes": 79.0,
13
+ "flops_per_problem": 4928863813.632
14
+ },
15
+ "add9": {
16
+ "acc": 0.666,
17
+ "reasoning_bytes": 89.0,
18
+ "flops_per_problem": 5503513706.496
19
+ },
20
+ "add10": {
21
+ "acc": 0.002,
22
+ "reasoning_bytes": 98.992,
23
+ "flops_per_problem": 6077280337.92
24
+ },
25
+ "reach14": {
26
+ "acc": 0.632,
27
+ "reasoning_bytes": 6.912,
28
+ "flops_per_problem": 4589956399.104
29
+ },
30
+ "reach16": {
31
+ "acc": 0.618,
32
+ "reasoning_bytes": 6.622,
33
+ "flops_per_problem": 5113023823.872
34
+ },
35
+ "reach18": {
36
+ "acc": 0.572,
37
+ "reasoning_bytes": 7.196,
38
+ "flops_per_problem": 5662765744.128
39
+ },
40
+ "parity24": {
41
+ "acc": 0.73,
42
+ "reasoning_bytes": 142.472,
43
+ "flops_per_problem": 7837708713.984
44
+ },
45
+ "parity28": {
46
+ "acc": 0.526,
47
+ "reasoning_bytes": 164.036,
48
+ "flops_per_problem": 8966693486.592
49
+ },
50
+ "parity32": {
51
+ "acc": 0.476,
52
+ "reasoning_bytes": 185.552,
53
+ "flops_per_problem": 10093558431.744
54
+ },
55
+ "cmp8": {
56
+ "acc": 0.996,
57
+ "reasoning_bytes": 5.816,
58
+ "flops_per_problem": 1360929226.752
59
+ },
60
+ "cmp9": {
61
+ "acc": 0.992,
62
+ "reasoning_bytes": 5.724,
63
+ "flops_per_problem": 1445192368.128
64
+ },
65
+ "cmp10": {
66
+ "acc": 0.996,
67
+ "reasoning_bytes": 5.76,
68
+ "flops_per_problem": 1535108382.72
69
+ },
70
+ "runmax12": {
71
+ "acc": 1.0,
72
+ "reasoning_bytes": 71.0,
73
+ "flops_per_problem": 4151328768.0
74
+ },
75
+ "runmax14": {
76
+ "acc": 1.0,
77
+ "reasoning_bytes": 82.964,
78
+ "flops_per_problem": 4768021905.408
79
+ },
80
+ "runmax16": {
81
+ "acc": 0.982,
82
+ "reasoning_bytes": 90.26,
83
+ "flops_per_problem": 5178561822.72
84
+ },
85
+ "count16": {
86
+ "acc": 1.0,
87
+ "reasoning_bytes": 79.03,
88
+ "flops_per_problem": 4726773596.16
89
+ },
90
+ "count18": {
91
+ "acc": 0.952,
92
+ "reasoning_bytes": 89.05,
93
+ "flops_per_problem": 5257613721.6
94
+ },
95
+ "count20": {
96
+ "acc": 0.828,
97
+ "reasoning_bytes": 97.15,
98
+ "flops_per_problem": 5703660748.8
99
+ },
100
+ "brackets16": {
101
+ "acc": 1.0,
102
+ "reasoning_bytes": 53.286,
103
+ "flops_per_problem": 3634002542.592
104
+ },
105
+ "brackets18": {
106
+ "acc": 0.984,
107
+ "reasoning_bytes": 56.806,
108
+ "flops_per_problem": 3877782700.032
109
+ },
110
+ "brackets20": {
111
+ "acc": 0.904,
112
+ "reasoning_bytes": 63.884,
113
+ "flops_per_problem": 4278783393.792
114
+ },
115
+ "sumd8": {
116
+ "acc": 0.998,
117
+ "reasoning_bytes": 58.566,
118
+ "flops_per_problem": 3381301444.608
119
+ },
120
+ "sumd9": {
121
+ "acc": 0.922,
122
+ "reasoning_bytes": 66.392,
123
+ "flops_per_problem": 3771173044.224
124
+ },
125
+ "sumd10": {
126
+ "acc": 0.46,
127
+ "reasoning_bytes": 71.216,
128
+ "flops_per_problem": 4028378775.552
129
+ },
130
+ "inc8": {
131
+ "acc": 1.0,
132
+ "reasoning_bytes": 63.0,
133
+ "flops_per_problem": 3798024192.0
134
+ },
135
+ "inc9": {
136
+ "acc": 0.79,
137
+ "reasoning_bytes": 70.52,
138
+ "flops_per_problem": 4215806853.12
139
+ },
140
+ "inc10": {
141
+ "acc": 0.004,
142
+ "reasoning_bytes": 71.16,
143
+ "flops_per_problem": 4291679010.816
144
+ },
145
+ "add.cmp6": {
146
+ "acc": 0.998,
147
+ "reasoning_bytes": 92.634,
148
+ "flops_per_problem": 5532573007.872
149
+ },
150
+ "add.cmp7": {
151
+ "acc": 0.99,
152
+ "reasoning_bytes": 105.488,
153
+ "flops_per_problem": 6229996240.896
154
+ },
155
+ "add.cmp8": {
156
+ "acc": 0.984,
157
+ "reasoning_bytes": 118.61,
158
+ "flops_per_problem": 6943494832.128
159
+ },
160
+ "add.runmax6": {
161
+ "acc": 1.0,
162
+ "reasoning_bytes": 120.832,
163
+ "flops_per_problem": 6572878331.904
164
+ },
165
+ "add.runmax7": {
166
+ "acc": 0.988,
167
+ "reasoning_bytes": 138.864,
168
+ "flops_per_problem": 7457552990.208
169
+ },
170
+ "add.runmax8": {
171
+ "acc": 0.536,
172
+ "reasoning_bytes": 148.744,
173
+ "flops_per_problem": 8017894047.744
174
+ },
175
+ "add.count6": {
176
+ "acc": 0.994,
177
+ "reasoning_bytes": 115.648,
178
+ "flops_per_problem": 6388100038.656
179
+ },
180
+ "add.count7": {
181
+ "acc": 0.948,
182
+ "reasoning_bytes": 132.09,
183
+ "flops_per_problem": 7202555412.48
184
+ },
185
+ "add.count8": {
186
+ "acc": 0.35,
187
+ "reasoning_bytes": 147.776,
188
+ "flops_per_problem": 7984330113.024
189
+ },
190
+ "add.find6": {
191
+ "acc": 0.96,
192
+ "reasoning_bytes": 107.696,
193
+ "flops_per_problem": 5992752218.112
194
+ },
195
+ "add.find7": {
196
+ "acc": 0.896,
197
+ "reasoning_bytes": 121.144,
198
+ "flops_per_problem": 6674983354.368
199
+ },
200
+ "add.find8": {
201
+ "acc": 0.49,
202
+ "reasoning_bytes": 137.482,
203
+ "flops_per_problem": 7485552377.856
204
+ },
205
+ "add.rev6": {
206
+ "acc": 0.994,
207
+ "reasoning_bytes": 123.256,
208
+ "flops_per_problem": 6795018584.064
209
+ },
210
+ "add.rev7": {
211
+ "acc": 0.698,
212
+ "reasoning_bytes": 145.56,
213
+ "flops_per_problem": 7911637696.512
214
+ },
215
+ "add.rev8": {
216
+ "acc": 0.01,
217
+ "reasoning_bytes": 155.684,
218
+ "flops_per_problem": 8477189996.544
219
+ },
220
+ "add.sumd6": {
221
+ "acc": 0.98,
222
+ "reasoning_bytes": 126.476,
223
+ "flops_per_problem": 6777794985.984
224
+ },
225
+ "add.sumd7": {
226
+ "acc": 0.818,
227
+ "reasoning_bytes": 146.392,
228
+ "flops_per_problem": 7745761198.08
229
+ },
230
+ "add.sumd8": {
231
+ "acc": 0.046,
232
+ "reasoning_bytes": 156.276,
233
+ "flops_per_problem": 8305572298.752
234
+ },
235
+ "add.inc6": {
236
+ "acc": 0.996,
237
+ "reasoning_bytes": 131.04,
238
+ "flops_per_problem": 7138695610.368
239
+ },
240
+ "add.inc7": {
241
+ "acc": 0.936,
242
+ "reasoning_bytes": 150.888,
243
+ "flops_per_problem": 8148263436.288
244
+ },
245
+ "add.inc8": {
246
+ "acc": 0.076,
247
+ "reasoning_bytes": 155.03,
248
+ "flops_per_problem": 8466325880.832
249
+ },
250
+ "runmax.cmp12": {
251
+ "acc": 1.0,
252
+ "reasoning_bytes": 87.0,
253
+ "flops_per_problem": 5122916352.0
254
+ },
255
+ "runmax.cmp14": {
256
+ "acc": 0.998,
257
+ "reasoning_bytes": 97.944,
258
+ "flops_per_problem": 5694563155.968
259
+ },
260
+ "runmax.cmp16": {
261
+ "acc": 0.994,
262
+ "reasoning_bytes": 105.084,
263
+ "flops_per_problem": 6098213634.048
264
+ },
265
+ "runmax.runmax12": {
266
+ "acc": 1.0,
267
+ "reasoning_bytes": 88.0,
268
+ "flops_per_problem": 5211242496.0
269
+ },
270
+ "runmax.runmax14": {
271
+ "acc": 0.996,
272
+ "reasoning_bytes": 99.604,
273
+ "flops_per_problem": 5812036927.488
274
+ },
275
+ "runmax.runmax16": {
276
+ "acc": 0.974,
277
+ "reasoning_bytes": 106.132,
278
+ "flops_per_problem": 6188659605.504
279
+ },
280
+ "runmax.count12": {
281
+ "acc": 1.0,
282
+ "reasoning_bytes": 88.0,
283
+ "flops_per_problem": 5255405568.0
284
+ },
285
+ "runmax.count14": {
286
+ "acc": 1.0,
287
+ "reasoning_bytes": 99.376,
288
+ "flops_per_problem": 5846130819.072
289
+ },
290
+ "runmax.count16": {
291
+ "acc": 1.0,
292
+ "reasoning_bytes": 106.228,
293
+ "flops_per_problem": 6237062332.416
294
+ },
295
+ "runmax.find12": {
296
+ "acc": 1.0,
297
+ "reasoning_bytes": 88.0,
298
+ "flops_per_problem": 5211242496.0
299
+ },
300
+ "runmax.find14": {
301
+ "acc": 1.0,
302
+ "reasoning_bytes": 99.46,
303
+ "flops_per_problem": 5805677445.12
304
+ },
305
+ "runmax.find16": {
306
+ "acc": 0.986,
307
+ "reasoning_bytes": 106.2,
308
+ "flops_per_problem": 6191662694.4
309
+ },
310
+ "runmax.rev12": {
311
+ "acc": 1.0,
312
+ "reasoning_bytes": 83.0,
313
+ "flops_per_problem": 4857937920.0
314
+ },
315
+ "runmax.rev14": {
316
+ "acc": 0.994,
317
+ "reasoning_bytes": 94.652,
318
+ "flops_per_problem": 5460852178.944
319
+ },
320
+ "runmax.rev16": {
321
+ "acc": 0.97,
322
+ "reasoning_bytes": 101.192,
323
+ "flops_per_problem": 5838004813.824
324
+ },
325
+ "runmax.sumd12": {
326
+ "acc": 1.0,
327
+ "reasoning_bytes": 86.0,
328
+ "flops_per_problem": 5034590208.0
329
+ },
330
+ "runmax.sumd14": {
331
+ "acc": 0.994,
332
+ "reasoning_bytes": 97.364,
333
+ "flops_per_problem": 5624785502.208
334
+ },
335
+ "runmax.sumd16": {
336
+ "acc": 0.972,
337
+ "reasoning_bytes": 104.096,
338
+ "flops_per_problem": 6010417446.912
339
+ },
340
+ "runmax.inc12": {
341
+ "acc": 1.0,
342
+ "reasoning_bytes": 87.0,
343
+ "flops_per_problem": 5065769336.832
344
+ },
345
+ "runmax.inc14": {
346
+ "acc": 0.994,
347
+ "reasoning_bytes": 98.52,
348
+ "flops_per_problem": 5666122137.6
349
+ },
350
+ "runmax.inc16": {
351
+ "acc": 0.974,
352
+ "reasoning_bytes": 105.216,
353
+ "flops_per_problem": 6051842408.448
354
+ },
355
+ "count.cmp16": {
356
+ "acc": 0.996,
357
+ "reasoning_bytes": 95.0,
358
+ "flops_per_problem": 5697036288.0
359
+ },
360
+ "count.cmp18": {
361
+ "acc": 0.994,
362
+ "reasoning_bytes": 104.906,
363
+ "flops_per_problem": 6222930149.376
364
+ },
365
+ "count.cmp20": {
366
+ "acc": 0.966,
367
+ "reasoning_bytes": 112.444,
368
+ "flops_per_problem": 6644245856.256
369
+ },
370
+ "count.runmax16": {
371
+ "acc": 0.998,
372
+ "reasoning_bytes": 96.004,
373
+ "flops_per_problem": 5785539084.288
374
+ },
375
+ "count.runmax18": {
376
+ "acc": 0.98,
377
+ "reasoning_bytes": 105.94,
378
+ "flops_per_problem": 6312669511.68
379
+ },
380
+ "count.runmax20": {
381
+ "acc": 0.868,
382
+ "reasoning_bytes": 113.724,
383
+ "flops_per_problem": 6744761008.128
384
+ },
385
+ "count.count16": {
386
+ "acc": 1.0,
387
+ "reasoning_bytes": 96.0,
388
+ "flops_per_problem": 5829525504.0
389
+ },
390
+ "count.count18": {
391
+ "acc": 1.0,
392
+ "reasoning_bytes": 105.96,
393
+ "flops_per_problem": 6357715845.12
394
+ },
395
+ "count.count20": {
396
+ "acc": 0.994,
397
+ "reasoning_bytes": 113.75,
398
+ "flops_per_problem": 6790072320.0
399
+ },
400
+ "count.find16": {
401
+ "acc": 1.0,
402
+ "reasoning_bytes": 96.0,
403
+ "flops_per_problem": 5785362432.0
404
+ },
405
+ "count.find18": {
406
+ "acc": 0.992,
407
+ "reasoning_bytes": 105.92,
408
+ "flops_per_problem": 6311786250.24
409
+ },
410
+ "count.find20": {
411
+ "acc": 0.952,
412
+ "reasoning_bytes": 113.562,
413
+ "flops_per_problem": 6737606590.464
414
+ },
415
+ "count.rev16": {
416
+ "acc": 0.996,
417
+ "reasoning_bytes": 91.01,
418
+ "flops_per_problem": 5432499486.72
419
+ },
420
+ "count.rev18": {
421
+ "acc": 0.97,
422
+ "reasoning_bytes": 100.99,
423
+ "flops_per_problem": 5961573089.28
424
+ },
425
+ "count.rev20": {
426
+ "acc": 0.864,
427
+ "reasoning_bytes": 108.49,
428
+ "flops_per_problem": 6381122273.28
429
+ },
430
+ "count.sumd16": {
431
+ "acc": 0.996,
432
+ "reasoning_bytes": 94.02,
433
+ "flops_per_problem": 5609593405.44
434
+ },
435
+ "count.sumd18": {
436
+ "acc": 0.97,
437
+ "reasoning_bytes": 103.79,
438
+ "flops_per_problem": 6129392762.88
439
+ },
440
+ "count.sumd20": {
441
+ "acc": 0.834,
442
+ "reasoning_bytes": 111.03,
443
+ "flops_per_problem": 6537459548.16
444
+ },
445
+ "count.inc16": {
446
+ "acc": 0.996,
447
+ "reasoning_bytes": 95.01,
448
+ "flops_per_problem": 5609151774.72
449
+ },
450
+ "count.inc18": {
451
+ "acc": 0.968,
452
+ "reasoning_bytes": 104.9,
453
+ "flops_per_problem": 6134250700.8
454
+ },
455
+ "count.inc20": {
456
+ "acc": 0.862,
457
+ "reasoning_bytes": 112.422,
458
+ "flops_per_problem": 6554771472.384
459
+ },
460
+ "parity.cmp16": {
461
+ "acc": 1.0,
462
+ "reasoning_bytes": 111.0,
463
+ "flops_per_problem": 6359482368.0
464
+ },
465
+ "parity.cmp20": {
466
+ "acc": 0.98,
467
+ "reasoning_bytes": 133.128,
468
+ "flops_per_problem": 7513375113.216
469
+ },
470
+ "parity.cmp24": {
471
+ "acc": 0.968,
472
+ "reasoning_bytes": 152.992,
473
+ "flops_per_problem": 8567282663.424
474
+ },
475
+ "parity.runmax16": {
476
+ "acc": 1.0,
477
+ "reasoning_bytes": 112.0,
478
+ "flops_per_problem": 6447808512.0
479
+ },
480
+ "parity.runmax20": {
481
+ "acc": 0.802,
482
+ "reasoning_bytes": 132.52,
483
+ "flops_per_problem": 7530687037.44
484
+ },
485
+ "parity.runmax24": {
486
+ "acc": 0.686,
487
+ "reasoning_bytes": 152.572,
488
+ "flops_per_problem": 8592897245.184
489
+ },
490
+ "parity.count16": {
491
+ "acc": 1.0,
492
+ "reasoning_bytes": 112.0,
493
+ "flops_per_problem": 6491971584.0
494
+ },
495
+ "parity.count20": {
496
+ "acc": 1.0,
497
+ "reasoning_bytes": 132.712,
498
+ "flops_per_problem": 7583329419.264
499
+ },
500
+ "parity.count24": {
501
+ "acc": 1.0,
502
+ "reasoning_bytes": 152.716,
503
+ "flops_per_problem": 8643419799.552
504
+ },
505
+ "parity.find16": {
506
+ "acc": 1.0,
507
+ "reasoning_bytes": 112.0,
508
+ "flops_per_problem": 6447808512.0
509
+ },
510
+ "parity.find20": {
511
+ "acc": 0.974,
512
+ "reasoning_bytes": 134.176,
513
+ "flops_per_problem": 7603821084.672
514
+ },
515
+ "parity.find24": {
516
+ "acc": 0.958,
517
+ "reasoning_bytes": 153.732,
518
+ "flops_per_problem": 8644126408.704
519
+ },
520
+ "parity.rev16": {
521
+ "acc": 1.0,
522
+ "reasoning_bytes": 107.0,
523
+ "flops_per_problem": 6094503936.0
524
+ },
525
+ "parity.rev20": {
526
+ "acc": 0.894,
527
+ "reasoning_bytes": 129.728,
528
+ "flops_per_problem": 7274894524.416
529
+ },
530
+ "parity.rev24": {
531
+ "acc": 0.716,
532
+ "reasoning_bytes": 149.0,
533
+ "flops_per_problem": 8302657536.0
534
+ },
535
+ "parity.sumd16": {
536
+ "acc": 1.0,
537
+ "reasoning_bytes": 110.0,
538
+ "flops_per_problem": 6271156224.0
539
+ },
540
+ "parity.sumd20": {
541
+ "acc": 0.884,
542
+ "reasoning_bytes": 132.344,
543
+ "flops_per_problem": 7434588192.768
544
+ },
545
+ "parity.sumd24": {
546
+ "acc": 0.82,
547
+ "reasoning_bytes": 151.952,
548
+ "flops_per_problem": 8477189996.544
549
+ },
550
+ "parity.inc16": {
551
+ "acc": 1.0,
552
+ "reasoning_bytes": 111.0,
553
+ "flops_per_problem": 6271156224.0
554
+ },
555
+ "parity.inc20": {
556
+ "acc": 0.91,
557
+ "reasoning_bytes": 134.112,
558
+ "flops_per_problem": 7468505432.064
559
+ },
560
+ "parity.inc24": {
561
+ "acc": 0.77,
562
+ "reasoning_bytes": 153.06,
563
+ "flops_per_problem": 8481959608.32
564
+ },
565
+ "sumd.cmp6": {
566
+ "acc": 1.0,
567
+ "reasoning_bytes": 62.088,
568
+ "flops_per_problem": 3713407746.048
569
+ },
570
+ "sumd.cmp7": {
571
+ "acc": 0.996,
572
+ "reasoning_bytes": 70.084,
573
+ "flops_per_problem": 4110875394.048
574
+ },
575
+ "sumd.cmp8": {
576
+ "acc": 0.954,
577
+ "reasoning_bytes": 71.416,
578
+ "flops_per_problem": 4213775351.808
579
+ },
580
+ "sumd.runmax6": {
581
+ "acc": 1.0,
582
+ "reasoning_bytes": 67.542,
583
+ "flops_per_problem": 3954449793.024
584
+ },
585
+ "sumd.runmax7": {
586
+ "acc": 0.998,
587
+ "reasoning_bytes": 75.48,
588
+ "flops_per_problem": 4349179330.56
589
+ },
590
+ "sumd.runmax8": {
591
+ "acc": 0.52,
592
+ "reasoning_bytes": 78.814,
593
+ "flops_per_problem": 4540582084.608
594
+ },
595
+ "sumd.count6": {
596
+ "acc": 1.0,
597
+ "reasoning_bytes": 66.428,
598
+ "flops_per_problem": 3949415202.816
599
+ },
600
+ "sumd.count7": {
601
+ "acc": 0.994,
602
+ "reasoning_bytes": 74.444,
603
+ "flops_per_problem": 4347589459.968
604
+ },
605
+ "sumd.count8": {
606
+ "acc": 0.454,
607
+ "reasoning_bytes": 75.612,
608
+ "flops_per_problem": 4443335000.064
609
+ },
610
+ "sumd.find6": {
611
+ "acc": 1.0,
612
+ "reasoning_bytes": 66.24,
613
+ "flops_per_problem": 3896949473.28
614
+ },
615
+ "sumd.find7": {
616
+ "acc": 1.0,
617
+ "reasoning_bytes": 74.43,
618
+ "flops_per_problem": 4302808104.96
619
+ },
620
+ "sumd.find8": {
621
+ "acc": 0.784,
622
+ "reasoning_bytes": 76.698,
623
+ "flops_per_problem": 4447133024.256
624
+ },
625
+ "sumd.rev6": {
626
+ "acc": 1.0,
627
+ "reasoning_bytes": 61.544,
628
+ "flops_per_problem": 3601145217.024
629
+ },
630
+ "sumd.rev7": {
631
+ "acc": 1.0,
632
+ "reasoning_bytes": 69.48,
633
+ "flops_per_problem": 3995874754.56
634
+ },
635
+ "sumd.rev8": {
636
+ "acc": 0.95,
637
+ "reasoning_bytes": 77.31,
638
+ "flops_per_problem": 4385834680.32
639
+ },
640
+ "sumd.sumd6": {
641
+ "acc": 0.996,
642
+ "reasoning_bytes": 65.744,
643
+ "flops_per_problem": 3795639386.112
644
+ },
645
+ "sumd.sumd7": {
646
+ "acc": 0.992,
647
+ "reasoning_bytes": 73.738,
648
+ "flops_per_problem": 4195315187.712
649
+ },
650
+ "sumd.sumd8": {
651
+ "acc": 0.672,
652
+ "reasoning_bytes": 78.888,
653
+ "flops_per_problem": 4467448037.376
654
+ },
655
+ "sumd.inc6": {
656
+ "acc": 1.0,
657
+ "reasoning_bytes": 68.538,
658
+ "flops_per_problem": 3910021742.592
659
+ },
660
+ "sumd.inc7": {
661
+ "acc": 0.972,
662
+ "reasoning_bytes": 76.48,
663
+ "flops_per_problem": 4305104584.704
664
+ },
665
+ "sumd.inc8": {
666
+ "acc": 0.648,
667
+ "reasoning_bytes": 82.42,
668
+ "flops_per_problem": 4611507978.24
669
+ },
670
+ "inc.cmp6": {
671
+ "acc": 1.0,
672
+ "reasoning_bytes": 78.624,
673
+ "flops_per_problem": 4576354172.928
674
+ },
675
+ "inc.cmp7": {
676
+ "acc": 1.0,
677
+ "reasoning_bytes": 89.77,
678
+ "flops_per_problem": 5156921917.44
679
+ },
680
+ "inc.cmp8": {
681
+ "acc": 0.982,
682
+ "reasoning_bytes": 99.694,
683
+ "flops_per_problem": 5683522387.968
684
+ },
685
+ "inc.runmax6": {
686
+ "acc": 1.0,
687
+ "reasoning_bytes": 104.0,
688
+ "flops_per_problem": 5520384000.0
689
+ },
690
+ "inc.runmax7": {
691
+ "acc": 1.0,
692
+ "reasoning_bytes": 120.0,
693
+ "flops_per_problem": 6271156224.0
694
+ },
695
+ "inc.runmax8": {
696
+ "acc": 0.902,
697
+ "reasoning_bytes": 129.406,
698
+ "flops_per_problem": 6730717151.232
699
+ },
700
+ "inc.count6": {
701
+ "acc": 0.98,
702
+ "reasoning_bytes": 99.0,
703
+ "flops_per_problem": 5343731712.0
704
+ },
705
+ "inc.count7": {
706
+ "acc": 0.898,
707
+ "reasoning_bytes": 113.866,
708
+ "flops_per_problem": 6044423012.352
709
+ },
710
+ "inc.count8": {
711
+ "acc": 0.754,
712
+ "reasoning_bytes": 114.088,
713
+ "flops_per_problem": 6098390286.336
714
+ },
715
+ "inc.find6": {
716
+ "acc": 0.998,
717
+ "reasoning_bytes": 91.88,
718
+ "flops_per_problem": 4985127567.36
719
+ },
720
+ "inc.find7": {
721
+ "acc": 0.894,
722
+ "reasoning_bytes": 102.828,
723
+ "flops_per_problem": 5512787951.616
724
+ },
725
+ "inc.find8": {
726
+ "acc": 0.824,
727
+ "reasoning_bytes": 106.948,
728
+ "flops_per_problem": 5738902880.256
729
+ },
730
+ "inc.rev6": {
731
+ "acc": 0.996,
732
+ "reasoning_bytes": 104.002,
733
+ "flops_per_problem": 5608798470.144
734
+ },
735
+ "inc.rev7": {
736
+ "acc": 0.9,
737
+ "reasoning_bytes": 123.996,
738
+ "flops_per_problem": 6580121075.712
739
+ },
740
+ "inc.rev8": {
741
+ "acc": 0.01,
742
+ "reasoning_bytes": 137.2,
743
+ "flops_per_problem": 7226138492.928
744
+ },
745
+ "inc.sumd6": {
746
+ "acc": 0.996,
747
+ "reasoning_bytes": 109.558,
748
+ "flops_per_problem": 5721590956.032
749
+ },
750
+ "inc.sumd7": {
751
+ "acc": 0.88,
752
+ "reasoning_bytes": 127.46,
753
+ "flops_per_problem": 6556273016.832
754
+ },
755
+ "inc.sumd8": {
756
+ "acc": 0.14,
757
+ "reasoning_bytes": 136.298,
758
+ "flops_per_problem": 6991367602.176
759
+ },
760
+ "inc.inc6": {
761
+ "acc": 1.0,
762
+ "reasoning_bytes": 113.0,
763
+ "flops_per_problem": 6006177792.0
764
+ },
765
+ "inc.inc7": {
766
+ "acc": 0.96,
767
+ "reasoning_bytes": 131.0,
768
+ "flops_per_problem": 6889704210.432
769
+ },
770
+ "inc.inc8": {
771
+ "acc": 0.096,
772
+ "reasoning_bytes": 142.294,
773
+ "flops_per_problem": 7458701230.08
774
+ },
775
+ "fsf6": {
776
+ "acc": 0.98,
777
+ "actions": 3.49,
778
+ "invalid": 0.034,
779
+ "reasoning_bytes": 42.172,
780
+ "flops_per_problem": 2810537902.08
781
+ },
782
+ "fsf7": {
783
+ "acc": 0.992,
784
+ "actions": 4.004,
785
+ "invalid": 0.056,
786
+ "reasoning_bytes": 48.584,
787
+ "flops_per_problem": 3174971572.224
788
+ },
789
+ "fsf8": {
790
+ "acc": 0.99,
791
+ "actions": 4.234,
792
+ "invalid": 0.052,
793
+ "reasoning_bytes": 53.68,
794
+ "flops_per_problem": 3478813507.584
795
+ },
796
+ "fs2mm6": {
797
+ "acc": 0.95,
798
+ "actions": 8.934,
799
+ "invalid": 0.128,
800
+ "reasoning_bytes": 112.492,
801
+ "flops_per_problem": 6647160619.008
802
+ },
803
+ "fs2mm7": {
804
+ "acc": 0.958,
805
+ "actions": 9.972,
806
+ "invalid": 0.126,
807
+ "reasoning_bytes": 127.926,
808
+ "flops_per_problem": 7463117537.28
809
+ },
810
+ "fs2mm8": {
811
+ "acc": 0.926,
812
+ "actions": 10.994,
813
+ "invalid": 0.212,
814
+ "reasoning_bytes": 143.69,
815
+ "flops_per_problem": 8279692738.56
816
+ },
817
+ "fs2md6": {
818
+ "acc": 0.0,
819
+ "actions": 9.012,
820
+ "invalid": 0.134,
821
+ "reasoning_bytes": 110.124,
822
+ "flops_per_problem": 6446306967.552
823
+ },
824
+ "fs2md7": {
825
+ "acc": 0.008,
826
+ "actions": 10.78,
827
+ "invalid": 0.362,
828
+ "reasoning_bytes": 135.712,
829
+ "flops_per_problem": 7745496219.648
830
+ },
831
+ "fs2md8": {
832
+ "acc": 0.022,
833
+ "actions": 13.636,
834
+ "invalid": 0.772,
835
+ "reasoning_bytes": 173.148,
836
+ "flops_per_problem": 9631259394.048
837
+ },
838
+ "fs2mc6": {
839
+ "acc": 1.0,
840
+ "actions": 5.338,
841
+ "invalid": 0.038,
842
+ "reasoning_bytes": 61.438,
843
+ "flops_per_problem": 4167669104.64
844
+ },
845
+ "fs2mc7": {
846
+ "acc": 1.0,
847
+ "actions": 5.94,
848
+ "invalid": 0.042,
849
+ "reasoning_bytes": 70.206,
850
+ "flops_per_problem": 4649134915.584
851
+ },
852
+ "fs2mc8": {
853
+ "acc": 0.996,
854
+ "actions": 6.492,
855
+ "invalid": 0.082,
856
+ "reasoning_bytes": 77.59,
857
+ "flops_per_problem": 5064267792.384
858
+ },
859
+ "fs2dm6": {
860
+ "acc": 0.966,
861
+ "actions": 8.968,
862
+ "invalid": 0.136,
863
+ "reasoning_bytes": 108.998,
864
+ "flops_per_problem": 6396314370.048
865
+ },
866
+ "fs2dm7": {
867
+ "acc": 0.966,
868
+ "actions": 10.138,
869
+ "invalid": 0.124,
870
+ "reasoning_bytes": 125.168,
871
+ "flops_per_problem": 7246541832.192
872
+ },
873
+ "fs2dm8": {
874
+ "acc": 0.962,
875
+ "actions": 11.432,
876
+ "invalid": 0.216,
877
+ "reasoning_bytes": 145.306,
878
+ "flops_per_problem": 8278367846.4
879
+ },
880
+ "fs2dd6": {
881
+ "acc": 0.15,
882
+ "actions": 9.286,
883
+ "invalid": 0.138,
884
+ "reasoning_bytes": 112.748,
885
+ "flops_per_problem": 6536046329.856
886
+ },
887
+ "fs2dd7": {
888
+ "acc": 0.132,
889
+ "actions": 10.312,
890
+ "invalid": 0.198,
891
+ "reasoning_bytes": 128.146,
892
+ "flops_per_problem": 7341139132.416
893
+ },
894
+ "fs2dd8": {
895
+ "acc": 0.098,
896
+ "actions": 11.582,
897
+ "invalid": 0.324,
898
+ "reasoning_bytes": 146.874,
899
+ "flops_per_problem": 8307868778.496
900
+ },
901
+ "fs2dc6": {
902
+ "acc": 1.0,
903
+ "actions": 5.534,
904
+ "invalid": 0.07,
905
+ "reasoning_bytes": 62.476,
906
+ "flops_per_problem": 4143467741.184
907
+ },
908
+ "fs2dc7": {
909
+ "acc": 1.0,
910
+ "actions": 6.012,
911
+ "invalid": 0.056,
912
+ "reasoning_bytes": 68.528,
913
+ "flops_per_problem": 4490766139.392
914
+ },
915
+ "fs2dc8": {
916
+ "acc": 0.998,
917
+ "actions": 6.632,
918
+ "invalid": 0.074,
919
+ "reasoning_bytes": 78.52,
920
+ "flops_per_problem": 5029555617.792
921
+ },
922
+ "fs2cm6": {
923
+ "acc": 0.28,
924
+ "actions": 6.496,
925
+ "invalid": 0.872,
926
+ "reasoning_bytes": 79.424,
927
+ "flops_per_problem": 5004559319.04
928
+ },
929
+ "fs2cm7": {
930
+ "acc": 0.334,
931
+ "actions": 6.712,
932
+ "invalid": 0.856,
933
+ "reasoning_bytes": 82.962,
934
+ "flops_per_problem": 5214952194.048
935
+ },
936
+ "fs2cm8": {
937
+ "acc": 0.38,
938
+ "actions": 7.086,
939
+ "invalid": 1.002,
940
+ "reasoning_bytes": 89.37,
941
+ "flops_per_problem": 5582565605.376
942
+ },
943
+ "fs2cd6": {
944
+ "acc": 1.0,
945
+ "actions": 5.578,
946
+ "invalid": 0.064,
947
+ "reasoning_bytes": 64.284,
948
+ "flops_per_problem": 4249017483.264
949
+ },
950
+ "fs2cd7": {
951
+ "acc": 0.994,
952
+ "actions": 5.992,
953
+ "invalid": 0.054,
954
+ "reasoning_bytes": 70.66,
955
+ "flops_per_problem": 4607179997.184
956
+ },
957
+ "fs2cd8": {
958
+ "acc": 0.998,
959
+ "actions": 6.538,
960
+ "invalid": 0.078,
961
+ "reasoning_bytes": 79.198,
962
+ "flops_per_problem": 5076810104.832
963
+ },
964
+ "fs2cc6": {
965
+ "acc": 1.0,
966
+ "actions": 2.0,
967
+ "invalid": 0.024,
968
+ "reasoning_bytes": 13.0,
969
+ "flops_per_problem": 1815543889.92
970
+ },
971
+ "fs2cc7": {
972
+ "acc": 1.0,
973
+ "actions": 2.0,
974
+ "invalid": 0.016,
975
+ "reasoning_bytes": 13.0,
976
+ "flops_per_problem": 1860325244.928
977
+ },
978
+ "fs2cc8": {
979
+ "acc": 1.0,
980
+ "actions": 2.0,
981
+ "invalid": 0.02,
982
+ "reasoning_bytes": 13.0,
983
+ "flops_per_problem": 1908816297.984
984
+ }
985
+ }