lamshiu commited on
Commit
b4f8eb0
·
verified ·
1 Parent(s): c38c201

Upload folder using huggingface_hub

Browse files
5m-s0-bpe4096-step-bytes3-v16-wfso4-none-docmask/ckpt.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2bee6792b8d56f10f6b02982002b1946166e07515865250101afe6de8bdbe079
3
+ size 23171663
5m-s0-bpe4096-step-bytes3-v16-wfso4-none-docmask/config.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "size": "5m",
3
+ "name": "5m-s0-bpe4096-step-bytes3-v16-wfso4-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
+ "fso=4"
20
+ ],
21
+ "holdout": [],
22
+ "dataset": "v16",
23
+ "vocab_size": 4096,
24
+ "n_layer": 6,
25
+ "n_head": 4,
26
+ "d_model": 256,
27
+ "loops": 1,
28
+ "params": 5787648,
29
+ "flops_per_token": 44163072,
30
+ "steps": 12207,
31
+ "n_docs": 1173417,
32
+ "stream_tokens": 190321453
33
+ }
5m-s0-bpe4096-step-bytes3-v16-wfso4-none-docmask/log.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
5m-s0-bpe4096-step-bytes3-v16-wfso4-none-docmask/summary.json ADDED
@@ -0,0 +1,796 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "params": 5787648,
3
+ "tokens": 399998976,
4
+ "train_flops": 17665183577014272,
5
+ "wall_s": 683.7998633384705,
6
+ "peak_mem_gb": 2.520964096,
7
+ "final_val_loss": 1.7814195990562438,
8
+ "final_val_bpb": 0.642172824061847,
9
+ "bytes_seen": 1600839241.8284984,
10
+ "add8": {
11
+ "acc": 0.996,
12
+ "reasoning_bytes": 79.0,
13
+ "flops_per_problem": 4928687161.344
14
+ },
15
+ "add9": {
16
+ "acc": 0.76,
17
+ "reasoning_bytes": 89.004,
18
+ "flops_per_problem": 5503248728.064
19
+ },
20
+ "add10": {
21
+ "acc": 0.0,
22
+ "reasoning_bytes": 98.648,
23
+ "flops_per_problem": 6061558284.288
24
+ },
25
+ "reach14": {
26
+ "acc": 0.634,
27
+ "reasoning_bytes": 7.89,
28
+ "flops_per_problem": 4632264622.08
29
+ },
30
+ "reach16": {
31
+ "acc": 0.564,
32
+ "reasoning_bytes": 7.75,
33
+ "flops_per_problem": 5156038656.0
34
+ },
35
+ "reach18": {
36
+ "acc": 0.602,
37
+ "reasoning_bytes": 7.104,
38
+ "flops_per_problem": 5657112870.912
39
+ },
40
+ "parity24": {
41
+ "acc": 0.52,
42
+ "reasoning_bytes": 143.0,
43
+ "flops_per_problem": 7861026816.0
44
+ },
45
+ "parity28": {
46
+ "acc": 0.354,
47
+ "reasoning_bytes": 166.952,
48
+ "flops_per_problem": 9095473004.544
49
+ },
50
+ "parity32": {
51
+ "acc": 0.516,
52
+ "reasoning_bytes": 188.48,
53
+ "flops_per_problem": 10222867906.56
54
+ },
55
+ "cmp8": {
56
+ "acc": 1.0,
57
+ "reasoning_bytes": 5.828,
58
+ "flops_per_problem": 1361459183.616
59
+ },
60
+ "cmp9": {
61
+ "acc": 1.0,
62
+ "reasoning_bytes": 5.72,
63
+ "flops_per_problem": 1445015715.84
64
+ },
65
+ "cmp10": {
66
+ "acc": 0.996,
67
+ "reasoning_bytes": 5.78,
68
+ "flops_per_problem": 1535991644.16
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": 82.808,
78
+ "flops_per_problem": 4761132466.176
79
+ },
80
+ "runmax16": {
81
+ "acc": 0.978,
82
+ "reasoning_bytes": 89.786,
83
+ "flops_per_problem": 5157628526.592
84
+ },
85
+ "count16": {
86
+ "acc": 0.998,
87
+ "reasoning_bytes": 79.0,
88
+ "flops_per_problem": 4725448704.0
89
+ },
90
+ "count18": {
91
+ "acc": 0.978,
92
+ "reasoning_bytes": 88.55,
93
+ "flops_per_problem": 5235532185.6
94
+ },
95
+ "count20": {
96
+ "acc": 0.826,
97
+ "reasoning_bytes": 95.83,
98
+ "flops_per_problem": 5645365493.76
99
+ },
100
+ "brackets16": {
101
+ "acc": 0.998,
102
+ "reasoning_bytes": 53.468,
103
+ "flops_per_problem": 3642040221.696
104
+ },
105
+ "brackets18": {
106
+ "acc": 0.982,
107
+ "reasoning_bytes": 56.766,
108
+ "flops_per_problem": 3876016177.152
109
+ },
110
+ "brackets20": {
111
+ "acc": 0.948,
112
+ "reasoning_bytes": 63.206,
113
+ "flops_per_problem": 4249017483.264
114
+ },
115
+ "sumd8": {
116
+ "acc": 0.99,
117
+ "reasoning_bytes": 58.566,
118
+ "flops_per_problem": 3381301444.608
119
+ },
120
+ "sumd9": {
121
+ "acc": 0.89,
122
+ "reasoning_bytes": 66.366,
123
+ "flops_per_problem": 3770024804.352
124
+ },
125
+ "sumd10": {
126
+ "acc": 0.556,
127
+ "reasoning_bytes": 72.68,
128
+ "flops_per_problem": 4092591882.24
129
+ },
130
+ "inc8": {
131
+ "acc": 0.998,
132
+ "reasoning_bytes": 62.98,
133
+ "flops_per_problem": 3796522647.552
134
+ },
135
+ "inc9": {
136
+ "acc": 0.862,
137
+ "reasoning_bytes": 70.968,
138
+ "flops_per_problem": 4238594998.272
139
+ },
140
+ "inc10": {
141
+ "acc": 0.028,
142
+ "reasoning_bytes": 78.632,
143
+ "flops_per_problem": 4646926761.984
144
+ },
145
+ "add.cmp6": {
146
+ "acc": 1.0,
147
+ "reasoning_bytes": 92.622,
148
+ "flops_per_problem": 5532043051.008
149
+ },
150
+ "add.cmp7": {
151
+ "acc": 0.994,
152
+ "reasoning_bytes": 105.524,
153
+ "flops_per_problem": 6231586111.488
154
+ },
155
+ "add.cmp8": {
156
+ "acc": 0.98,
157
+ "reasoning_bytes": 118.658,
158
+ "flops_per_problem": 6945967964.16
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": 138.848,
168
+ "flops_per_problem": 7456846381.056
169
+ },
170
+ "add.runmax8": {
171
+ "acc": 0.558,
172
+ "reasoning_bytes": 156.476,
173
+ "flops_per_problem": 8323679158.272
174
+ },
175
+ "add.count6": {
176
+ "acc": 0.99,
177
+ "reasoning_bytes": 115.648,
178
+ "flops_per_problem": 6388100038.656
179
+ },
180
+ "add.count7": {
181
+ "acc": 0.932,
182
+ "reasoning_bytes": 132.088,
183
+ "flops_per_problem": 7202467086.336
184
+ },
185
+ "add.count8": {
186
+ "acc": 0.35,
187
+ "reasoning_bytes": 145.598,
188
+ "flops_per_problem": 7887436333.056
189
+ },
190
+ "add.find6": {
191
+ "acc": 0.98,
192
+ "reasoning_bytes": 107.696,
193
+ "flops_per_problem": 5992752218.112
194
+ },
195
+ "add.find7": {
196
+ "acc": 0.94,
197
+ "reasoning_bytes": 121.248,
198
+ "flops_per_problem": 6679576313.856
199
+ },
200
+ "add.find8": {
201
+ "acc": 0.486,
202
+ "reasoning_bytes": 137.234,
203
+ "flops_per_problem": 7473893326.848
204
+ },
205
+ "add.rev6": {
206
+ "acc": 0.99,
207
+ "reasoning_bytes": 123.248,
208
+ "flops_per_problem": 6794576953.344
209
+ },
210
+ "add.rev7": {
211
+ "acc": 0.714,
212
+ "reasoning_bytes": 145.866,
213
+ "flops_per_problem": 7924974944.256
214
+ },
215
+ "add.rev8": {
216
+ "acc": 0.05,
217
+ "reasoning_bytes": 162.142,
218
+ "flops_per_problem": 8742698385.408
219
+ },
220
+ "add.sumd6": {
221
+ "acc": 0.984,
222
+ "reasoning_bytes": 126.476,
223
+ "flops_per_problem": 6777794985.984
224
+ },
225
+ "add.sumd7": {
226
+ "acc": 0.908,
227
+ "reasoning_bytes": 146.2,
228
+ "flops_per_problem": 7737458540.544
229
+ },
230
+ "add.sumd8": {
231
+ "acc": 0.188,
232
+ "reasoning_bytes": 165.146,
233
+ "flops_per_problem": 8663028203.52
234
+ },
235
+ "add.inc6": {
236
+ "acc": 1.0,
237
+ "reasoning_bytes": 131.04,
238
+ "flops_per_problem": 7138695610.368
239
+ },
240
+ "add.inc7": {
241
+ "acc": 0.962,
242
+ "reasoning_bytes": 151.06,
243
+ "flops_per_problem": 8155417853.952
244
+ },
245
+ "add.inc8": {
246
+ "acc": 0.104,
247
+ "reasoning_bytes": 169.122,
248
+ "flops_per_problem": 9082489061.376
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": 1.0,
257
+ "reasoning_bytes": 98.7,
258
+ "flops_per_problem": 5727950438.4
259
+ },
260
+ "runmax.cmp16": {
261
+ "acc": 0.994,
262
+ "reasoning_bytes": 105.042,
263
+ "flops_per_problem": 6096358785.024
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": 1.0,
272
+ "reasoning_bytes": 99.544,
273
+ "flops_per_problem": 5809387143.168
274
+ },
275
+ "runmax.runmax16": {
276
+ "acc": 0.974,
277
+ "reasoning_bytes": 106.012,
278
+ "flops_per_problem": 6183360036.864
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": 99.828,
288
+ "flops_per_problem": 5866092527.616
289
+ },
290
+ "runmax.count16": {
291
+ "acc": 1.0,
292
+ "reasoning_bytes": 106.024,
293
+ "flops_per_problem": 6228053065.728
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": 99.64,
303
+ "flops_per_problem": 5813626798.08
304
+ },
305
+ "runmax.find16": {
306
+ "acc": 0.99,
307
+ "reasoning_bytes": 106.012,
308
+ "flops_per_problem": 6183360036.864
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": 1.0,
317
+ "reasoning_bytes": 94.88,
318
+ "flops_per_problem": 5470921359.36
319
+ },
320
+ "runmax.rev16": {
321
+ "acc": 0.976,
322
+ "reasoning_bytes": 101.092,
323
+ "flops_per_problem": 5833588506.624
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": 0.998,
332
+ "reasoning_bytes": 96.392,
333
+ "flops_per_problem": 5581858996.224
334
+ },
335
+ "runmax.sumd16": {
336
+ "acc": 0.972,
337
+ "reasoning_bytes": 104.012,
338
+ "flops_per_problem": 6006707748.864
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": 0.998,
347
+ "reasoning_bytes": 98.556,
348
+ "flops_per_problem": 5667800334.336
349
+ },
350
+ "runmax.inc16": {
351
+ "acc": 0.972,
352
+ "reasoning_bytes": 105.036,
353
+ "flops_per_problem": 6043981381.632
354
+ },
355
+ "count.cmp16": {
356
+ "acc": 0.998,
357
+ "reasoning_bytes": 95.0,
358
+ "flops_per_problem": 5697036288.0
359
+ },
360
+ "count.cmp18": {
361
+ "acc": 0.972,
362
+ "reasoning_bytes": 103.54,
363
+ "flops_per_problem": 6162603393.024
364
+ },
365
+ "count.cmp20": {
366
+ "acc": 0.952,
367
+ "reasoning_bytes": 110.04,
368
+ "flops_per_problem": 6538077831.168
369
+ },
370
+ "count.runmax16": {
371
+ "acc": 0.998,
372
+ "reasoning_bytes": 96.0,
373
+ "flops_per_problem": 5785362432.0
374
+ },
375
+ "count.runmax18": {
376
+ "acc": 0.96,
377
+ "reasoning_bytes": 104.94,
378
+ "flops_per_problem": 6268506439.68
379
+ },
380
+ "count.runmax20": {
381
+ "acc": 0.766,
382
+ "reasoning_bytes": 111.15,
383
+ "flops_per_problem": 6631085260.8
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": 0.998,
392
+ "reasoning_bytes": 105.296,
393
+ "flops_per_problem": 6328391565.312
394
+ },
395
+ "count.count20": {
396
+ "acc": 0.998,
397
+ "reasoning_bytes": 111.128,
398
+ "flops_per_problem": 6674276745.216
399
+ },
400
+ "count.find16": {
401
+ "acc": 0.998,
402
+ "reasoning_bytes": 96.0,
403
+ "flops_per_problem": 5785362432.0
404
+ },
405
+ "count.find18": {
406
+ "acc": 0.98,
407
+ "reasoning_bytes": 104.7,
408
+ "flops_per_problem": 6257907302.4
409
+ },
410
+ "count.find20": {
411
+ "acc": 0.948,
412
+ "reasoning_bytes": 111.11,
413
+ "flops_per_problem": 6629318737.92
414
+ },
415
+ "count.rev16": {
416
+ "acc": 0.998,
417
+ "reasoning_bytes": 91.0,
418
+ "flops_per_problem": 5432057856.0
419
+ },
420
+ "count.rev18": {
421
+ "acc": 0.958,
422
+ "reasoning_bytes": 100.08,
423
+ "flops_per_problem": 5921384693.76
424
+ },
425
+ "count.rev20": {
426
+ "acc": 0.77,
427
+ "reasoning_bytes": 106.23,
428
+ "flops_per_problem": 6281313730.56
429
+ },
430
+ "count.sumd16": {
431
+ "acc": 0.998,
432
+ "reasoning_bytes": 94.0,
433
+ "flops_per_problem": 5608710144.0
434
+ },
435
+ "count.sumd18": {
436
+ "acc": 0.94,
437
+ "reasoning_bytes": 102.55,
438
+ "flops_per_problem": 6074630553.6
439
+ },
440
+ "count.sumd20": {
441
+ "acc": 0.77,
442
+ "reasoning_bytes": 109.12,
443
+ "flops_per_problem": 6453108080.64
444
+ },
445
+ "count.inc16": {
446
+ "acc": 0.996,
447
+ "reasoning_bytes": 95.0,
448
+ "flops_per_problem": 5608710144.0
449
+ },
450
+ "count.inc18": {
451
+ "acc": 0.966,
452
+ "reasoning_bytes": 104.07,
453
+ "flops_per_problem": 6097772003.328
454
+ },
455
+ "count.inc20": {
456
+ "acc": 0.78,
457
+ "reasoning_bytes": 110.43,
458
+ "flops_per_problem": 6466798632.96
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": 0.964,
467
+ "reasoning_bytes": 129.18,
468
+ "flops_per_problem": 7339019304.96
469
+ },
470
+ "parity.cmp24": {
471
+ "acc": 0.938,
472
+ "reasoning_bytes": 147.564,
473
+ "flops_per_problem": 8327565508.608
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": 0.834,
482
+ "reasoning_bytes": 132.724,
483
+ "flops_per_problem": 7539696304.128
484
+ },
485
+ "parity.runmax24": {
486
+ "acc": 0.726,
487
+ "reasoning_bytes": 153.436,
488
+ "flops_per_problem": 8631054139.392
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": 130.96,
498
+ "flops_per_problem": 7505955717.12
499
+ },
500
+ "parity.count24": {
501
+ "acc": 0.998,
502
+ "reasoning_bytes": 150.7,
503
+ "flops_per_problem": 8554387046.4
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": 0.94,
512
+ "reasoning_bytes": 132.04,
513
+ "flops_per_problem": 7509488762.88
514
+ },
515
+ "parity.find24": {
516
+ "acc": 0.938,
517
+ "reasoning_bytes": 152.896,
518
+ "flops_per_problem": 8607206080.512
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": 0.868,
527
+ "reasoning_bytes": 129.596,
528
+ "flops_per_problem": 7269064998.912
529
+ },
530
+ "parity.rev24": {
531
+ "acc": 0.688,
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": 0.778,
542
+ "reasoning_bytes": 129.752,
543
+ "flops_per_problem": 7320117510.144
544
+ },
545
+ "parity.sumd24": {
546
+ "acc": 0.67,
547
+ "reasoning_bytes": 151.328,
548
+ "flops_per_problem": 8449632239.616
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": 0.79,
557
+ "reasoning_bytes": 131.856,
558
+ "flops_per_problem": 7368873541.632
559
+ },
560
+ "parity.inc24": {
561
+ "acc": 0.668,
562
+ "reasoning_bytes": 152.88,
563
+ "flops_per_problem": 8474010255.36
564
+ },
565
+ "sumd.cmp6": {
566
+ "acc": 0.998,
567
+ "reasoning_bytes": 62.088,
568
+ "flops_per_problem": 3713407746.048
569
+ },
570
+ "sumd.cmp7": {
571
+ "acc": 0.992,
572
+ "reasoning_bytes": 70.084,
573
+ "flops_per_problem": 4110875394.048
574
+ },
575
+ "sumd.cmp8": {
576
+ "acc": 0.958,
577
+ "reasoning_bytes": 74.796,
578
+ "flops_per_problem": 4363046535.168
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": 0.996,
587
+ "reasoning_bytes": 75.48,
588
+ "flops_per_problem": 4349179330.56
589
+ },
590
+ "sumd.runmax8": {
591
+ "acc": 0.758,
592
+ "reasoning_bytes": 81.49,
593
+ "flops_per_problem": 4658762465.28
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": 0.996,
602
+ "reasoning_bytes": 74.428,
603
+ "flops_per_problem": 4346882850.816
604
+ },
605
+ "sumd.count8": {
606
+ "acc": 0.676,
607
+ "reasoning_bytes": 79.316,
608
+ "flops_per_problem": 4606915018.752
609
+ },
610
+ "sumd.find6": {
611
+ "acc": 1.0,
612
+ "reasoning_bytes": 66.26,
613
+ "flops_per_problem": 3897832734.72
614
+ },
615
+ "sumd.find7": {
616
+ "acc": 1.0,
617
+ "reasoning_bytes": 74.412,
618
+ "flops_per_problem": 4302013169.664
619
+ },
620
+ "sumd.find8": {
621
+ "acc": 0.866,
622
+ "reasoning_bytes": 79.372,
623
+ "flops_per_problem": 4565225078.784
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": 0.998,
632
+ "reasoning_bytes": 69.48,
633
+ "flops_per_problem": 3995874754.56
634
+ },
635
+ "sumd.rev8": {
636
+ "acc": 0.956,
637
+ "reasoning_bytes": 77.408,
638
+ "flops_per_problem": 4389897682.944
639
+ },
640
+ "sumd.sumd6": {
641
+ "acc": 1.0,
642
+ "reasoning_bytes": 65.744,
643
+ "flops_per_problem": 3795639386.112
644
+ },
645
+ "sumd.sumd7": {
646
+ "acc": 0.986,
647
+ "reasoning_bytes": 73.728,
648
+ "flops_per_problem": 4194696904.704
649
+ },
650
+ "sumd.sumd8": {
651
+ "acc": 0.696,
652
+ "reasoning_bytes": 80.102,
653
+ "flops_per_problem": 4520355397.632
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": 0.97,
662
+ "reasoning_bytes": 76.486,
663
+ "flops_per_problem": 4305281236.992
664
+ },
665
+ "sumd.inc8": {
666
+ "acc": 0.718,
667
+ "reasoning_bytes": 83.436,
668
+ "flops_per_problem": 4657084268.544
669
+ },
670
+ "inc.cmp6": {
671
+ "acc": 0.998,
672
+ "reasoning_bytes": 78.624,
673
+ "flops_per_problem": 4576354172.928
674
+ },
675
+ "inc.cmp7": {
676
+ "acc": 1.0,
677
+ "reasoning_bytes": 89.818,
678
+ "flops_per_problem": 5159041744.896
679
+ },
680
+ "inc.cmp8": {
681
+ "acc": 0.982,
682
+ "reasoning_bytes": 100.692,
683
+ "flops_per_problem": 5727862112.256
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": 0.992,
692
+ "reasoning_bytes": 120.0,
693
+ "flops_per_problem": 6271156224.0
694
+ },
695
+ "inc.runmax8": {
696
+ "acc": 0.906,
697
+ "reasoning_bytes": 133.786,
698
+ "flops_per_problem": 6924151406.592
699
+ },
700
+ "inc.count6": {
701
+ "acc": 0.99,
702
+ "reasoning_bytes": 99.0,
703
+ "flops_per_problem": 5343731712.0
704
+ },
705
+ "inc.count7": {
706
+ "acc": 0.844,
707
+ "reasoning_bytes": 114.002,
708
+ "flops_per_problem": 6050429190.144
709
+ },
710
+ "inc.count8": {
711
+ "acc": 0.786,
712
+ "reasoning_bytes": 114.06,
713
+ "flops_per_problem": 6097153720.32
714
+ },
715
+ "inc.find6": {
716
+ "acc": 1.0,
717
+ "reasoning_bytes": 91.88,
718
+ "flops_per_problem": 4985127567.36
719
+ },
720
+ "inc.find7": {
721
+ "acc": 0.924,
722
+ "reasoning_bytes": 102.672,
723
+ "flops_per_problem": 5505898512.384
724
+ },
725
+ "inc.find8": {
726
+ "acc": 0.636,
727
+ "reasoning_bytes": 113.272,
728
+ "flops_per_problem": 6020309975.04
729
+ },
730
+ "inc.rev6": {
731
+ "acc": 0.998,
732
+ "reasoning_bytes": 104.0,
733
+ "flops_per_problem": 5608710144.0
734
+ },
735
+ "inc.rev7": {
736
+ "acc": 0.912,
737
+ "reasoning_bytes": 123.97,
738
+ "flops_per_problem": 6578796183.552
739
+ },
740
+ "inc.rev8": {
741
+ "acc": 0.002,
742
+ "reasoning_bytes": 136.128,
743
+ "flops_per_problem": 7173672763.392
744
+ },
745
+ "inc.sumd6": {
746
+ "acc": 0.992,
747
+ "reasoning_bytes": 109.554,
748
+ "flops_per_problem": 5721414303.744
749
+ },
750
+ "inc.sumd7": {
751
+ "acc": 0.812,
752
+ "reasoning_bytes": 127.45,
753
+ "flops_per_problem": 6555478081.536
754
+ },
755
+ "inc.sumd8": {
756
+ "acc": 0.066,
757
+ "reasoning_bytes": 142.848,
758
+ "flops_per_problem": 7280459071.488
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": 0.976,
767
+ "reasoning_bytes": 131.0,
768
+ "flops_per_problem": 6889439232.0
769
+ },
770
+ "inc.inc8": {
771
+ "acc": 0.07,
772
+ "reasoning_bytes": 148.52,
773
+ "flops_per_problem": 7750530809.856
774
+ },
775
+ "fso6": {
776
+ "acc": 0.536,
777
+ "actions": 4.53,
778
+ "invalid": 0.334,
779
+ "reasoning_bytes": 21.406,
780
+ "flops_per_problem": 1967818162.176
781
+ },
782
+ "fso7": {
783
+ "acc": 0.494,
784
+ "actions": 5.162,
785
+ "invalid": 0.372,
786
+ "reasoning_bytes": 23.984,
787
+ "flops_per_problem": 2169290096.64
788
+ },
789
+ "fso8": {
790
+ "acc": 0.474,
791
+ "actions": 5.502,
792
+ "invalid": 0.34,
793
+ "reasoning_bytes": 25.306,
794
+ "flops_per_problem": 2303722487.808
795
+ }
796
+ }