Upload folder using huggingface_hub
Browse files
5m-s0-bpe4096-step-bytes3-v14-none-docmask/ckpt.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2f5094cf0af010f38a7f950d5daeb7f83f6561c258c36edfd0b212999da1fa9a
|
| 3 |
+
size 23171663
|
5m-s0-bpe4096-step-bytes3-v14-none-docmask/config.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"size": "5m",
|
| 3 |
+
"name": "5m-s0-bpe4096-step-bytes3-v14-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 |
+
"dataset": "v14",
|
| 21 |
+
"vocab_size": 4096,
|
| 22 |
+
"n_layer": 6,
|
| 23 |
+
"n_head": 4,
|
| 24 |
+
"d_model": 256,
|
| 25 |
+
"loops": 1,
|
| 26 |
+
"params": 5787648,
|
| 27 |
+
"flops_per_token": 44163072,
|
| 28 |
+
"steps": 12207,
|
| 29 |
+
"n_docs": 1172574,
|
| 30 |
+
"stream_tokens": 193496432
|
| 31 |
+
}
|
5m-s0-bpe4096-step-bytes3-v14-none-docmask/log.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
5m-s0-bpe4096-step-bytes3-v14-none-docmask/summary.json
ADDED
|
@@ -0,0 +1,796 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"params": 5787648,
|
| 3 |
+
"tokens": 399998976,
|
| 4 |
+
"train_flops": 17665183577014272,
|
| 5 |
+
"wall_s": 721.5269265174866,
|
| 6 |
+
"peak_mem_gb": 2.520964096,
|
| 7 |
+
"final_val_loss": 1.7836309611797332,
|
| 8 |
+
"final_val_bpb": 0.6429699841810109,
|
| 9 |
+
"bytes_seen": 1600839241.8284984,
|
| 10 |
+
"add8": {
|
| 11 |
+
"acc": 0.998,
|
| 12 |
+
"reasoning_bytes": 79.0,
|
| 13 |
+
"flops_per_problem": 4928775487.488
|
| 14 |
+
},
|
| 15 |
+
"add9": {
|
| 16 |
+
"acc": 0.672,
|
| 17 |
+
"reasoning_bytes": 89.0,
|
| 18 |
+
"flops_per_problem": 5502983749.632
|
| 19 |
+
},
|
| 20 |
+
"add10": {
|
| 21 |
+
"acc": 0.0,
|
| 22 |
+
"reasoning_bytes": 98.998,
|
| 23 |
+
"flops_per_problem": 6076750381.056
|
| 24 |
+
},
|
| 25 |
+
"reach14": {
|
| 26 |
+
"acc": 0.682,
|
| 27 |
+
"reasoning_bytes": 8.034,
|
| 28 |
+
"flops_per_problem": 4638624104.448
|
| 29 |
+
},
|
| 30 |
+
"reach16": {
|
| 31 |
+
"acc": 0.576,
|
| 32 |
+
"reasoning_bytes": 7.826,
|
| 33 |
+
"flops_per_problem": 5159395049.472
|
| 34 |
+
},
|
| 35 |
+
"reach18": {
|
| 36 |
+
"acc": 0.578,
|
| 37 |
+
"reasoning_bytes": 7.988,
|
| 38 |
+
"flops_per_problem": 5696506331.136
|
| 39 |
+
},
|
| 40 |
+
"parity24": {
|
| 41 |
+
"acc": 0.49,
|
| 42 |
+
"reasoning_bytes": 143.024,
|
| 43 |
+
"flops_per_problem": 7862086729.728
|
| 44 |
+
},
|
| 45 |
+
"parity28": {
|
| 46 |
+
"acc": 0.5,
|
| 47 |
+
"reasoning_bytes": 167.012,
|
| 48 |
+
"flops_per_problem": 9098122788.864
|
| 49 |
+
},
|
| 50 |
+
"parity32": {
|
| 51 |
+
"acc": 0.51,
|
| 52 |
+
"reasoning_bytes": 190.94,
|
| 53 |
+
"flops_per_problem": 10331509063.68
|
| 54 |
+
},
|
| 55 |
+
"cmp8": {
|
| 56 |
+
"acc": 0.998,
|
| 57 |
+
"reasoning_bytes": 5.78,
|
| 58 |
+
"flops_per_problem": 1359339356.16
|
| 59 |
+
},
|
| 60 |
+
"cmp9": {
|
| 61 |
+
"acc": 1.0,
|
| 62 |
+
"reasoning_bytes": 5.72,
|
| 63 |
+
"flops_per_problem": 1445015715.84
|
| 64 |
+
},
|
| 65 |
+
"cmp10": {
|
| 66 |
+
"acc": 0.994,
|
| 67 |
+
"reasoning_bytes": 5.78,
|
| 68 |
+
"flops_per_problem": 1536609927.168
|
| 69 |
+
},
|
| 70 |
+
"runmax12": {
|
| 71 |
+
"acc": 1.0,
|
| 72 |
+
"reasoning_bytes": 71.0,
|
| 73 |
+
"flops_per_problem": 4151328768.0
|
| 74 |
+
},
|
| 75 |
+
"runmax14": {
|
| 76 |
+
"acc": 0.998,
|
| 77 |
+
"reasoning_bytes": 83.0,
|
| 78 |
+
"flops_per_problem": 4769611776.0
|
| 79 |
+
},
|
| 80 |
+
"runmax16": {
|
| 81 |
+
"acc": 0.998,
|
| 82 |
+
"reasoning_bytes": 94.408,
|
| 83 |
+
"flops_per_problem": 5361750245.376
|
| 84 |
+
},
|
| 85 |
+
"count16": {
|
| 86 |
+
"acc": 0.998,
|
| 87 |
+
"reasoning_bytes": 78.99,
|
| 88 |
+
"flops_per_problem": 4725007073.28
|
| 89 |
+
},
|
| 90 |
+
"count18": {
|
| 91 |
+
"acc": 0.95,
|
| 92 |
+
"reasoning_bytes": 88.24,
|
| 93 |
+
"flops_per_problem": 5221841633.28
|
| 94 |
+
},
|
| 95 |
+
"count20": {
|
| 96 |
+
"acc": 0.782,
|
| 97 |
+
"reasoning_bytes": 94.88,
|
| 98 |
+
"flops_per_problem": 5603410575.36
|
| 99 |
+
},
|
| 100 |
+
"brackets16": {
|
| 101 |
+
"acc": 0.996,
|
| 102 |
+
"reasoning_bytes": 53.464,
|
| 103 |
+
"flops_per_problem": 3641951895.552
|
| 104 |
+
},
|
| 105 |
+
"brackets18": {
|
| 106 |
+
"acc": 0.974,
|
| 107 |
+
"reasoning_bytes": 56.682,
|
| 108 |
+
"flops_per_problem": 3872483131.392
|
| 109 |
+
},
|
| 110 |
+
"brackets20": {
|
| 111 |
+
"acc": 0.87,
|
| 112 |
+
"reasoning_bytes": 63.688,
|
| 113 |
+
"flops_per_problem": 4270304083.968
|
| 114 |
+
},
|
| 115 |
+
"sumd8": {
|
| 116 |
+
"acc": 0.982,
|
| 117 |
+
"reasoning_bytes": 58.566,
|
| 118 |
+
"flops_per_problem": 3381301444.608
|
| 119 |
+
},
|
| 120 |
+
"sumd9": {
|
| 121 |
+
"acc": 0.874,
|
| 122 |
+
"reasoning_bytes": 66.43,
|
| 123 |
+
"flops_per_problem": 3772851240.96
|
| 124 |
+
},
|
| 125 |
+
"sumd10": {
|
| 126 |
+
"acc": 0.618,
|
| 127 |
+
"reasoning_bytes": 74.418,
|
| 128 |
+
"flops_per_problem": 4169965584.384
|
| 129 |
+
},
|
| 130 |
+
"inc8": {
|
| 131 |
+
"acc": 0.998,
|
| 132 |
+
"reasoning_bytes": 63.0,
|
| 133 |
+
"flops_per_problem": 3797935865.856
|
| 134 |
+
},
|
| 135 |
+
"inc9": {
|
| 136 |
+
"acc": 0.716,
|
| 137 |
+
"reasoning_bytes": 71.0,
|
| 138 |
+
"flops_per_problem": 4239301607.424
|
| 139 |
+
},
|
| 140 |
+
"inc10": {
|
| 141 |
+
"acc": 0.016,
|
| 142 |
+
"reasoning_bytes": 78.84,
|
| 143 |
+
"flops_per_problem": 4656819290.112
|
| 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": 0.992,
|
| 152 |
+
"reasoning_bytes": 105.514,
|
| 153 |
+
"flops_per_problem": 6231144480.768
|
| 154 |
+
},
|
| 155 |
+
"add.cmp8": {
|
| 156 |
+
"acc": 0.942,
|
| 157 |
+
"reasoning_bytes": 117.988,
|
| 158 |
+
"flops_per_problem": 6921678274.56
|
| 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": 0.986,
|
| 167 |
+
"reasoning_bytes": 138.848,
|
| 168 |
+
"flops_per_problem": 7456846381.056
|
| 169 |
+
},
|
| 170 |
+
"add.runmax8": {
|
| 171 |
+
"acc": 0.682,
|
| 172 |
+
"reasoning_bytes": 156.284,
|
| 173 |
+
"flops_per_problem": 8317143023.616
|
| 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": 0.954,
|
| 182 |
+
"reasoning_bytes": 132.082,
|
| 183 |
+
"flops_per_problem": 7202202107.904
|
| 184 |
+
},
|
| 185 |
+
"add.count8": {
|
| 186 |
+
"acc": 0.426,
|
| 187 |
+
"reasoning_bytes": 147.246,
|
| 188 |
+
"flops_per_problem": 7963926773.76
|
| 189 |
+
},
|
| 190 |
+
"add.find6": {
|
| 191 |
+
"acc": 0.972,
|
| 192 |
+
"reasoning_bytes": 107.696,
|
| 193 |
+
"flops_per_problem": 5992752218.112
|
| 194 |
+
},
|
| 195 |
+
"add.find7": {
|
| 196 |
+
"acc": 0.93,
|
| 197 |
+
"reasoning_bytes": 121.306,
|
| 198 |
+
"flops_per_problem": 6682137772.032
|
| 199 |
+
},
|
| 200 |
+
"add.find8": {
|
| 201 |
+
"acc": 0.59,
|
| 202 |
+
"reasoning_bytes": 136.226,
|
| 203 |
+
"flops_per_problem": 7429995233.28
|
| 204 |
+
},
|
| 205 |
+
"add.rev6": {
|
| 206 |
+
"acc": 1.0,
|
| 207 |
+
"reasoning_bytes": 123.248,
|
| 208 |
+
"flops_per_problem": 6794576953.344
|
| 209 |
+
},
|
| 210 |
+
"add.rev7": {
|
| 211 |
+
"acc": 0.922,
|
| 212 |
+
"reasoning_bytes": 145.852,
|
| 213 |
+
"flops_per_problem": 7925063270.4
|
| 214 |
+
},
|
| 215 |
+
"add.rev8": {
|
| 216 |
+
"acc": 0.088,
|
| 217 |
+
"reasoning_bytes": 164.576,
|
| 218 |
+
"flops_per_problem": 8866266660.864
|
| 219 |
+
},
|
| 220 |
+
"add.sumd6": {
|
| 221 |
+
"acc": 0.994,
|
| 222 |
+
"reasoning_bytes": 126.476,
|
| 223 |
+
"flops_per_problem": 6777794985.984
|
| 224 |
+
},
|
| 225 |
+
"add.sumd7": {
|
| 226 |
+
"acc": 0.864,
|
| 227 |
+
"reasoning_bytes": 146.344,
|
| 228 |
+
"flops_per_problem": 7743729696.768
|
| 229 |
+
},
|
| 230 |
+
"add.sumd8": {
|
| 231 |
+
"acc": 0.128,
|
| 232 |
+
"reasoning_bytes": 163.37,
|
| 233 |
+
"flops_per_problem": 8588745916.416
|
| 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": 0.95,
|
| 242 |
+
"reasoning_bytes": 151.042,
|
| 243 |
+
"flops_per_problem": 8154711244.8
|
| 244 |
+
},
|
| 245 |
+
"add.inc8": {
|
| 246 |
+
"acc": 0.096,
|
| 247 |
+
"reasoning_bytes": 169.508,
|
| 248 |
+
"flops_per_problem": 9107750338.56
|
| 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": 99.0,
|
| 258 |
+
"flops_per_problem": 5741199360.0
|
| 259 |
+
},
|
| 260 |
+
"runmax.cmp16": {
|
| 261 |
+
"acc": 0.992,
|
| 262 |
+
"reasoning_bytes": 105.3,
|
| 263 |
+
"flops_per_problem": 6107752857.6
|
| 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": 99.952,
|
| 273 |
+
"flops_per_problem": 5827405676.544
|
| 274 |
+
},
|
| 275 |
+
"runmax.runmax16": {
|
| 276 |
+
"acc": 0.998,
|
| 277 |
+
"reasoning_bytes": 106.756,
|
| 278 |
+
"flops_per_problem": 6216217362.432
|
| 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": 99.952,
|
| 288 |
+
"flops_per_problem": 5871568748.544
|
| 289 |
+
},
|
| 290 |
+
"runmax.count16": {
|
| 291 |
+
"acc": 0.994,
|
| 292 |
+
"reasoning_bytes": 106.288,
|
| 293 |
+
"flops_per_problem": 6239712116.736
|
| 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": 99.832,
|
| 303 |
+
"flops_per_problem": 5822106107.904
|
| 304 |
+
},
|
| 305 |
+
"runmax.find16": {
|
| 306 |
+
"acc": 0.996,
|
| 307 |
+
"reasoning_bytes": 106.224,
|
| 308 |
+
"flops_per_problem": 6192722608.128
|
| 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": 94.988,
|
| 318 |
+
"flops_per_problem": 5475690971.136
|
| 319 |
+
},
|
| 320 |
+
"runmax.rev16": {
|
| 321 |
+
"acc": 0.988,
|
| 322 |
+
"reasoning_bytes": 102.16,
|
| 323 |
+
"flops_per_problem": 5880754667.52
|
| 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": 1.0,
|
| 332 |
+
"reasoning_bytes": 98.0,
|
| 333 |
+
"flops_per_problem": 5652873216.0
|
| 334 |
+
},
|
| 335 |
+
"runmax.sumd16": {
|
| 336 |
+
"acc": 0.992,
|
| 337 |
+
"reasoning_bytes": 107.844,
|
| 338 |
+
"flops_per_problem": 6175940640.768
|
| 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": 1.0,
|
| 347 |
+
"reasoning_bytes": 98.82,
|
| 348 |
+
"flops_per_problem": 5679459385.344
|
| 349 |
+
},
|
| 350 |
+
"runmax.inc16": {
|
| 351 |
+
"acc": 0.996,
|
| 352 |
+
"reasoning_bytes": 106.124,
|
| 353 |
+
"flops_per_problem": 6092825739.264
|
| 354 |
+
},
|
| 355 |
+
"count.cmp16": {
|
| 356 |
+
"acc": 1.0,
|
| 357 |
+
"reasoning_bytes": 95.01,
|
| 358 |
+
"flops_per_problem": 5697477918.72
|
| 359 |
+
},
|
| 360 |
+
"count.cmp18": {
|
| 361 |
+
"acc": 0.984,
|
| 362 |
+
"reasoning_bytes": 104.03,
|
| 363 |
+
"flops_per_problem": 6184243298.304
|
| 364 |
+
},
|
| 365 |
+
"count.cmp20": {
|
| 366 |
+
"acc": 0.95,
|
| 367 |
+
"reasoning_bytes": 110.078,
|
| 368 |
+
"flops_per_problem": 6539756027.904
|
| 369 |
+
},
|
| 370 |
+
"count.runmax16": {
|
| 371 |
+
"acc": 0.996,
|
| 372 |
+
"reasoning_bytes": 96.01,
|
| 373 |
+
"flops_per_problem": 5785804062.72
|
| 374 |
+
},
|
| 375 |
+
"count.runmax18": {
|
| 376 |
+
"acc": 0.984,
|
| 377 |
+
"reasoning_bytes": 105.72,
|
| 378 |
+
"flops_per_problem": 6302953635.84
|
| 379 |
+
},
|
| 380 |
+
"count.runmax20": {
|
| 381 |
+
"acc": 0.774,
|
| 382 |
+
"reasoning_bytes": 111.292,
|
| 383 |
+
"flops_per_problem": 6637356417.024
|
| 384 |
+
},
|
| 385 |
+
"count.count16": {
|
| 386 |
+
"acc": 0.998,
|
| 387 |
+
"reasoning_bytes": 95.99,
|
| 388 |
+
"flops_per_problem": 5829083873.28
|
| 389 |
+
},
|
| 390 |
+
"count.count18": {
|
| 391 |
+
"acc": 0.974,
|
| 392 |
+
"reasoning_bytes": 104.6,
|
| 393 |
+
"flops_per_problem": 6297654067.2
|
| 394 |
+
},
|
| 395 |
+
"count.count20": {
|
| 396 |
+
"acc": 0.942,
|
| 397 |
+
"reasoning_bytes": 111.03,
|
| 398 |
+
"flops_per_problem": 6669948764.16
|
| 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": 0.988,
|
| 407 |
+
"reasoning_bytes": 105.37,
|
| 408 |
+
"flops_per_problem": 6287496560.64
|
| 409 |
+
},
|
| 410 |
+
"count.find20": {
|
| 411 |
+
"acc": 0.934,
|
| 412 |
+
"reasoning_bytes": 111.156,
|
| 413 |
+
"flops_per_problem": 6631350239.232
|
| 414 |
+
},
|
| 415 |
+
"count.rev16": {
|
| 416 |
+
"acc": 0.996,
|
| 417 |
+
"reasoning_bytes": 91.0,
|
| 418 |
+
"flops_per_problem": 5432057856.0
|
| 419 |
+
},
|
| 420 |
+
"count.rev18": {
|
| 421 |
+
"acc": 0.938,
|
| 422 |
+
"reasoning_bytes": 99.73,
|
| 423 |
+
"flops_per_problem": 5905927618.56
|
| 424 |
+
},
|
| 425 |
+
"count.rev20": {
|
| 426 |
+
"acc": 0.764,
|
| 427 |
+
"reasoning_bytes": 106.12,
|
| 428 |
+
"flops_per_problem": 6276455792.64
|
| 429 |
+
},
|
| 430 |
+
"count.sumd16": {
|
| 431 |
+
"acc": 0.996,
|
| 432 |
+
"reasoning_bytes": 94.01,
|
| 433 |
+
"flops_per_problem": 5609151774.72
|
| 434 |
+
},
|
| 435 |
+
"count.sumd18": {
|
| 436 |
+
"acc": 0.958,
|
| 437 |
+
"reasoning_bytes": 103.35,
|
| 438 |
+
"flops_per_problem": 6109961011.2
|
| 439 |
+
},
|
| 440 |
+
"count.sumd20": {
|
| 441 |
+
"acc": 0.768,
|
| 442 |
+
"reasoning_bytes": 109.19,
|
| 443 |
+
"flops_per_problem": 6456199495.68
|
| 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": 0.952,
|
| 452 |
+
"reasoning_bytes": 104.126,
|
| 453 |
+
"flops_per_problem": 6100068483.072
|
| 454 |
+
},
|
| 455 |
+
"count.inc20": {
|
| 456 |
+
"acc": 0.776,
|
| 457 |
+
"reasoning_bytes": 110.128,
|
| 458 |
+
"flops_per_problem": 6453461385.216
|
| 459 |
+
},
|
| 460 |
+
"parity.cmp16": {
|
| 461 |
+
"acc": 0.996,
|
| 462 |
+
"reasoning_bytes": 111.06,
|
| 463 |
+
"flops_per_problem": 6362132152.32
|
| 464 |
+
},
|
| 465 |
+
"parity.cmp20": {
|
| 466 |
+
"acc": 0.956,
|
| 467 |
+
"reasoning_bytes": 135.0,
|
| 468 |
+
"flops_per_problem": 7596048384.0
|
| 469 |
+
},
|
| 470 |
+
"parity.cmp24": {
|
| 471 |
+
"acc": 0.956,
|
| 472 |
+
"reasoning_bytes": 153.072,
|
| 473 |
+
"flops_per_problem": 8570815709.184
|
| 474 |
+
},
|
| 475 |
+
"parity.runmax16": {
|
| 476 |
+
"acc": 0.99,
|
| 477 |
+
"reasoning_bytes": 112.036,
|
| 478 |
+
"flops_per_problem": 6449398382.592
|
| 479 |
+
},
|
| 480 |
+
"parity.runmax20": {
|
| 481 |
+
"acc": 0.916,
|
| 482 |
+
"reasoning_bytes": 136.0,
|
| 483 |
+
"flops_per_problem": 7684374528.0
|
| 484 |
+
},
|
| 485 |
+
"parity.runmax24": {
|
| 486 |
+
"acc": 0.474,
|
| 487 |
+
"reasoning_bytes": 154.0,
|
| 488 |
+
"flops_per_problem": 8655962112.0
|
| 489 |
+
},
|
| 490 |
+
"parity.count16": {
|
| 491 |
+
"acc": 0.998,
|
| 492 |
+
"reasoning_bytes": 112.0,
|
| 493 |
+
"flops_per_problem": 6491971584.0
|
| 494 |
+
},
|
| 495 |
+
"parity.count20": {
|
| 496 |
+
"acc": 0.95,
|
| 497 |
+
"reasoning_bytes": 135.976,
|
| 498 |
+
"flops_per_problem": 7727477686.272
|
| 499 |
+
},
|
| 500 |
+
"parity.count24": {
|
| 501 |
+
"acc": 0.814,
|
| 502 |
+
"reasoning_bytes": 154.012,
|
| 503 |
+
"flops_per_problem": 8700655140.864
|
| 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": 0.966,
|
| 512 |
+
"reasoning_bytes": 135.928,
|
| 513 |
+
"flops_per_problem": 7681194786.816
|
| 514 |
+
},
|
| 515 |
+
"parity.find24": {
|
| 516 |
+
"acc": 0.87,
|
| 517 |
+
"reasoning_bytes": 154.012,
|
| 518 |
+
"flops_per_problem": 8656492068.864
|
| 519 |
+
},
|
| 520 |
+
"parity.rev16": {
|
| 521 |
+
"acc": 0.984,
|
| 522 |
+
"reasoning_bytes": 107.036,
|
| 523 |
+
"flops_per_problem": 6096093806.592
|
| 524 |
+
},
|
| 525 |
+
"parity.rev20": {
|
| 526 |
+
"acc": 0.78,
|
| 527 |
+
"reasoning_bytes": 130.976,
|
| 528 |
+
"flops_per_problem": 7330010038.272
|
| 529 |
+
},
|
| 530 |
+
"parity.rev24": {
|
| 531 |
+
"acc": 0.486,
|
| 532 |
+
"reasoning_bytes": 149.492,
|
| 533 |
+
"flops_per_problem": 8324385767.424
|
| 534 |
+
},
|
| 535 |
+
"parity.sumd16": {
|
| 536 |
+
"acc": 0.986,
|
| 537 |
+
"reasoning_bytes": 110.048,
|
| 538 |
+
"flops_per_problem": 6273276051.456
|
| 539 |
+
},
|
| 540 |
+
"parity.sumd20": {
|
| 541 |
+
"acc": 0.81,
|
| 542 |
+
"reasoning_bytes": 133.928,
|
| 543 |
+
"flops_per_problem": 7504542498.816
|
| 544 |
+
},
|
| 545 |
+
"parity.sumd24": {
|
| 546 |
+
"acc": 0.478,
|
| 547 |
+
"reasoning_bytes": 152.372,
|
| 548 |
+
"flops_per_problem": 8495738486.784
|
| 549 |
+
},
|
| 550 |
+
"parity.inc16": {
|
| 551 |
+
"acc": 0.984,
|
| 552 |
+
"reasoning_bytes": 111.036,
|
| 553 |
+
"flops_per_problem": 6272746094.592
|
| 554 |
+
},
|
| 555 |
+
"parity.inc20": {
|
| 556 |
+
"acc": 0.81,
|
| 557 |
+
"reasoning_bytes": 134.976,
|
| 558 |
+
"flops_per_problem": 7506662326.272
|
| 559 |
+
},
|
| 560 |
+
"parity.inc24": {
|
| 561 |
+
"acc": 0.514,
|
| 562 |
+
"reasoning_bytes": 153.9,
|
| 563 |
+
"flops_per_problem": 8519056588.8
|
| 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": 1.0,
|
| 572 |
+
"reasoning_bytes": 70.076,
|
| 573 |
+
"flops_per_problem": 4110522089.472
|
| 574 |
+
},
|
| 575 |
+
"sumd.cmp8": {
|
| 576 |
+
"acc": 0.988,
|
| 577 |
+
"reasoning_bytes": 77.626,
|
| 578 |
+
"flops_per_problem": 4488028028.928
|
| 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": 0.912,
|
| 592 |
+
"reasoning_bytes": 83.054,
|
| 593 |
+
"flops_per_problem": 4727921836.032
|
| 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": 0.998,
|
| 602 |
+
"reasoning_bytes": 74.46,
|
| 603 |
+
"flops_per_problem": 4348296069.12
|
| 604 |
+
},
|
| 605 |
+
"sumd.count8": {
|
| 606 |
+
"acc": 0.95,
|
| 607 |
+
"reasoning_bytes": 82.416,
|
| 608 |
+
"flops_per_problem": 4743820541.952
|
| 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": 74.428,
|
| 618 |
+
"flops_per_problem": 4302719778.816
|
| 619 |
+
},
|
| 620 |
+
"sumd.find8": {
|
| 621 |
+
"acc": 0.982,
|
| 622 |
+
"reasoning_bytes": 82.18,
|
| 623 |
+
"flops_per_problem": 4689234984.96
|
| 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": 0.962,
|
| 637 |
+
"reasoning_bytes": 77.498,
|
| 638 |
+
"flops_per_problem": 4394049011.712
|
| 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": 0.996,
|
| 647 |
+
"reasoning_bytes": 73.738,
|
| 648 |
+
"flops_per_problem": 4195315187.712
|
| 649 |
+
},
|
| 650 |
+
"sumd.sumd8": {
|
| 651 |
+
"acc": 0.832,
|
| 652 |
+
"reasoning_bytes": 80.116,
|
| 653 |
+
"flops_per_problem": 4524065095.68
|
| 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": 0.998,
|
| 662 |
+
"reasoning_bytes": 76.48,
|
| 663 |
+
"flops_per_problem": 4305016258.56
|
| 664 |
+
},
|
| 665 |
+
"sumd.inc8": {
|
| 666 |
+
"acc": 0.876,
|
| 667 |
+
"reasoning_bytes": 84.014,
|
| 668 |
+
"flops_per_problem": 4682345545.728
|
| 669 |
+
},
|
| 670 |
+
"inc.cmp6": {
|
| 671 |
+
"acc": 1.0,
|
| 672 |
+
"reasoning_bytes": 78.612,
|
| 673 |
+
"flops_per_problem": 4575824216.064
|
| 674 |
+
},
|
| 675 |
+
"inc.cmp7": {
|
| 676 |
+
"acc": 0.996,
|
| 677 |
+
"reasoning_bytes": 89.766,
|
| 678 |
+
"flops_per_problem": 5156745265.152
|
| 679 |
+
},
|
| 680 |
+
"inc.cmp8": {
|
| 681 |
+
"acc": 0.954,
|
| 682 |
+
"reasoning_bytes": 100.424,
|
| 683 |
+
"flops_per_problem": 5721237651.456
|
| 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": 119.992,
|
| 693 |
+
"flops_per_problem": 6270802919.424
|
| 694 |
+
},
|
| 695 |
+
"inc.runmax8": {
|
| 696 |
+
"acc": 0.866,
|
| 697 |
+
"reasoning_bytes": 135.432,
|
| 698 |
+
"flops_per_problem": 6997020475.392
|
| 699 |
+
},
|
| 700 |
+
"inc.count6": {
|
| 701 |
+
"acc": 0.982,
|
| 702 |
+
"reasoning_bytes": 99.0,
|
| 703 |
+
"flops_per_problem": 5343731712.0
|
| 704 |
+
},
|
| 705 |
+
"inc.count7": {
|
| 706 |
+
"acc": 0.892,
|
| 707 |
+
"reasoning_bytes": 113.974,
|
| 708 |
+
"flops_per_problem": 6049192624.128
|
| 709 |
+
},
|
| 710 |
+
"inc.count8": {
|
| 711 |
+
"acc": 0.664,
|
| 712 |
+
"reasoning_bytes": 120.884,
|
| 713 |
+
"flops_per_problem": 6398522523.648
|
| 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": 0.926,
|
| 722 |
+
"reasoning_bytes": 102.672,
|
| 723 |
+
"flops_per_problem": 5505898512.384
|
| 724 |
+
},
|
| 725 |
+
"inc.find8": {
|
| 726 |
+
"acc": 0.458,
|
| 727 |
+
"reasoning_bytes": 119.938,
|
| 728 |
+
"flops_per_problem": 6312581185.536
|
| 729 |
+
},
|
| 730 |
+
"inc.rev6": {
|
| 731 |
+
"acc": 1.0,
|
| 732 |
+
"reasoning_bytes": 104.0,
|
| 733 |
+
"flops_per_problem": 5608710144.0
|
| 734 |
+
},
|
| 735 |
+
"inc.rev7": {
|
| 736 |
+
"acc": 0.902,
|
| 737 |
+
"reasoning_bytes": 123.772,
|
| 738 |
+
"flops_per_problem": 6569080307.712
|
| 739 |
+
},
|
| 740 |
+
"inc.rev8": {
|
| 741 |
+
"acc": 0.0,
|
| 742 |
+
"reasoning_bytes": 134.572,
|
| 743 |
+
"flops_per_problem": 7095327473.664
|
| 744 |
+
},
|
| 745 |
+
"inc.sumd6": {
|
| 746 |
+
"acc": 0.99,
|
| 747 |
+
"reasoning_bytes": 109.558,
|
| 748 |
+
"flops_per_problem": 5721590956.032
|
| 749 |
+
},
|
| 750 |
+
"inc.sumd7": {
|
| 751 |
+
"acc": 0.898,
|
| 752 |
+
"reasoning_bytes": 127.446,
|
| 753 |
+
"flops_per_problem": 6555654733.824
|
| 754 |
+
},
|
| 755 |
+
"inc.sumd8": {
|
| 756 |
+
"acc": 0.074,
|
| 757 |
+
"reasoning_bytes": 140.126,
|
| 758 |
+
"flops_per_problem": 7159893884.928
|
| 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": 0.886,
|
| 767 |
+
"reasoning_bytes": 130.898,
|
| 768 |
+
"flops_per_problem": 6884934598.656
|
| 769 |
+
},
|
| 770 |
+
"inc.inc8": {
|
| 771 |
+
"acc": 0.0,
|
| 772 |
+
"reasoning_bytes": 144.93,
|
| 773 |
+
"flops_per_problem": 7579531395.072
|
| 774 |
+
},
|
| 775 |
+
"fss6": {
|
| 776 |
+
"acc": 0.78,
|
| 777 |
+
"actions": 4.03,
|
| 778 |
+
"invalid": 0.196,
|
| 779 |
+
"reasoning_bytes": 40.016,
|
| 780 |
+
"flops_per_problem": 2755687366.656
|
| 781 |
+
},
|
| 782 |
+
"fss7": {
|
| 783 |
+
"acc": 0.736,
|
| 784 |
+
"actions": 4.746,
|
| 785 |
+
"invalid": 0.292,
|
| 786 |
+
"reasoning_bytes": 47.73,
|
| 787 |
+
"flops_per_problem": 3190870278.144
|
| 788 |
+
},
|
| 789 |
+
"fss8": {
|
| 790 |
+
"acc": 0.69,
|
| 791 |
+
"actions": 4.93,
|
| 792 |
+
"invalid": 0.342,
|
| 793 |
+
"reasoning_bytes": 49.254,
|
| 794 |
+
"flops_per_problem": 3314791858.176
|
| 795 |
+
}
|
| 796 |
+
}
|