Upload folder using huggingface_hub
Browse files
5m-s0-bpe4096-step-bytes3-v15-none-docmask/ckpt.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4a72007161fd2349de3f1d1f978d1cc57e573297f2236d3049fbb56a15f1214c
|
| 3 |
+
size 23171663
|
5m-s0-bpe4096-step-bytes3-v15-none-docmask/config.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"size": "5m",
|
| 3 |
+
"name": "5m-s0-bpe4096-step-bytes3-v15-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": "v15",
|
| 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": 193600261
|
| 31 |
+
}
|
5m-s0-bpe4096-step-bytes3-v15-none-docmask/log.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
5m-s0-bpe4096-step-bytes3-v15-none-docmask/summary.json
ADDED
|
@@ -0,0 +1,796 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"params": 5787648,
|
| 3 |
+
"tokens": 399998976,
|
| 4 |
+
"train_flops": 17665183577014272,
|
| 5 |
+
"wall_s": 735.696070432663,
|
| 6 |
+
"peak_mem_gb": 2.520964096,
|
| 7 |
+
"final_val_loss": 1.7855027914047241,
|
| 8 |
+
"final_val_bpb": 0.643644748566888,
|
| 9 |
+
"bytes_seen": 1600839241.8284984,
|
| 10 |
+
"add8": {
|
| 11 |
+
"acc": 0.994,
|
| 12 |
+
"reasoning_bytes": 79.0,
|
| 13 |
+
"flops_per_problem": 4928775487.488
|
| 14 |
+
},
|
| 15 |
+
"add9": {
|
| 16 |
+
"acc": 0.888,
|
| 17 |
+
"reasoning_bytes": 89.0,
|
| 18 |
+
"flops_per_problem": 5502895423.488
|
| 19 |
+
},
|
| 20 |
+
"add10": {
|
| 21 |
+
"acc": 0.006,
|
| 22 |
+
"reasoning_bytes": 98.98,
|
| 23 |
+
"flops_per_problem": 6075955445.76
|
| 24 |
+
},
|
| 25 |
+
"reach14": {
|
| 26 |
+
"acc": 0.666,
|
| 27 |
+
"reasoning_bytes": 7.812,
|
| 28 |
+
"flops_per_problem": 4628819902.464
|
| 29 |
+
},
|
| 30 |
+
"reach16": {
|
| 31 |
+
"acc": 0.592,
|
| 32 |
+
"reasoning_bytes": 7.732,
|
| 33 |
+
"flops_per_problem": 5155243720.704
|
| 34 |
+
},
|
| 35 |
+
"reach18": {
|
| 36 |
+
"acc": 0.564,
|
| 37 |
+
"reasoning_bytes": 8.24,
|
| 38 |
+
"flops_per_problem": 5707547099.136
|
| 39 |
+
},
|
| 40 |
+
"parity24": {
|
| 41 |
+
"acc": 0.918,
|
| 42 |
+
"reasoning_bytes": 143.0,
|
| 43 |
+
"flops_per_problem": 7861026816.0
|
| 44 |
+
},
|
| 45 |
+
"parity28": {
|
| 46 |
+
"acc": 0.626,
|
| 47 |
+
"reasoning_bytes": 167.108,
|
| 48 |
+
"flops_per_problem": 9102362443.776
|
| 49 |
+
},
|
| 50 |
+
"parity32": {
|
| 51 |
+
"acc": 0.442,
|
| 52 |
+
"reasoning_bytes": 191.872,
|
| 53 |
+
"flops_per_problem": 10372580720.64
|
| 54 |
+
},
|
| 55 |
+
"cmp8": {
|
| 56 |
+
"acc": 0.996,
|
| 57 |
+
"reasoning_bytes": 5.816,
|
| 58 |
+
"flops_per_problem": 1360929226.752
|
| 59 |
+
},
|
| 60 |
+
"cmp9": {
|
| 61 |
+
"acc": 1.0,
|
| 62 |
+
"reasoning_bytes": 5.696,
|
| 63 |
+
"flops_per_problem": 1443955802.112
|
| 64 |
+
},
|
| 65 |
+
"cmp10": {
|
| 66 |
+
"acc": 1.0,
|
| 67 |
+
"reasoning_bytes": 5.768,
|
| 68 |
+
"flops_per_problem": 1535461687.296
|
| 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": 82.604,
|
| 78 |
+
"flops_per_problem": 4752123199.488
|
| 79 |
+
},
|
| 80 |
+
"runmax16": {
|
| 81 |
+
"acc": 0.974,
|
| 82 |
+
"reasoning_bytes": 89.912,
|
| 83 |
+
"flops_per_problem": 5163193073.664
|
| 84 |
+
},
|
| 85 |
+
"count16": {
|
| 86 |
+
"acc": 1.0,
|
| 87 |
+
"reasoning_bytes": 79.0,
|
| 88 |
+
"flops_per_problem": 4725448704.0
|
| 89 |
+
},
|
| 90 |
+
"count18": {
|
| 91 |
+
"acc": 0.986,
|
| 92 |
+
"reasoning_bytes": 88.62,
|
| 93 |
+
"flops_per_problem": 5238623600.64
|
| 94 |
+
},
|
| 95 |
+
"count20": {
|
| 96 |
+
"acc": 0.852,
|
| 97 |
+
"reasoning_bytes": 95.53,
|
| 98 |
+
"flops_per_problem": 5632116572.16
|
| 99 |
+
},
|
| 100 |
+
"brackets16": {
|
| 101 |
+
"acc": 0.996,
|
| 102 |
+
"reasoning_bytes": 53.418,
|
| 103 |
+
"flops_per_problem": 3641068634.112
|
| 104 |
+
},
|
| 105 |
+
"brackets18": {
|
| 106 |
+
"acc": 0.98,
|
| 107 |
+
"reasoning_bytes": 56.732,
|
| 108 |
+
"flops_per_problem": 3874867937.28
|
| 109 |
+
},
|
| 110 |
+
"brackets20": {
|
| 111 |
+
"acc": 0.958,
|
| 112 |
+
"reasoning_bytes": 63.512,
|
| 113 |
+
"flops_per_problem": 4262619709.44
|
| 114 |
+
},
|
| 115 |
+
"sumd8": {
|
| 116 |
+
"acc": 0.996,
|
| 117 |
+
"reasoning_bytes": 58.566,
|
| 118 |
+
"flops_per_problem": 3381301444.608
|
| 119 |
+
},
|
| 120 |
+
"sumd9": {
|
| 121 |
+
"acc": 0.936,
|
| 122 |
+
"reasoning_bytes": 66.364,
|
| 123 |
+
"flops_per_problem": 3769936478.208
|
| 124 |
+
},
|
| 125 |
+
"sumd10": {
|
| 126 |
+
"acc": 0.52,
|
| 127 |
+
"reasoning_bytes": 71.264,
|
| 128 |
+
"flops_per_problem": 4030586929.152
|
| 129 |
+
},
|
| 130 |
+
"inc8": {
|
| 131 |
+
"acc": 0.99,
|
| 132 |
+
"reasoning_bytes": 63.0,
|
| 133 |
+
"flops_per_problem": 3798024192.0
|
| 134 |
+
},
|
| 135 |
+
"inc9": {
|
| 136 |
+
"acc": 0.774,
|
| 137 |
+
"reasoning_bytes": 71.0,
|
| 138 |
+
"flops_per_problem": 4239654912.0
|
| 139 |
+
},
|
| 140 |
+
"inc10": {
|
| 141 |
+
"acc": 0.008,
|
| 142 |
+
"reasoning_bytes": 77.724,
|
| 143 |
+
"flops_per_problem": 4573969367.04
|
| 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": 1.0,
|
| 152 |
+
"reasoning_bytes": 105.532,
|
| 153 |
+
"flops_per_problem": 6231939416.064
|
| 154 |
+
},
|
| 155 |
+
"add.cmp8": {
|
| 156 |
+
"acc": 0.988,
|
| 157 |
+
"reasoning_bytes": 118.798,
|
| 158 |
+
"flops_per_problem": 6951797489.664
|
| 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": 0.966,
|
| 167 |
+
"reasoning_bytes": 138.848,
|
| 168 |
+
"flops_per_problem": 7456846381.056
|
| 169 |
+
},
|
| 170 |
+
"add.runmax8": {
|
| 171 |
+
"acc": 0.67,
|
| 172 |
+
"reasoning_bytes": 156.914,
|
| 173 |
+
"flops_per_problem": 8343022583.808
|
| 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": 0.958,
|
| 182 |
+
"reasoning_bytes": 132.09,
|
| 183 |
+
"flops_per_problem": 7202555412.48
|
| 184 |
+
},
|
| 185 |
+
"add.count8": {
|
| 186 |
+
"acc": 0.4,
|
| 187 |
+
"reasoning_bytes": 148.824,
|
| 188 |
+
"flops_per_problem": 8033174470.656
|
| 189 |
+
},
|
| 190 |
+
"add.find6": {
|
| 191 |
+
"acc": 0.954,
|
| 192 |
+
"reasoning_bytes": 107.696,
|
| 193 |
+
"flops_per_problem": 5992752218.112
|
| 194 |
+
},
|
| 195 |
+
"add.find7": {
|
| 196 |
+
"acc": 0.902,
|
| 197 |
+
"reasoning_bytes": 121.2,
|
| 198 |
+
"flops_per_problem": 6677456486.4
|
| 199 |
+
},
|
| 200 |
+
"add.find8": {
|
| 201 |
+
"acc": 0.572,
|
| 202 |
+
"reasoning_bytes": 137.2,
|
| 203 |
+
"flops_per_problem": 7474423283.712
|
| 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": 0.812,
|
| 212 |
+
"reasoning_bytes": 145.866,
|
| 213 |
+
"flops_per_problem": 7925681553.408
|
| 214 |
+
},
|
| 215 |
+
"add.rev8": {
|
| 216 |
+
"acc": 0.004,
|
| 217 |
+
"reasoning_bytes": 160.746,
|
| 218 |
+
"flops_per_problem": 8679721844.736
|
| 219 |
+
},
|
| 220 |
+
"add.sumd6": {
|
| 221 |
+
"acc": 0.998,
|
| 222 |
+
"reasoning_bytes": 126.476,
|
| 223 |
+
"flops_per_problem": 6777794985.984
|
| 224 |
+
},
|
| 225 |
+
"add.sumd7": {
|
| 226 |
+
"acc": 0.858,
|
| 227 |
+
"reasoning_bytes": 146.368,
|
| 228 |
+
"flops_per_problem": 7745054588.928
|
| 229 |
+
},
|
| 230 |
+
"add.sumd8": {
|
| 231 |
+
"acc": 0.13,
|
| 232 |
+
"reasoning_bytes": 165.748,
|
| 233 |
+
"flops_per_problem": 8691380895.744
|
| 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": 0.97,
|
| 242 |
+
"reasoning_bytes": 151.06,
|
| 243 |
+
"flops_per_problem": 8155771158.528
|
| 244 |
+
},
|
| 245 |
+
"add.inc8": {
|
| 246 |
+
"acc": 0.074,
|
| 247 |
+
"reasoning_bytes": 170.718,
|
| 248 |
+
"flops_per_problem": 9159244480.512
|
| 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": 96.18,
|
| 258 |
+
"flops_per_problem": 5616659496.96
|
| 259 |
+
},
|
| 260 |
+
"runmax.cmp16": {
|
| 261 |
+
"acc": 0.992,
|
| 262 |
+
"reasoning_bytes": 105.0,
|
| 263 |
+
"flops_per_problem": 6094503936.0
|
| 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": 99.58,
|
| 273 |
+
"flops_per_problem": 5810977013.76
|
| 274 |
+
},
|
| 275 |
+
"runmax.runmax16": {
|
| 276 |
+
"acc": 0.972,
|
| 277 |
+
"reasoning_bytes": 106.192,
|
| 278 |
+
"flops_per_problem": 6191309389.824
|
| 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": 99.316,
|
| 288 |
+
"flops_per_problem": 5843481034.752
|
| 289 |
+
},
|
| 290 |
+
"runmax.count16": {
|
| 291 |
+
"acc": 0.998,
|
| 292 |
+
"reasoning_bytes": 106.036,
|
| 293 |
+
"flops_per_problem": 6228583022.592
|
| 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": 98.944,
|
| 303 |
+
"flops_per_problem": 5782889299.968
|
| 304 |
+
},
|
| 305 |
+
"runmax.find16": {
|
| 306 |
+
"acc": 0.982,
|
| 307 |
+
"reasoning_bytes": 106.008,
|
| 308 |
+
"flops_per_problem": 6183183384.576
|
| 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": 94.544,
|
| 318 |
+
"flops_per_problem": 5456082567.168
|
| 319 |
+
},
|
| 320 |
+
"runmax.rev16": {
|
| 321 |
+
"acc": 0.976,
|
| 322 |
+
"reasoning_bytes": 101.18,
|
| 323 |
+
"flops_per_problem": 5837474856.96
|
| 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": 96.464,
|
| 333 |
+
"flops_per_problem": 5585038737.408
|
| 334 |
+
},
|
| 335 |
+
"runmax.sumd16": {
|
| 336 |
+
"acc": 0.968,
|
| 337 |
+
"reasoning_bytes": 104.06,
|
| 338 |
+
"flops_per_problem": 6008827576.32
|
| 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": 97.968,
|
| 348 |
+
"flops_per_problem": 5641832448.0
|
| 349 |
+
},
|
| 350 |
+
"runmax.inc16": {
|
| 351 |
+
"acc": 0.968,
|
| 352 |
+
"reasoning_bytes": 105.06,
|
| 353 |
+
"flops_per_problem": 6044687990.784
|
| 354 |
+
},
|
| 355 |
+
"count.cmp16": {
|
| 356 |
+
"acc": 1.0,
|
| 357 |
+
"reasoning_bytes": 95.0,
|
| 358 |
+
"flops_per_problem": 5697036288.0
|
| 359 |
+
},
|
| 360 |
+
"count.cmp18": {
|
| 361 |
+
"acc": 0.996,
|
| 362 |
+
"reasoning_bytes": 104.73,
|
| 363 |
+
"flops_per_problem": 6215157448.704
|
| 364 |
+
},
|
| 365 |
+
"count.cmp20": {
|
| 366 |
+
"acc": 0.984,
|
| 367 |
+
"reasoning_bytes": 111.04,
|
| 368 |
+
"flops_per_problem": 6582240903.168
|
| 369 |
+
},
|
| 370 |
+
"count.runmax16": {
|
| 371 |
+
"acc": 1.0,
|
| 372 |
+
"reasoning_bytes": 96.0,
|
| 373 |
+
"flops_per_problem": 5785362432.0
|
| 374 |
+
},
|
| 375 |
+
"count.runmax18": {
|
| 376 |
+
"acc": 0.996,
|
| 377 |
+
"reasoning_bytes": 105.93,
|
| 378 |
+
"flops_per_problem": 6312227880.96
|
| 379 |
+
},
|
| 380 |
+
"count.runmax20": {
|
| 381 |
+
"acc": 0.874,
|
| 382 |
+
"reasoning_bytes": 112.99,
|
| 383 |
+
"flops_per_problem": 6712345313.28
|
| 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": 105.92,
|
| 393 |
+
"flops_per_problem": 6355949322.24
|
| 394 |
+
},
|
| 395 |
+
"count.count20": {
|
| 396 |
+
"acc": 0.998,
|
| 397 |
+
"reasoning_bytes": 112.976,
|
| 398 |
+
"flops_per_problem": 6755890102.272
|
| 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": 0.996,
|
| 407 |
+
"reasoning_bytes": 105.58,
|
| 408 |
+
"flops_per_problem": 6296770805.76
|
| 409 |
+
},
|
| 410 |
+
"count.find20": {
|
| 411 |
+
"acc": 0.94,
|
| 412 |
+
"reasoning_bytes": 111.944,
|
| 413 |
+
"flops_per_problem": 6666150739.968
|
| 414 |
+
},
|
| 415 |
+
"count.rev16": {
|
| 416 |
+
"acc": 1.0,
|
| 417 |
+
"reasoning_bytes": 91.0,
|
| 418 |
+
"flops_per_problem": 5432057856.0
|
| 419 |
+
},
|
| 420 |
+
"count.rev18": {
|
| 421 |
+
"acc": 0.992,
|
| 422 |
+
"reasoning_bytes": 100.75,
|
| 423 |
+
"flops_per_problem": 5950973952.0
|
| 424 |
+
},
|
| 425 |
+
"count.rev20": {
|
| 426 |
+
"acc": 0.824,
|
| 427 |
+
"reasoning_bytes": 106.73,
|
| 428 |
+
"flops_per_problem": 6303395266.56
|
| 429 |
+
},
|
| 430 |
+
"count.sumd16": {
|
| 431 |
+
"acc": 1.0,
|
| 432 |
+
"reasoning_bytes": 94.0,
|
| 433 |
+
"flops_per_problem": 5608710144.0
|
| 434 |
+
},
|
| 435 |
+
"count.sumd18": {
|
| 436 |
+
"acc": 0.994,
|
| 437 |
+
"reasoning_bytes": 103.78,
|
| 438 |
+
"flops_per_problem": 6128951132.16
|
| 439 |
+
},
|
| 440 |
+
"count.sumd20": {
|
| 441 |
+
"acc": 0.86,
|
| 442 |
+
"reasoning_bytes": 110.48,
|
| 443 |
+
"flops_per_problem": 6513169858.56
|
| 444 |
+
},
|
| 445 |
+
"count.inc16": {
|
| 446 |
+
"acc": 1.0,
|
| 447 |
+
"reasoning_bytes": 95.0,
|
| 448 |
+
"flops_per_problem": 5608798470.144
|
| 449 |
+
},
|
| 450 |
+
"count.inc18": {
|
| 451 |
+
"acc": 0.986,
|
| 452 |
+
"reasoning_bytes": 104.56,
|
| 453 |
+
"flops_per_problem": 6119323582.464
|
| 454 |
+
},
|
| 455 |
+
"count.inc20": {
|
| 456 |
+
"acc": 0.814,
|
| 457 |
+
"reasoning_bytes": 110.56,
|
| 458 |
+
"flops_per_problem": 6472539832.32
|
| 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": 0.978,
|
| 467 |
+
"reasoning_bytes": 132.204,
|
| 468 |
+
"flops_per_problem": 7472568434.688
|
| 469 |
+
},
|
| 470 |
+
"parity.cmp24": {
|
| 471 |
+
"acc": 0.952,
|
| 472 |
+
"reasoning_bytes": 153.012,
|
| 473 |
+
"flops_per_problem": 8568165924.864
|
| 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": 0.83,
|
| 482 |
+
"reasoning_bytes": 134.248,
|
| 483 |
+
"flops_per_problem": 7607000825.856
|
| 484 |
+
},
|
| 485 |
+
"parity.runmax24": {
|
| 486 |
+
"acc": 0.582,
|
| 487 |
+
"reasoning_bytes": 154.096,
|
| 488 |
+
"flops_per_problem": 8660201766.912
|
| 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": 134.342,
|
| 498 |
+
"flops_per_problem": 7655315226.624
|
| 499 |
+
},
|
| 500 |
+
"parity.count24": {
|
| 501 |
+
"acc": 0.984,
|
| 502 |
+
"reasoning_bytes": 154.026,
|
| 503 |
+
"flops_per_problem": 8701273423.872
|
| 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": 0.934,
|
| 512 |
+
"reasoning_bytes": 133.456,
|
| 513 |
+
"flops_per_problem": 7572023672.832
|
| 514 |
+
},
|
| 515 |
+
"parity.find24": {
|
| 516 |
+
"acc": 0.894,
|
| 517 |
+
"reasoning_bytes": 154.084,
|
| 518 |
+
"flops_per_problem": 8659671810.048
|
| 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": 0.946,
|
| 527 |
+
"reasoning_bytes": 130.952,
|
| 528 |
+
"flops_per_problem": 7328950124.544
|
| 529 |
+
},
|
| 530 |
+
"parity.rev24": {
|
| 531 |
+
"acc": 0.762,
|
| 532 |
+
"reasoning_bytes": 150.14,
|
| 533 |
+
"flops_per_problem": 8353003438.08
|
| 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": 0.832,
|
| 542 |
+
"reasoning_bytes": 132.644,
|
| 543 |
+
"flops_per_problem": 7447837114.368
|
| 544 |
+
},
|
| 545 |
+
"parity.sumd24": {
|
| 546 |
+
"acc": 0.726,
|
| 547 |
+
"reasoning_bytes": 152.912,
|
| 548 |
+
"flops_per_problem": 8519586545.664
|
| 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": 0.896,
|
| 557 |
+
"reasoning_bytes": 134.52,
|
| 558 |
+
"flops_per_problem": 7486523965.44
|
| 559 |
+
},
|
| 560 |
+
"parity.inc24": {
|
| 561 |
+
"acc": 0.594,
|
| 562 |
+
"reasoning_bytes": 153.544,
|
| 563 |
+
"flops_per_problem": 8503334535.168
|
| 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": 0.994,
|
| 572 |
+
"reasoning_bytes": 70.084,
|
| 573 |
+
"flops_per_problem": 4110875394.048
|
| 574 |
+
},
|
| 575 |
+
"sumd.cmp8": {
|
| 576 |
+
"acc": 0.984,
|
| 577 |
+
"reasoning_bytes": 76.578,
|
| 578 |
+
"flops_per_problem": 4441745129.472
|
| 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": 1.0,
|
| 587 |
+
"reasoning_bytes": 75.48,
|
| 588 |
+
"flops_per_problem": 4349179330.56
|
| 589 |
+
},
|
| 590 |
+
"sumd.runmax8": {
|
| 591 |
+
"acc": 0.83,
|
| 592 |
+
"reasoning_bytes": 81.672,
|
| 593 |
+
"flops_per_problem": 4667948384.256
|
| 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": 1.0,
|
| 602 |
+
"reasoning_bytes": 74.46,
|
| 603 |
+
"flops_per_problem": 4348296069.12
|
| 604 |
+
},
|
| 605 |
+
"sumd.count8": {
|
| 606 |
+
"acc": 0.858,
|
| 607 |
+
"reasoning_bytes": 81.06,
|
| 608 |
+
"flops_per_problem": 4684288720.896
|
| 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": 0.99,
|
| 617 |
+
"reasoning_bytes": 74.428,
|
| 618 |
+
"flops_per_problem": 4302719778.816
|
| 619 |
+
},
|
| 620 |
+
"sumd.find8": {
|
| 621 |
+
"acc": 0.956,
|
| 622 |
+
"reasoning_bytes": 81.914,
|
| 623 |
+
"flops_per_problem": 4677487607.808
|
| 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": 0.936,
|
| 637 |
+
"reasoning_bytes": 76.652,
|
| 638 |
+
"flops_per_problem": 4356687052.8
|
| 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": 1.0,
|
| 647 |
+
"reasoning_bytes": 73.738,
|
| 648 |
+
"flops_per_problem": 4195315187.712
|
| 649 |
+
},
|
| 650 |
+
"sumd.sumd8": {
|
| 651 |
+
"acc": 0.848,
|
| 652 |
+
"reasoning_bytes": 80.65,
|
| 653 |
+
"flops_per_problem": 4547383197.696
|
| 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": 0.996,
|
| 662 |
+
"reasoning_bytes": 76.48,
|
| 663 |
+
"flops_per_problem": 4305192910.848
|
| 664 |
+
},
|
| 665 |
+
"sumd.inc8": {
|
| 666 |
+
"acc": 0.792,
|
| 667 |
+
"reasoning_bytes": 82.834,
|
| 668 |
+
"flops_per_problem": 4632176295.936
|
| 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": 89.792,
|
| 678 |
+
"flops_per_problem": 5157893505.024
|
| 679 |
+
},
|
| 680 |
+
"inc.cmp8": {
|
| 681 |
+
"acc": 0.984,
|
| 682 |
+
"reasoning_bytes": 100.824,
|
| 683 |
+
"flops_per_problem": 5733691637.76
|
| 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": 0.982,
|
| 692 |
+
"reasoning_bytes": 120.0,
|
| 693 |
+
"flops_per_problem": 6271156224.0
|
| 694 |
+
},
|
| 695 |
+
"inc.runmax8": {
|
| 696 |
+
"acc": 0.894,
|
| 697 |
+
"reasoning_bytes": 134.696,
|
| 698 |
+
"flops_per_problem": 6964339802.112
|
| 699 |
+
},
|
| 700 |
+
"inc.count6": {
|
| 701 |
+
"acc": 0.984,
|
| 702 |
+
"reasoning_bytes": 99.0,
|
| 703 |
+
"flops_per_problem": 5343731712.0
|
| 704 |
+
},
|
| 705 |
+
"inc.count7": {
|
| 706 |
+
"acc": 0.936,
|
| 707 |
+
"reasoning_bytes": 114.0,
|
| 708 |
+
"flops_per_problem": 6050340864.0
|
| 709 |
+
},
|
| 710 |
+
"inc.count8": {
|
| 711 |
+
"acc": 0.616,
|
| 712 |
+
"reasoning_bytes": 123.188,
|
| 713 |
+
"flops_per_problem": 6500274241.536
|
| 714 |
+
},
|
| 715 |
+
"inc.find6": {
|
| 716 |
+
"acc": 0.99,
|
| 717 |
+
"reasoning_bytes": 91.88,
|
| 718 |
+
"flops_per_problem": 4985127567.36
|
| 719 |
+
},
|
| 720 |
+
"inc.find7": {
|
| 721 |
+
"acc": 0.924,
|
| 722 |
+
"reasoning_bytes": 102.6,
|
| 723 |
+
"flops_per_problem": 5502718771.2
|
| 724 |
+
},
|
| 725 |
+
"inc.find8": {
|
| 726 |
+
"acc": 0.704,
|
| 727 |
+
"reasoning_bytes": 109.646,
|
| 728 |
+
"flops_per_problem": 5858054848.512
|
| 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": 0.868,
|
| 737 |
+
"reasoning_bytes": 124.0,
|
| 738 |
+
"flops_per_problem": 6580209401.856
|
| 739 |
+
},
|
| 740 |
+
"inc.rev8": {
|
| 741 |
+
"acc": 0.0,
|
| 742 |
+
"reasoning_bytes": 134.898,
|
| 743 |
+
"flops_per_problem": 7112727724.032
|
| 744 |
+
},
|
| 745 |
+
"inc.sumd6": {
|
| 746 |
+
"acc": 0.994,
|
| 747 |
+
"reasoning_bytes": 109.556,
|
| 748 |
+
"flops_per_problem": 5721414303.744
|
| 749 |
+
},
|
| 750 |
+
"inc.sumd7": {
|
| 751 |
+
"acc": 0.866,
|
| 752 |
+
"reasoning_bytes": 127.49,
|
| 753 |
+
"flops_per_problem": 6557597908.992
|
| 754 |
+
},
|
| 755 |
+
"inc.sumd8": {
|
| 756 |
+
"acc": 0.172,
|
| 757 |
+
"reasoning_bytes": 142.956,
|
| 758 |
+
"flops_per_problem": 7285405335.552
|
| 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": 0.986,
|
| 767 |
+
"reasoning_bytes": 131.0,
|
| 768 |
+
"flops_per_problem": 6889439232.0
|
| 769 |
+
},
|
| 770 |
+
"inc.inc8": {
|
| 771 |
+
"acc": 0.204,
|
| 772 |
+
"reasoning_bytes": 144.448,
|
| 773 |
+
"flops_per_problem": 7562837753.856
|
| 774 |
+
},
|
| 775 |
+
"fsf6": {
|
| 776 |
+
"acc": 0.976,
|
| 777 |
+
"actions": 3.324,
|
| 778 |
+
"invalid": 0.056,
|
| 779 |
+
"reasoning_bytes": 39.668,
|
| 780 |
+
"flops_per_problem": 2694830653.44
|
| 781 |
+
},
|
| 782 |
+
"fsf7": {
|
| 783 |
+
"acc": 0.964,
|
| 784 |
+
"actions": 3.728,
|
| 785 |
+
"invalid": 0.06,
|
| 786 |
+
"reasoning_bytes": 44.884,
|
| 787 |
+
"flops_per_problem": 3000792416.256
|
| 788 |
+
},
|
| 789 |
+
"fsf8": {
|
| 790 |
+
"acc": 0.952,
|
| 791 |
+
"actions": 3.566,
|
| 792 |
+
"invalid": 0.09,
|
| 793 |
+
"reasoning_bytes": 42.816,
|
| 794 |
+
"flops_per_problem": 2942408835.072
|
| 795 |
+
}
|
| 796 |
+
}
|