Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
c22ea1c
·
verified ·
1 Parent(s): 0b9d493

surface arm bert/mse_solo_o0_norm: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/bert/mse_solo_o0_norm/closes/sarm_bert_mse_solo_o0_norm_close1_s13_bert.safetensors +3 -0
  2. surface/bert/mse_solo_o0_norm/closes/sarm_bert_mse_solo_o0_norm_close2_s13_bert.safetensors +3 -0
  3. surface/bert/mse_solo_o0_norm/closes/sarm_bert_mse_solo_o0_norm_close3_s13_bert.safetensors +3 -0
  4. surface/bert/mse_solo_o0_norm/closes/sarm_bert_mse_solo_o0_norm_close4_s13_bert.safetensors +3 -0
  5. surface/bert/mse_solo_o0_norm/manifest.json +20 -0
  6. surface/bert/mse_solo_o0_norm/readout.md +127 -0
  7. surface/bert/mse_solo_o0_norm/results.json +0 -0
  8. surface/bert/mse_solo_o0_norm/s13_bert.safetensors +3 -0
  9. surface/bert/mse_solo_o0_norm/stats.pt +3 -0
  10. surface/bert/mse_solo_o0_norm/train.log +160 -0
surface/bert/mse_solo_o0_norm/closes/sarm_bert_mse_solo_o0_norm_close1_s13_bert.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f77cff3f463dbb1542ac4f3bafc7f167024d34af11500e9fbd04259b661b3a0d
3
+ size 54819320
surface/bert/mse_solo_o0_norm/closes/sarm_bert_mse_solo_o0_norm_close2_s13_bert.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d152c35f1991c75ff29000a41b9fb4a56978985a106a10bebec429ff44e24ede
3
+ size 54819320
surface/bert/mse_solo_o0_norm/closes/sarm_bert_mse_solo_o0_norm_close3_s13_bert.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:22d56b276648fde23353dfe9637dfff1b4bdebed214e40361fa82ae7f6f6c6f3
3
+ size 54819320
surface/bert/mse_solo_o0_norm/closes/sarm_bert_mse_solo_o0_norm_close4_s13_bert.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb4a72da7410e1d43f0fe73b77a913d8a1238ef29f748f39001af61c2f35fba1
3
+ size 54819320
surface/bert/mse_solo_o0_norm/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "sarm_bert_mse_solo_o0_norm",
3
+ "run": "mse_solo_o0_norm",
4
+ "convention": "bert",
5
+ "member": "s13_bert",
6
+ "final_anchor": "sarm_bert_mse_solo_o0_norm_close4_s13_bert.safetensors",
7
+ "closes": [
8
+ "sarm_bert_mse_solo_o0_norm_close1_s13_bert.safetensors",
9
+ "sarm_bert_mse_solo_o0_norm_close2_s13_bert.safetensors",
10
+ "sarm_bert_mse_solo_o0_norm_close3_s13_bert.safetensors",
11
+ "sarm_bert_mse_solo_o0_norm_close4_s13_bert.safetensors"
12
+ ],
13
+ "finished": true,
14
+ "form": "mse",
15
+ "base": "solo",
16
+ "steps": 4000,
17
+ "trunk_step": 245674,
18
+ "shipped": "2026-10-09 14:22 local",
19
+ "note": "every close's anchor ships (mids included); the final close's anchor is the run's s13_bert.safetensors"
20
+ }
surface/bert/mse_solo_o0_norm/readout.md ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # The surface arm's readout: bert-base-uncased's spelling (the WordPiece convention)
2
+
3
+ A surface arm makes the trunk read a tokenizer's spelling of a text as it reads the text's own bytes. Every number below is read on the dual-extraction read's held-out captions (512 of each COCO draw, never trained on; draw 2 is the repeat), one document a row, at every token's closing byte, as whitened Procrustes alignment (1 = the same geometry up to a rotation; two folds by caption). The pre-registered bars are in the plan; B3 and B4 (against Qwen's own states, the mood axis) are the instrument's read.
4
+
5
+ ## The runs
6
+
7
+ - sarm_bert_mse_solo_o0_norm: per-site MSE; alone on the bare trunk; blocks 0-31; 4000 of 4000 steps; seed offset 0
8
+
9
+ ## The verdicts (the final close of each run)
10
+
11
+ ### sarm_bert_mse_solo_o0_norm
12
+
13
+ - B1 the gap: NOT MET over every served block >= 12 (the lowest 0.764 at block 31 on draw 1, 0.778 at block 31 on draw 2; bar 0.85); over blocks 12-28 alone: MET
14
+ - B2 silence, whitened: NOT MET (the lowest A alignment 0.9819 at block 31 on draw 1, 0.9845 on draw 2; bar 0.995)
15
+ - B2 silence, in bits: MET (web +0.00050 bpb; bar +0.012)
16
+
17
+ ## The closes
18
+
19
+ ### sarm_bert_mse_solo_o0_norm
20
+
21
+ | close | step | B1 lowest (draw 1 / 2) | B1 at 12 / 20 / 28 (draw 1) | B2 lowest (draw 1 / 2) | web bits | gate mean / widest |
22
+ |---|---|---|---|---|---|---|
23
+ | 0 | 0 | 0.710 / 0.733 | 0.883 / 0.874 / 0.818 | 1.0000 / 1.0000 | +0.00000 | 0.047 / 0.047 |
24
+ | 1 | 1000 | 0.737 / 0.757 | 0.903 / 0.901 / 0.849 | 0.9890 / 0.9885 | +0.00051 | 0.163 / 0.217 |
25
+ | 2 | 2000 | 0.753 / 0.769 | 0.909 / 0.910 / 0.861 | 0.9862 / 0.9864 | +0.00062 | 0.258 / 0.516 |
26
+ | 3 | 3000 | 0.763 / 0.774 | 0.913 / 0.915 / 0.868 | 0.9816 / 0.9844 | +0.00057 | 0.349 / 0.698 |
27
+ | 4 | 4000 | 0.764 / 0.778 | 0.915 / 0.916 / 0.870 | 0.9819 / 0.9845 | +0.00050 | 0.433 / 0.849 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### sarm_bert_mse_solo_o0_norm, draw1 (512 captions, 4,966 sites), close 4 at step 4000
32
+
33
+ | block | gap before | armed B vs masked A | armed B vs armed A | A armed vs masked | mse | cos |
34
+ |---|---|---|---|---|---|---|
35
+ | 0 | 0.802 | 0.802 | 0.803 | 0.9994 | 0.253 | 0.903 |
36
+ | 4 | 0.873 | 0.885 | 0.886 | 0.9978 | 0.121 | 0.946 |
37
+ | 8 | 0.876 | 0.898 | 0.899 | 0.9975 | 0.117 | 0.949 |
38
+ | 12 | 0.883 | 0.915 | 0.918 | 0.9962 | 0.111 | 0.951 |
39
+ | 16 | 0.883 | 0.921 | 0.924 | 0.9955 | 0.107 | 0.949 |
40
+ | 18 | 0.877 | 0.918 | 0.921 | 0.9951 | 0.112 | 0.947 |
41
+ | 20 | 0.874 | 0.916 | 0.920 | 0.9920 | 0.117 | 0.945 |
42
+ | 22 | 0.868 | 0.912 | 0.916 | 0.9932 | 0.139 | 0.940 |
43
+ | 24 | 0.861 | 0.909 | 0.914 | 0.9928 | 0.148 | 0.935 |
44
+ | 26 | 0.838 | 0.886 | 0.893 | 0.9899 | 0.185 | 0.921 |
45
+ | 28 | 0.818 | 0.870 | 0.876 | 0.9893 | 0.216 | 0.908 |
46
+ | 29 | 0.804 | 0.857 | 0.862 | 0.9868 | 0.233 | 0.899 |
47
+ | 30 | 0.765 | 0.822 | 0.829 | 0.9860 | 0.264 | 0.892 |
48
+ | 31 | 0.710 | 0.764 | 0.770 | 0.9819 | 0.308 | 0.883 |
49
+
50
+ ### sarm_bert_mse_solo_o0_norm, draw2 (512 captions, 5,109 sites), close 4 at step 4000
51
+
52
+ | block | gap before | armed B vs masked A | armed B vs armed A | A armed vs masked | mse | cos |
53
+ |---|---|---|---|---|---|---|
54
+ | 0 | 0.790 | 0.790 | 0.791 | 0.9994 | 0.241 | 0.908 |
55
+ | 4 | 0.868 | 0.877 | 0.878 | 0.9980 | 0.118 | 0.948 |
56
+ | 8 | 0.876 | 0.894 | 0.895 | 0.9976 | 0.114 | 0.950 |
57
+ | 12 | 0.884 | 0.912 | 0.912 | 0.9959 | 0.110 | 0.951 |
58
+ | 16 | 0.881 | 0.918 | 0.921 | 0.9964 | 0.106 | 0.949 |
59
+ | 18 | 0.876 | 0.914 | 0.918 | 0.9951 | 0.111 | 0.947 |
60
+ | 20 | 0.873 | 0.914 | 0.918 | 0.9952 | 0.117 | 0.945 |
61
+ | 22 | 0.869 | 0.911 | 0.914 | 0.9948 | 0.136 | 0.941 |
62
+ | 24 | 0.860 | 0.903 | 0.908 | 0.9912 | 0.145 | 0.936 |
63
+ | 26 | 0.839 | 0.883 | 0.888 | 0.9926 | 0.179 | 0.923 |
64
+ | 28 | 0.825 | 0.872 | 0.878 | 0.9863 | 0.208 | 0.910 |
65
+ | 29 | 0.813 | 0.860 | 0.867 | 0.9880 | 0.222 | 0.902 |
66
+ | 30 | 0.776 | 0.827 | 0.832 | 0.9868 | 0.249 | 0.896 |
67
+ | 31 | 0.733 | 0.778 | 0.785 | 0.9845 | 0.279 | 0.887 |
68
+
69
+ The member's share in bits per byte (all on minus the member masked): web +0.00050
70
+
71
+ ## The training curves
72
+
73
+ ### sarm_bert_mse_solo_o0_norm
74
+
75
+ | step | task loss | mse | cos | InfoNCE | quiet on A | quiet off-domain |
76
+ |---|---|---|---|---|---|---|
77
+ | 1 | 0.6411 | 0.641 | 0.690 | None | 0.00000 | 0.00000 |
78
+ | 50 | 0.6489 | 0.649 | 0.687 | None | 0.00035 | 0.00025 |
79
+ | 250 | 0.5755 | 0.576 | 0.722 | None | 0.00177 | 0.00071 |
80
+ | 500 | 0.4996 | 0.500 | 0.760 | None | 0.00138 | 0.00051 |
81
+ | 1000 | 0.3456 | 0.346 | 0.841 | None | 0.00194 | 0.00088 |
82
+ | 1500 | 0.2903 | 0.290 | 0.869 | None | 0.00167 | 0.00088 |
83
+ | 2000 | 0.2657 | 0.266 | 0.880 | None | 0.00180 | 0.00109 |
84
+ | 2500 | 0.2527 | 0.253 | 0.887 | None | 0.00185 | 0.00084 |
85
+ | 3000 | 0.2493 | 0.249 | 0.889 | None | 0.00199 | 0.00088 |
86
+ | 3500 | 0.2326 | 0.233 | 0.897 | None | 0.00158 | 0.00082 |
87
+ | 4000 | 0.2263 | 0.226 | 0.900 | None | 0.00155 | 0.00080 |
88
+ | 1 | 0.6411 | 0.641 | 0.690 | None | 0.00000 | 0.00000 |
89
+ | 50 | 0.6489 | 0.649 | 0.687 | None | 0.00035 | 0.00025 |
90
+ | 250 | 0.5755 | 0.576 | 0.722 | None | 0.00177 | 0.00071 |
91
+ | 500 | 0.4996 | 0.500 | 0.760 | None | 0.00138 | 0.00051 |
92
+ | 1000 | 0.3456 | 0.346 | 0.841 | None | 0.00194 | 0.00088 |
93
+ | 1500 | 0.2903 | 0.290 | 0.869 | None | 0.00167 | 0.00088 |
94
+ | 2000 | 0.2657 | 0.266 | 0.880 | None | 0.00180 | 0.00109 |
95
+ | 2500 | 0.2527 | 0.253 | 0.887 | None | 0.00185 | 0.00084 |
96
+ | 3000 | 0.2493 | 0.249 | 0.889 | None | 0.00199 | 0.00088 |
97
+ | 3500 | 0.2326 | 0.233 | 0.897 | None | 0.00158 | 0.00082 |
98
+ | 4000 | 0.2263 | 0.226 | 0.900 | None | 0.00155 | 0.00080 |
99
+ | 1 | 0.6411 | 0.641 | 0.690 | None | 0.00000 | 0.00000 |
100
+ | 50 | 0.6489 | 0.649 | 0.687 | None | 0.00035 | 0.00025 |
101
+ | 250 | 0.5755 | 0.576 | 0.722 | None | 0.00177 | 0.00071 |
102
+ | 500 | 0.4996 | 0.500 | 0.760 | None | 0.00138 | 0.00051 |
103
+ | 1000 | 0.3456 | 0.346 | 0.841 | None | 0.00194 | 0.00088 |
104
+ | 1500 | 0.2903 | 0.290 | 0.869 | None | 0.00167 | 0.00088 |
105
+ | 2000 | 0.2657 | 0.266 | 0.880 | None | 0.00180 | 0.00109 |
106
+ | 2500 | 0.2527 | 0.253 | 0.887 | None | 0.00185 | 0.00084 |
107
+ | 3000 | 0.2493 | 0.249 | 0.889 | None | 0.00199 | 0.00088 |
108
+ | 3500 | 0.2326 | 0.233 | 0.897 | None | 0.00158 | 0.00082 |
109
+ | 4000 | 0.2263 | 0.226 | 0.900 | None | 0.00155 | 0.00080 |
110
+ | 1 | 0.6411 | 0.641 | 0.690 | None | 0.00000 | 0.00000 |
111
+ | 50 | 0.6489 | 0.649 | 0.687 | None | 0.00035 | 0.00025 |
112
+ | 250 | 0.5755 | 0.576 | 0.722 | None | 0.00177 | 0.00071 |
113
+ | 500 | 0.4996 | 0.500 | 0.760 | None | 0.00138 | 0.00051 |
114
+ | 1000 | 0.3456 | 0.346 | 0.841 | None | 0.00194 | 0.00088 |
115
+ | 1500 | 0.2903 | 0.290 | 0.869 | None | 0.00167 | 0.00088 |
116
+ | 2000 | 0.2657 | 0.266 | 0.880 | None | 0.00180 | 0.00109 |
117
+ | 2500 | 0.2527 | 0.253 | 0.887 | None | 0.00185 | 0.00084 |
118
+ | 3000 | 0.2493 | 0.249 | 0.889 | None | 0.00199 | 0.00088 |
119
+ | 3500 | 0.2326 | 0.233 | 0.897 | None | 0.00158 | 0.00082 |
120
+ | 4000 | 0.2263 | 0.226 | 0.900 | None | 0.00155 | 0.00080 |
121
+
122
+ ## The recipe
123
+
124
+ - per step: the teacher (member masked) reads the A rows: block outputs at the sites + logits; the armed model reads the B rows: per-site MSE of standardized LayerNorm'd block outputs against the teacher's at each token's closing byte, mean over served blocks; quiet on A: lambda x KL(teacher || armed) on the A rows; quiet off-domain: lambda x KL(member masked || armed) on one web row + one partner row
125
+ - lambda 2.0; pure Adam lr 0.001 wd 0; 2 x 4096-byte paired rows a step; the arm LAWFUL_16x8 at every block ({'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}); the trunk mini-beatrix-3 at step 245,674; alephllm 0.10.7, amoe-lora 0.2.11, torch 2.8.0+cu128
126
+ - the surface: bert-base-uncased's spelling (the WordPiece convention; sha256 9114256b104cca80..., vocab 30,522; A = the text normalized; sites = tokens); documents = one caption-like field of a caption-pack row (caption, short, long, tags, medium), sources drawn by {'dff': 0.161, 'scf': 0.161, 'danbooru': 0.164, 'cc12m': 0.406, 'coco': 0.108}
127
+ - the test set: the first 512 captions of each draw of coco_caps_2x2048.json (every caption of the file kept out of training)
surface/bert/mse_solo_o0_norm/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/bert/mse_solo_o0_norm/s13_bert.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb4a72da7410e1d43f0fe73b77a913d8a1238ef29f748f39001af61c2f35fba1
3
+ size 54819320
surface/bert/mse_solo_o0_norm/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:086c36d130998f2eb569e7ef49863c95788d0bbf0999061e92eb48ed8ee916e5
3
+ size 264365
surface/bert/mse_solo_o0_norm/train.log ADDED
@@ -0,0 +1,160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 0.5s] THE JOB: sarm_bert_mse_solo_o0_norm: the surface arm s13_bert (wordpiece spelling by bert-base-uncased; A = the text normalized; sites = tokens) alone on the bare trunk; form mse; blocks all; 4000 steps, a close every 1000; lambda 2.0; pure Adam lr 0.001; the compiled chain with recompute; card 30.5 GB with 30.4 GB free; host RSS 0.6 GB
2
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
3
+ [ 16.8s] mini-beatrix-3 @ 245,674: trunk loaded from /workspace/qarm/dl/mini-beatrix-3/checkpoints/step_00245674.safetensors and frozen (376.0M); alephllm 0.10.7 at /workspace/qarm/alephllm/geolip/alephllm; amoe-lora 0.2.11; torch 2.8.0+cu128
4
+ [ 17.3s] tokenizer bert-base-uncased (wordpiece spelling; vocab 30522; sha256 9114256b104cca80): 'a taco truck' -> ['a', 'ta', '##co', 'truck']; A 12 bytes, B 15 bytes; sites [(1, 4, 4), (2, 6, 9)]
5
+ [ 17.3s] the read's captions: 512 of each draw of /workspace/qarm/dl/coco_caps_2x2048.json (4,090 captions of both draws kept out of training)
6
+ [ 30.3s] caption pack (/workspace/qarm/dl/mini-beatrix-3/caption_pack/pack_v2, train): dff 121,153 rows, scf 42,080 rows, danbooru 603,026 rows, cc12m 1,242,176 rows, coco 19,227 rows; host RSS 3.6 GB
7
+ [ 30.5s] streams: web text; partners none (two web rows a quiet chunk); held-out sets ['web'] x 8 batches
8
+ [ 30.6s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 108), 4,966 sites
9
+ [ 30.6s] read rows draw2: 512 captions (0 unspellable), A rows (512, 151), B rows (512, 161), 5,109 sites
10
+ [arms] attached 's13_bert' for phase 'wordpiece_spelling': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
11
+ [ 42.3s] BIRTH: the member attached over the trunk: web rows read exactly as before; the hand-run trunk against the model's forward: armed 0.0e+00, masked 0.0e+00 (gates: 0.0)
12
+ [ 46.8s] standardization statistics from 12,934 teacher sites on 8 paired batches (a mean and a spread per block and dimension); sigma range at block 20 0.741..1.136; saved /workspace/qarm/out/anchors/sarm_bert_mse_solo_o0_norm_stats.pt
13
+ [ 250.2s] CLOSE 0 (step 0): the member's alignment read on draw 1 (512 captions, 4,966 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
14
+ [ 250.2s] block 8: 0.876 -> 0.876 / 0.876; A 1.0000; mse 0.170 cos 0.925
15
+ [ 250.2s] block 12: 0.883 -> 0.883 / 0.883; A 1.0000; mse 0.172 cos 0.923
16
+ [ 250.2s] block 16: 0.883 -> 0.883 / 0.883; A 1.0000; mse 0.175 cos 0.917
17
+ [ 250.2s] block 18: 0.877 -> 0.877 / 0.877; A 1.0000; mse 0.185 cos 0.912
18
+ [ 250.2s] block 20: 0.874 -> 0.874 / 0.874; A 1.0000; mse 0.197 cos 0.907
19
+ [ 250.2s] block 24: 0.861 -> 0.861 / 0.861; A 1.0000; mse 0.246 cos 0.891
20
+ [ 250.2s] block 28: 0.818 -> 0.818 / 0.818; A 1.0000; mse 0.341 cos 0.849
21
+ [ 250.2s] block 31: 0.710 -> 0.710 / 0.710; A 1.0000; mse 0.425 cos 0.820
22
+ [ 250.2s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.884 / 0.881 / 0.873 / 0.860 / 0.825
23
+ [ 250.2s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.710 against the bar .85; B2: the lowest A alignment 1.0000 against .995; fineweb share +0.00000 +- 0.00000 against +.012
24
+ [ 250.2s] gate mean 0.047 / widest 0.047; compiled chain: 2 cache entries (limit 256), 0 limit warnings; docs drawn 256 (skipped {'test': 2, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 2.3 GB; host RSS 10.7 GB
25
+ [stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
26
+ [ 257.1s] step 1/4000: task 0.6411 (mse 0.641 cos 0.690); quiet A 0.00000 off 0.00000; 1,604 sites; 6.89 s/step; the job: 4.3 min spent, about 459.3 min of steps left; card peak 13.7 GB
27
+ [ 378.4s] step 50/4000: task 0.6489 (mse 0.649 cos 0.687); quiet A 0.00035 off 0.00025; 1,605 sites; 2.56 s/step; the job: 6.3 min spent, about 168.8 min of steps left; card peak 13.9 GB
28
+ [ 501.8s] step 100/4000: task 0.6255 (mse 0.626 cos 0.698); quiet A 0.00086 off 0.00048; 1,587 sites; 2.52 s/step; the job: 8.4 min spent, about 163.5 min of steps left; card peak 13.9 GB
29
+ [ 625.3s] step 150/4000: task 0.6060 (mse 0.606 cos 0.708); quiet A 0.00123 off 0.00063; 1,615 sites; 2.50 s/step; the job: 10.4 min spent, about 160.5 min of steps left; card peak 13.9 GB
30
+ [ 749.0s] step 200/4000: task 0.5947 (mse 0.595 cos 0.713); quiet A 0.00157 off 0.00067; 1,615 sites; 2.49 s/step; the job: 12.5 min spent, about 157.9 min of steps left; card peak 13.9 GB
31
+ [ 872.4s] step 250/4000: task 0.5755 (mse 0.576 cos 0.722); quiet A 0.00177 off 0.00071; 1,588 sites; 2.49 s/step; the job: 14.5 min spent, about 155.6 min of steps left; card peak 13.9 GB
32
+ [ 995.7s] step 300/4000: task 0.5537 (mse 0.554 cos 0.734); quiet A 0.00193 off 0.00086; 1,602 sites; 2.48 s/step; the job: 16.6 min spent, about 153.2 min of steps left; card peak 13.9 GB
33
+ [ 1119.7s] step 350/4000: task 0.5411 (mse 0.541 cos 0.740); quiet A 0.00198 off 0.00082; 1,587 sites; 2.48 s/step; the job: 18.7 min spent, about 151.1 min of steps left; card peak 13.9 GB
34
+ [ 1243.0s] step 400/4000: task 0.5084 (mse 0.508 cos 0.755); quiet A 0.00169 off 0.00060; 1,615 sites; 2.48 s/step; the job: 20.7 min spent, about 148.9 min of steps left; card peak 13.9 GB
35
+ [ 1366.4s] step 450/4000: task 0.5130 (mse 0.513 cos 0.753); quiet A 0.00156 off 0.00056; 1,654 sites; 2.48 s/step; the job: 22.8 min spent, about 146.8 min of steps left; card peak 13.9 GB
36
+ [ 1489.8s] step 500/4000: task 0.4996 (mse 0.500 cos 0.760); quiet A 0.00138 off 0.00051; 1,617 sites; 2.48 s/step; the job: 24.8 min spent, about 144.6 min of steps left; card peak 13.9 GB
37
+ [ 1613.2s] step 550/4000: task 0.4743 (mse 0.474 cos 0.772); quiet A 0.00135 off 0.00054; 1,564 sites; 2.48 s/step; the job: 26.9 min spent, about 142.5 min of steps left; card peak 13.9 GB
38
+ [ 1736.5s] step 600/4000: task 0.4646 (mse 0.465 cos 0.778); quiet A 0.00143 off 0.00062; 1,624 sites; 2.48 s/step; the job: 28.9 min spent, about 140.4 min of steps left; card peak 13.9 GB
39
+ [ 1860.3s] step 650/4000: task 0.4519 (mse 0.452 cos 0.785); quiet A 0.00130 off 0.00056; 1,626 sites; 2.48 s/step; the job: 31.0 min spent, about 138.3 min of steps left; card peak 13.9 GB
40
+ [transformers] Token indices sequence length is longer than the specified maximum sequence length for this model (558 > 512). Running this sequence through the model will result in indexing errors
41
+ [ 1983.8s] step 700/4000: task 0.4369 (mse 0.437 cos 0.792); quiet A 0.00121 off 0.00058; 1,582 sites; 2.48 s/step; the job: 33.1 min spent, about 136.2 min of steps left; card peak 13.9 GB
42
+ [ 2107.3s] step 750/4000: task 0.4145 (mse 0.414 cos 0.804); quiet A 0.00130 off 0.00065; 1,604 sites; 2.48 s/step; the job: 35.1 min spent, about 134.1 min of steps left; card peak 13.9 GB
43
+ [ 2230.8s] step 800/4000: task 0.3910 (mse 0.391 cos 0.815); quiet A 0.00135 off 0.00067; 1,630 sites; 2.48 s/step; the job: 37.2 min spent, about 132.0 min of steps left; card peak 13.9 GB
44
+ [ 2354.4s] step 850/4000: task 0.3751 (mse 0.375 cos 0.824); quiet A 0.00138 off 0.00067; 1,563 sites; 2.48 s/step; the job: 39.2 min spent, about 130.0 min of steps left; card peak 13.9 GB
45
+ [ 2478.0s] step 900/4000: task 0.3698 (mse 0.370 cos 0.827); quiet A 0.00153 off 0.00080; 1,351 sites; 2.48 s/step; the job: 41.3 min spent, about 127.9 min of steps left; card peak 13.9 GB
46
+ [ 2601.4s] step 950/4000: task 0.3529 (mse 0.353 cos 0.837); quiet A 0.00179 off 0.00070; 1,626 sites; 2.48 s/step; the job: 43.4 min spent, about 125.8 min of steps left; card peak 13.9 GB
47
+ [ 2724.9s] step 1000/4000: task 0.3456 (mse 0.346 cos 0.841); quiet A 0.00194 off 0.00088; 1,628 sites; 2.47 s/step; the job: 45.4 min spent, about 123.7 min of steps left; card peak 13.9 GB
48
+ [ 2725.2s] anchor saved: /workspace/qarm/out/anchors/sarm_bert_mse_solo_o0_norm_close1_s13_bert.safetensors
49
+ [ 2879.9s] CLOSE 1 (step 1000): the member's alignment read on draw 1 (512 captions, 4,966 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
50
+ [ 2879.9s] block 8: 0.876 -> 0.888 / 0.887; A 0.9973; mse 0.136 cos 0.940
51
+ [ 2879.9s] block 12: 0.883 -> 0.903 / 0.905; A 0.9968; mse 0.131 cos 0.941
52
+ [ 2879.9s] block 16: 0.883 -> 0.907 / 0.909; A 0.9961; mse 0.130 cos 0.938
53
+ [ 2879.9s] block 18: 0.877 -> 0.904 / 0.906; A 0.9944; mse 0.136 cos 0.935
54
+ [ 2879.9s] block 20: 0.874 -> 0.901 / 0.903; A 0.9950; mse 0.144 cos 0.933
55
+ [ 2879.9s] block 24: 0.861 -> 0.891 / 0.894; A 0.9948; mse 0.182 cos 0.920
56
+ [ 2879.9s] block 28: 0.818 -> 0.849 / 0.852; A 0.9911; mse 0.263 cos 0.888
57
+ [ 2879.9s] block 31: 0.710 -> 0.737 / 0.742; A 0.9890; mse 0.360 cos 0.859
58
+ [ 2879.9s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.901 / 0.904 / 0.899 / 0.887 / 0.853
59
+ [ 2879.9s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.737 against the bar .85; B2: the lowest A alignment 0.9890 against .995; fineweb share +0.00051 +- 0.00017 against +.012
60
+ [ 2879.9s] gate mean 0.163 / widest 0.217; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 32,523 (skipped {'test': 44, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 13.2 GB
61
+ [ 3003.3s] step 1050/4000: task 0.3337 (mse 0.334 cos 0.847); quiet A 0.00171 off 0.00080; 1,613 sites; 2.47 s/step; the job: 50.1 min spent, about 121.7 min of steps left; card peak 13.9 GB
62
+ [ 3126.7s] step 1100/4000: task 0.3236 (mse 0.324 cos 0.851); quiet A 0.00166 off 0.00079; 1,592 sites; 2.47 s/step; the job: 52.1 min spent, about 119.6 min of steps left; card peak 13.9 GB
63
+ [ 3250.3s] step 1150/4000: task 0.3137 (mse 0.314 cos 0.856); quiet A 0.00176 off 0.00083; 1,613 sites; 2.47 s/step; the job: 54.2 min spent, about 117.5 min of steps left; card peak 13.9 GB
64
+ [ 3374.0s] step 1200/4000: task 0.3178 (mse 0.318 cos 0.854); quiet A 0.00174 off 0.00087; 1,600 sites; 2.47 s/step; the job: 56.2 min spent, about 115.5 min of steps left; card peak 13.9 GB
65
+ [ 3497.5s] step 1250/4000: task 0.3148 (mse 0.315 cos 0.856); quiet A 0.00167 off 0.00083; 1,563 sites; 2.47 s/step; the job: 58.3 min spent, about 113.4 min of steps left; card peak 13.9 GB
66
+ [ 3621.0s] step 1300/4000: task 0.3086 (mse 0.309 cos 0.860); quiet A 0.00173 off 0.00080; 1,609 sites; 2.47 s/step; the job: 60.3 min spent, about 111.3 min of steps left; card peak 13.9 GB
67
+ [ 3744.8s] step 1350/4000: task 0.3032 (mse 0.303 cos 0.862); quiet A 0.00204 off 0.00087; 1,581 sites; 2.47 s/step; the job: 62.4 min spent, about 109.3 min of steps left; card peak 13.9 GB
68
+ [ 3868.5s] step 1400/4000: task 0.3006 (mse 0.301 cos 0.863); quiet A 0.00220 off 0.00094; 1,610 sites; 2.47 s/step; the job: 64.5 min spent, about 107.2 min of steps left; card peak 13.9 GB
69
+ [ 3992.3s] step 1450/4000: task 0.3001 (mse 0.300 cos 0.864); quiet A 0.00183 off 0.00085; 1,611 sites; 2.47 s/step; the job: 66.5 min spent, about 105.1 min of steps left; card peak 13.9 GB
70
+ [ 4115.9s] step 1500/4000: task 0.2903 (mse 0.290 cos 0.869); quiet A 0.00167 off 0.00088; 1,551 sites; 2.47 s/step; the job: 68.6 min spent, about 103.1 min of steps left; card peak 13.9 GB
71
+ [ 4239.3s] step 1550/4000: task 0.2878 (mse 0.288 cos 0.870); quiet A 0.00167 off 0.00082; 1,577 sites; 2.47 s/step; the job: 70.7 min spent, about 101.0 min of steps left; card peak 13.9 GB
72
+ [ 4362.7s] step 1600/4000: task 0.2849 (mse 0.285 cos 0.871); quiet A 0.00172 off 0.00076; 1,614 sites; 2.47 s/step; the job: 72.7 min spent, about 98.9 min of steps left; card peak 13.9 GB
73
+ [ 4486.3s] step 1650/4000: task 0.2840 (mse 0.284 cos 0.871); quiet A 0.00165 off 0.00078; 1,647 sites; 2.47 s/step; the job: 74.8 min spent, about 96.9 min of steps left; card peak 13.9 GB
74
+ [ 4609.7s] step 1700/4000: task 0.2849 (mse 0.285 cos 0.872); quiet A 0.00180 off 0.00086; 1,594 sites; 2.47 s/step; the job: 76.8 min spent, about 94.8 min of steps left; card peak 13.9 GB
75
+ [ 4733.3s] step 1750/4000: task 0.2738 (mse 0.274 cos 0.876); quiet A 0.00171 off 0.00081; 1,604 sites; 2.47 s/step; the job: 78.9 min spent, about 92.7 min of steps left; card peak 13.9 GB
76
+ [ 4856.8s] step 1800/4000: task 0.2708 (mse 0.271 cos 0.878); quiet A 0.00168 off 0.00078; 1,644 sites; 2.47 s/step; the job: 80.9 min spent, about 90.7 min of steps left; card peak 13.9 GB
77
+ [ 4980.2s] step 1850/4000: task 0.2804 (mse 0.280 cos 0.874); quiet A 0.00167 off 0.00084; 1,581 sites; 2.47 s/step; the job: 83.0 min spent, about 88.6 min of steps left; card peak 13.9 GB
78
+ [ 5103.7s] step 1900/4000: task 0.2724 (mse 0.272 cos 0.877); quiet A 0.00159 off 0.00080; 1,666 sites; 2.47 s/step; the job: 85.1 min spent, about 86.6 min of steps left; card peak 13.9 GB
79
+ [ 5227.3s] step 1950/4000: task 0.2694 (mse 0.269 cos 0.879); quiet A 0.00159 off 0.00085; 1,612 sites; 2.47 s/step; the job: 87.1 min spent, about 84.5 min of steps left; card peak 13.9 GB
80
+ [ 5351.0s] step 2000/4000: task 0.2657 (mse 0.266 cos 0.880); quiet A 0.00180 off 0.00109; 1,636 sites; 2.47 s/step; the job: 89.2 min spent, about 82.4 min of steps left; card peak 13.9 GB
81
+ [ 5351.2s] anchor saved: /workspace/qarm/out/anchors/sarm_bert_mse_solo_o0_norm_close2_s13_bert.safetensors
82
+ [ 5506.0s] CLOSE 2 (step 2000): the member's alignment read on draw 1 (512 captions, 4,966 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
83
+ [ 5506.0s] block 8: 0.876 -> 0.893 / 0.894; A 0.9973; mse 0.124 cos 0.946
84
+ [ 5506.0s] block 12: 0.883 -> 0.909 / 0.912; A 0.9963; mse 0.119 cos 0.947
85
+ [ 5506.0s] block 16: 0.883 -> 0.915 / 0.918; A 0.9962; mse 0.115 cos 0.946
86
+ [ 5506.0s] block 18: 0.877 -> 0.913 / 0.916; A 0.9950; mse 0.121 cos 0.943
87
+ [ 5506.0s] block 20: 0.874 -> 0.910 / 0.914; A 0.9933; mse 0.127 cos 0.941
88
+ [ 5506.0s] block 24: 0.861 -> 0.902 / 0.907; A 0.9935; mse 0.161 cos 0.930
89
+ [ 5506.0s] block 28: 0.818 -> 0.861 / 0.867; A 0.9889; mse 0.236 cos 0.900
90
+ [ 5506.0s] block 31: 0.710 -> 0.753 / 0.756; A 0.9862; mse 0.329 cos 0.873
91
+ [ 5506.0s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.906 / 0.912 / 0.907 / 0.895 / 0.864
92
+ [ 5506.0s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.753 against the bar .85; B2: the lowest A alignment 0.9862 against .995; fineweb share +0.00062 +- 0.00017 against +.012
93
+ [ 5506.0s] gate mean 0.258 / widest 0.516; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 64,565 (skipped {'test': 97, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 13.2 GB
94
+ [ 5630.0s] step 2050/4000: task 0.2654 (mse 0.265 cos 0.880); quiet A 0.00162 off 0.00081; 1,608 sites; 2.48 s/step; the job: 93.8 min spent, about 80.4 min of steps left; card peak 13.9 GB
95
+ [ 5753.5s] step 2100/4000: task 0.2716 (mse 0.272 cos 0.878); quiet A 0.00168 off 0.00086; 1,613 sites; 2.47 s/step; the job: 95.9 min spent, about 78.3 min of steps left; card peak 13.9 GB
96
+ [ 5877.1s] step 2150/4000: task 0.2603 (mse 0.260 cos 0.883); quiet A 0.00162 off 0.00080; 1,589 sites; 2.47 s/step; the job: 98.0 min spent, about 76.2 min of steps left; card peak 13.9 GB
97
+ [ 6000.6s] step 2200/4000: task 0.2691 (mse 0.269 cos 0.879); quiet A 0.00166 off 0.00077; 1,600 sites; 2.47 s/step; the job: 100.0 min spent, about 74.2 min of steps left; card peak 13.9 GB
98
+ [ 6124.1s] step 2250/4000: task 0.2706 (mse 0.271 cos 0.879); quiet A 0.00173 off 0.00085; 1,638 sites; 2.47 s/step; the job: 102.1 min spent, about 72.1 min of steps left; card peak 13.9 GB
99
+ [ 6247.5s] step 2300/4000: task 0.2609 (mse 0.261 cos 0.883); quiet A 0.00171 off 0.00082; 1,600 sites; 2.47 s/step; the job: 104.1 min spent, about 70.1 min of steps left; card peak 13.9 GB
100
+ [ 6371.1s] step 2350/4000: task 0.2552 (mse 0.255 cos 0.886); quiet A 0.00159 off 0.00082; 1,623 sites; 2.47 s/step; the job: 106.2 min spent, about 68.0 min of steps left; card peak 13.9 GB
101
+ [ 6494.5s] step 2400/4000: task 0.2484 (mse 0.248 cos 0.889); quiet A 0.00161 off 0.00091; 1,591 sites; 2.47 s/step; the job: 108.2 min spent, about 65.9 min of steps left; card peak 13.9 GB
102
+ [ 6618.0s] step 2450/4000: task 0.2504 (mse 0.250 cos 0.888); quiet A 0.00156 off 0.00085; 1,600 sites; 2.47 s/step; the job: 110.3 min spent, about 63.9 min of steps left; card peak 13.9 GB
103
+ [ 6741.7s] step 2500/4000: task 0.2527 (mse 0.253 cos 0.887); quiet A 0.00185 off 0.00084; 1,612 sites; 2.47 s/step; the job: 112.4 min spent, about 61.8 min of steps left; card peak 13.9 GB
104
+ [ 6865.3s] step 2550/4000: task 0.2569 (mse 0.257 cos 0.885); quiet A 0.00158 off 0.00082; 1,627 sites; 2.47 s/step; the job: 114.4 min spent, about 59.8 min of steps left; card peak 13.9 GB
105
+ [ 6989.1s] step 2600/4000: task 0.2574 (mse 0.257 cos 0.885); quiet A 0.00172 off 0.00085; 1,603 sites; 2.47 s/step; the job: 116.5 min spent, about 57.7 min of steps left; card peak 13.9 GB
106
+ [ 7112.7s] step 2650/4000: task 0.2469 (mse 0.247 cos 0.890); quiet A 0.00157 off 0.00085; 1,609 sites; 2.47 s/step; the job: 118.5 min spent, about 55.6 min of steps left; card peak 13.9 GB
107
+ [ 7236.2s] step 2700/4000: task 0.2507 (mse 0.251 cos 0.888); quiet A 0.00168 off 0.00081; 1,606 sites; 2.47 s/step; the job: 120.6 min spent, about 53.6 min of steps left; card peak 13.9 GB
108
+ [ 7359.9s] step 2750/4000: task 0.2454 (mse 0.245 cos 0.890); quiet A 0.00164 off 0.00084; 1,638 sites; 2.47 s/step; the job: 122.7 min spent, about 51.5 min of steps left; card peak 13.9 GB
109
+ [ 7483.4s] step 2800/4000: task 0.2441 (mse 0.244 cos 0.891); quiet A 0.00163 off 0.00079; 1,574 sites; 2.47 s/step; the job: 124.7 min spent, about 49.5 min of steps left; card peak 13.9 GB
110
+ [ 7606.8s] step 2850/4000: task 0.2452 (mse 0.245 cos 0.891); quiet A 0.00164 off 0.00078; 1,597 sites; 2.47 s/step; the job: 126.8 min spent, about 47.4 min of steps left; card peak 13.9 GB
111
+ [ 7730.8s] step 2900/4000: task 0.2465 (mse 0.246 cos 0.890); quiet A 0.00202 off 0.00079; 1,245 sites; 2.47 s/step; the job: 128.8 min spent, about 45.3 min of steps left; card peak 13.9 GB
112
+ [ 7854.3s] step 2950/4000: task 0.2377 (mse 0.238 cos 0.894); quiet A 0.00166 off 0.00087; 1,619 sites; 2.47 s/step; the job: 130.9 min spent, about 43.3 min of steps left; card peak 13.9 GB
113
+ [ 7977.9s] step 3000/4000: task 0.2493 (mse 0.249 cos 0.889); quiet A 0.00199 off 0.00088; 1,626 sites; 2.47 s/step; the job: 133.0 min spent, about 41.2 min of steps left; card peak 13.9 GB
114
+ [ 7978.1s] anchor saved: /workspace/qarm/out/anchors/sarm_bert_mse_solo_o0_norm_close3_s13_bert.safetensors
115
+ [ 8133.3s] CLOSE 3 (step 3000): the member's alignment read on draw 1 (512 captions, 4,966 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
116
+ [ 8133.3s] block 8: 0.876 -> 0.897 / 0.898; A 0.9972; mse 0.119 cos 0.948
117
+ [ 8133.3s] block 12: 0.883 -> 0.913 / 0.915; A 0.9960; mse 0.113 cos 0.950
118
+ [ 8133.3s] block 16: 0.883 -> 0.919 / 0.922; A 0.9941; mse 0.108 cos 0.949
119
+ [ 8133.3s] block 18: 0.877 -> 0.917 / 0.921; A 0.9943; mse 0.113 cos 0.946
120
+ [ 8133.3s] block 20: 0.874 -> 0.915 / 0.919; A 0.9932; mse 0.119 cos 0.945
121
+ [ 8133.3s] block 24: 0.861 -> 0.908 / 0.912; A 0.9927; mse 0.151 cos 0.934
122
+ [ 8133.3s] block 28: 0.818 -> 0.868 / 0.874; A 0.9891; mse 0.223 cos 0.905
123
+ [ 8133.3s] block 31: 0.710 -> 0.763 / 0.766; A 0.9816; mse 0.317 cos 0.879
124
+ [ 8133.3s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.909 / 0.916 / 0.912 / 0.900 / 0.869
125
+ [ 8133.3s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.763 against the bar .85; B2: the lowest A alignment 0.9816 against .995; fineweb share +0.00057 +- 0.00017 against +.012
126
+ [ 8133.3s] gate mean 0.349 / widest 0.698; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 96,466 (skipped {'test': 156, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 13.2 GB
127
+ [ 8257.0s] step 3050/4000: task 0.2342 (mse 0.234 cos 0.896); quiet A 0.00161 off 0.00077; 1,573 sites; 2.47 s/step; the job: 137.6 min spent, about 39.1 min of steps left; card peak 13.9 GB
128
+ [ 8380.7s] step 3100/4000: task 0.2426 (mse 0.243 cos 0.892); quiet A 0.00161 off 0.00088; 1,614 sites; 2.47 s/step; the job: 139.7 min spent, about 37.1 min of steps left; card peak 13.9 GB
129
+ [ 8504.4s] step 3150/4000: task 0.2390 (mse 0.239 cos 0.894); quiet A 0.00156 off 0.00075; 1,616 sites; 2.47 s/step; the job: 141.7 min spent, about 35.0 min of steps left; card peak 13.9 GB
130
+ [ 8628.3s] step 3200/4000: task 0.2405 (mse 0.240 cos 0.893); quiet A 0.00155 off 0.00089; 1,632 sites; 2.47 s/step; the job: 143.8 min spent, about 33.0 min of steps left; card peak 13.9 GB
131
+ [ 8751.9s] step 3250/4000: task 0.2316 (mse 0.232 cos 0.897); quiet A 0.00164 off 0.00086; 1,664 sites; 2.47 s/step; the job: 145.9 min spent, about 30.9 min of steps left; card peak 13.9 GB
132
+ [ 8875.4s] step 3300/4000: task 0.2295 (mse 0.230 cos 0.898); quiet A 0.00155 off 0.00080; 1,599 sites; 2.47 s/step; the job: 147.9 min spent, about 28.8 min of steps left; card peak 13.9 GB
133
+ [ 8998.8s] step 3350/4000: task 0.2372 (mse 0.237 cos 0.895); quiet A 0.00158 off 0.00075; 1,597 sites; 2.47 s/step; the job: 150.0 min spent, about 26.8 min of steps left; card peak 13.9 GB
134
+ [ 9122.5s] step 3400/4000: task 0.2358 (mse 0.236 cos 0.896); quiet A 0.00149 off 0.00077; 1,594 sites; 2.47 s/step; the job: 152.0 min spent, about 24.7 min of steps left; card peak 13.9 GB
135
+ [ 9246.0s] step 3450/4000: task 0.2331 (mse 0.233 cos 0.897); quiet A 0.00157 off 0.00085; 1,594 sites; 2.47 s/step; the job: 154.1 min spent, about 22.7 min of steps left; card peak 13.9 GB
136
+ [ 9369.8s] step 3500/4000: task 0.2326 (mse 0.233 cos 0.897); quiet A 0.00158 off 0.00082; 1,603 sites; 2.47 s/step; the job: 156.2 min spent, about 20.6 min of steps left; card peak 13.9 GB
137
+ [ 9493.5s] step 3550/4000: task 0.2318 (mse 0.232 cos 0.897); quiet A 0.00168 off 0.00072; 1,595 sites; 2.47 s/step; the job: 158.2 min spent, about 18.5 min of steps left; card peak 13.9 GB
138
+ [ 9617.2s] step 3600/4000: task 0.2272 (mse 0.227 cos 0.899); quiet A 0.00151 off 0.00079; 1,616 sites; 2.47 s/step; the job: 160.3 min spent, about 16.5 min of steps left; card peak 13.9 GB
139
+ [ 9740.7s] step 3650/4000: task 0.2328 (mse 0.233 cos 0.897); quiet A 0.00166 off 0.00084; 1,579 sites; 2.47 s/step; the job: 162.3 min spent, about 14.4 min of steps left; card peak 13.9 GB
140
+ [ 9864.3s] step 3700/4000: task 0.2294 (mse 0.229 cos 0.898); quiet A 0.00170 off 0.00095; 1,594 sites; 2.47 s/step; the job: 164.4 min spent, about 12.4 min of steps left; card peak 13.9 GB
141
+ [ 9987.8s] step 3750/4000: task 0.2212 (mse 0.221 cos 0.902); quiet A 0.00156 off 0.00079; 1,578 sites; 2.47 s/step; the job: 166.5 min spent, about 10.3 min of steps left; card peak 13.9 GB
142
+ [10111.7s] step 3800/4000: task 0.2230 (mse 0.223 cos 0.901); quiet A 0.00155 off 0.00071; 1,606 sites; 2.47 s/step; the job: 168.5 min spent, about 8.2 min of steps left; card peak 13.9 GB
143
+ [10235.3s] step 3850/4000: task 0.2276 (mse 0.228 cos 0.899); quiet A 0.00156 off 0.00074; 1,613 sites; 2.47 s/step; the job: 170.6 min spent, about 6.2 min of steps left; card peak 13.9 GB
144
+ [10359.0s] step 3900/4000: task 0.2236 (mse 0.224 cos 0.901); quiet A 0.00156 off 0.00075; 1,570 sites; 2.47 s/step; the job: 172.6 min spent, about 4.1 min of steps left; card peak 13.9 GB
145
+ [10482.5s] step 3950/4000: task 0.2177 (mse 0.218 cos 0.904); quiet A 0.00150 off 0.00074; 1,593 sites; 2.47 s/step; the job: 174.7 min spent, about 2.1 min of steps left; card peak 13.9 GB
146
+ [10606.0s] step 4000/4000: task 0.2263 (mse 0.226 cos 0.900); quiet A 0.00155 off 0.00080; 1,585 sites; 2.47 s/step; the job: 176.8 min spent, about 0.0 min of steps left; card peak 13.9 GB
147
+ [10606.3s] anchor saved: /workspace/qarm/out/anchors/sarm_bert_mse_solo_o0_norm_close4_s13_bert.safetensors
148
+ [10761.4s] CLOSE 4 (step 4000): the member's alignment read on draw 1 (512 captions, 4,966 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
149
+ [10761.4s] block 8: 0.876 -> 0.898 / 0.899; A 0.9975; mse 0.117 cos 0.949
150
+ [10761.4s] block 12: 0.883 -> 0.915 / 0.918; A 0.9962; mse 0.111 cos 0.951
151
+ [10761.4s] block 16: 0.883 -> 0.921 / 0.924; A 0.9955; mse 0.107 cos 0.949
152
+ [10761.4s] block 18: 0.877 -> 0.918 / 0.921; A 0.9951; mse 0.112 cos 0.947
153
+ [10761.4s] block 20: 0.874 -> 0.916 / 0.920; A 0.9920; mse 0.117 cos 0.945
154
+ [10761.4s] block 24: 0.861 -> 0.909 / 0.914; A 0.9928; mse 0.148 cos 0.935
155
+ [10761.4s] block 28: 0.818 -> 0.870 / 0.876; A 0.9893; mse 0.216 cos 0.908
156
+ [10761.4s] block 31: 0.710 -> 0.764 / 0.770; A 0.9819; mse 0.308 cos 0.883
157
+ [10761.4s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.912 / 0.918 / 0.914 / 0.903 / 0.872
158
+ [10761.4s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.764 against the bar .85; B2: the lowest A alignment 0.9819 against .995; fineweb share +0.00050 +- 0.00010 against +.012
159
+ [10761.4s] gate mean 0.433 / widest 0.849; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 128,603 (skipped {'test': 203, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 13.2 GB
160
+ [10761.4s] DONE sarm_bert_mse_solo_o0_norm: 4000 steps in 164.8 min of steps (2.47 s/step); 4 closes; 10761 s in all; card peak 13.9 GB; host RSS 13.2 GB; compiled chain: 5 cache entries (limit 256), 0 limit warnings