Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
c001e66
·
verified ·
1 Parent(s): a147b4f

surface arm t5/mse_solo_o0: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/t5/mse_solo_o0/closes/sarm_t5_mse_solo_o0_close1_s11_t5.safetensors +3 -0
  2. surface/t5/mse_solo_o0/closes/sarm_t5_mse_solo_o0_close2_s11_t5.safetensors +3 -0
  3. surface/t5/mse_solo_o0/closes/sarm_t5_mse_solo_o0_close3_s11_t5.safetensors +3 -0
  4. surface/t5/mse_solo_o0/closes/sarm_t5_mse_solo_o0_close4_s11_t5.safetensors +3 -0
  5. surface/t5/mse_solo_o0/manifest.json +20 -0
  6. surface/t5/mse_solo_o0/readout.md +127 -0
  7. surface/t5/mse_solo_o0/results.json +0 -0
  8. surface/t5/mse_solo_o0/s11_t5.safetensors +3 -0
  9. surface/t5/mse_solo_o0/stats.pt +3 -0
  10. surface/t5/mse_solo_o0/train.log +160 -0
surface/t5/mse_solo_o0/closes/sarm_t5_mse_solo_o0_close1_s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dafde8f75c492c382e7f1ccc1ab23bc4a607502e6f2a9081c4f78be6452b848b
3
+ size 54819296
surface/t5/mse_solo_o0/closes/sarm_t5_mse_solo_o0_close2_s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ebfc42cffa6b8a09604ce1eab726f4647fe482d85c712b25d84bbfe44fbd503
3
+ size 54819296
surface/t5/mse_solo_o0/closes/sarm_t5_mse_solo_o0_close3_s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b62eda3230f88f5a87f84d4509e9e1bd0c80911e9068a7ef735a7044bfea2830
3
+ size 54819296
surface/t5/mse_solo_o0/closes/sarm_t5_mse_solo_o0_close4_s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:26542d345a09a9c04197533546046e118f397b6fce01bf73061a8bf340284adf
3
+ size 54819296
surface/t5/mse_solo_o0/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "sarm_t5_mse_solo_o0",
3
+ "run": "mse_solo_o0",
4
+ "convention": "t5",
5
+ "member": "s11_t5",
6
+ "final_anchor": "sarm_t5_mse_solo_o0_close4_s11_t5.safetensors",
7
+ "closes": [
8
+ "sarm_t5_mse_solo_o0_close1_s11_t5.safetensors",
9
+ "sarm_t5_mse_solo_o0_close2_s11_t5.safetensors",
10
+ "sarm_t5_mse_solo_o0_close3_s11_t5.safetensors",
11
+ "sarm_t5_mse_solo_o0_close4_s11_t5.safetensors"
12
+ ],
13
+ "finished": true,
14
+ "form": "mse",
15
+ "base": "solo",
16
+ "steps": 4000,
17
+ "trunk_step": 245674,
18
+ "shipped": "2026-10-09 07:55 local",
19
+ "note": "every close's anchor ships (mids included); the final close's anchor is the run's s11_t5.safetensors"
20
+ }
surface/t5/mse_solo_o0/readout.md ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # The surface arm's readout: t5-v1_1-xxl's spelling (the sentence-piece 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_t5_mse_solo_o0: 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_t5_mse_solo_o0
12
+
13
+ - B1 the gap: MET over every served block >= 12 (the lowest 0.947 at block 12 on draw 1, 0.940 at block 31 on draw 2; bar 0.85)
14
+ - B2 silence, whitened: NOT MET (the lowest A alignment 0.9917 at block 1 on draw 1, 0.9889 on draw 2; bar 0.995)
15
+ - B2 silence, in bits: MET (web +0.00010 bpb; bar +0.012)
16
+
17
+ ## The closes
18
+
19
+ ### sarm_t5_mse_solo_o0
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 | 0.360 / 0.389 | 0.647 / 0.519 / 0.435 | 1.0000 / 1.0000 | +0.00000 | 0.047 / 0.047 |
24
+ | 1 | 1000 | 0.830 / 0.821 | 0.849 / 0.868 / 0.860 | 0.9905 / 0.9880 | +0.00031 | 0.107 / 0.148 |
25
+ | 2 | 2000 | 0.918 / 0.906 | 0.918 / 0.941 / 0.934 | 0.9878 / 0.9823 | +0.00054 | 0.142 / 0.287 |
26
+ | 3 | 3000 | 0.937 / 0.930 | 0.937 / 0.956 / 0.950 | 0.9896 / 0.9868 | +0.00035 | 0.168 / 0.394 |
27
+ | 4 | 4000 | 0.947 / 0.940 | 0.947 / 0.964 / 0.958 | 0.9917 / 0.9889 | +0.00010 | 0.189 / 0.482 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### sarm_t5_mse_solo_o0, draw1 (512 captions, 5,707 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 | 0.868 | 0.866 | 0.866 | 0.9996 | 0.376 | 0.798 |
36
+ | 4 | 0.797 | 0.922 | 0.921 | 0.9967 | 0.096 | 0.949 |
37
+ | 8 | 0.735 | 0.932 | 0.933 | 0.9925 | 0.082 | 0.958 |
38
+ | 12 | 0.647 | 0.947 | 0.948 | 0.9974 | 0.070 | 0.965 |
39
+ | 16 | 0.572 | 0.957 | 0.958 | 0.9959 | 0.058 | 0.972 |
40
+ | 18 | 0.545 | 0.958 | 0.959 | 0.9967 | 0.057 | 0.972 |
41
+ | 20 | 0.519 | 0.964 | 0.963 | 0.9966 | 0.055 | 0.974 |
42
+ | 22 | 0.495 | 0.964 | 0.965 | 0.9959 | 0.054 | 0.976 |
43
+ | 24 | 0.470 | 0.964 | 0.962 | 0.9931 | 0.051 | 0.977 |
44
+ | 26 | 0.448 | 0.962 | 0.962 | 0.9958 | 0.050 | 0.977 |
45
+ | 28 | 0.435 | 0.958 | 0.957 | 0.9938 | 0.052 | 0.974 |
46
+ | 29 | 0.421 | 0.955 | 0.955 | 0.9961 | 0.053 | 0.973 |
47
+ | 30 | 0.396 | 0.952 | 0.952 | 0.9933 | 0.054 | 0.970 |
48
+ | 31 | 0.360 | 0.949 | 0.949 | 0.9932 | 0.053 | 0.967 |
49
+
50
+ ### sarm_t5_mse_solo_o0, draw2 (512 captions, 5,988 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 | 0.865 | 0.863 | 0.863 | 0.9996 | 0.400 | 0.789 |
55
+ | 4 | 0.803 | 0.920 | 0.921 | 0.9977 | 0.104 | 0.946 |
56
+ | 8 | 0.750 | 0.936 | 0.937 | 0.9957 | 0.088 | 0.955 |
57
+ | 12 | 0.661 | 0.946 | 0.947 | 0.9968 | 0.076 | 0.962 |
58
+ | 16 | 0.586 | 0.952 | 0.954 | 0.9962 | 0.063 | 0.969 |
59
+ | 18 | 0.555 | 0.958 | 0.958 | 0.9967 | 0.062 | 0.970 |
60
+ | 20 | 0.526 | 0.958 | 0.959 | 0.9962 | 0.059 | 0.972 |
61
+ | 22 | 0.508 | 0.963 | 0.963 | 0.9960 | 0.058 | 0.974 |
62
+ | 24 | 0.489 | 0.961 | 0.961 | 0.9968 | 0.056 | 0.975 |
63
+ | 26 | 0.469 | 0.961 | 0.960 | 0.9963 | 0.054 | 0.974 |
64
+ | 28 | 0.457 | 0.958 | 0.958 | 0.9953 | 0.057 | 0.972 |
65
+ | 29 | 0.446 | 0.954 | 0.956 | 0.9955 | 0.059 | 0.970 |
66
+ | 30 | 0.426 | 0.947 | 0.948 | 0.9914 | 0.062 | 0.967 |
67
+ | 31 | 0.389 | 0.940 | 0.942 | 0.9889 | 0.059 | 0.964 |
68
+
69
+ The member's share in bits per byte (all on minus the member masked): web +0.00010
70
+
71
+ ## The training curves
72
+
73
+ ### sarm_t5_mse_solo_o0
74
+
75
+ | step | task loss | mse | cos | InfoNCE | quiet on A | quiet off-domain |
76
+ |---|---|---|---|---|---|---|
77
+ | 1 | 1.5939 | 1.594 | 0.338 | None | 0.00000 | 0.00000 |
78
+ | 50 | 1.5763 | 1.576 | 0.342 | None | 0.00030 | 0.00023 |
79
+ | 250 | 1.1622 | 1.162 | 0.464 | None | 0.00141 | 0.00093 |
80
+ | 500 | 0.7611 | 0.761 | 0.635 | None | 0.00059 | 0.00035 |
81
+ | 1000 | 0.4833 | 0.483 | 0.763 | None | 0.00061 | 0.00041 |
82
+ | 1500 | 0.3160 | 0.316 | 0.843 | None | 0.00078 | 0.00057 |
83
+ | 2000 | 0.2150 | 0.215 | 0.892 | None | 0.00089 | 0.00065 |
84
+ | 2500 | 0.1629 | 0.163 | 0.918 | None | 0.00086 | 0.00055 |
85
+ | 3000 | 0.1362 | 0.136 | 0.932 | None | 0.00098 | 0.00071 |
86
+ | 3500 | 0.0812 | 0.081 | 0.959 | None | 0.00088 | 0.00073 |
87
+ | 4000 | 0.0699 | 0.070 | 0.965 | None | 0.00072 | 0.00051 |
88
+ | 1 | 1.5939 | 1.594 | 0.338 | None | 0.00000 | 0.00000 |
89
+ | 50 | 1.5763 | 1.576 | 0.342 | None | 0.00030 | 0.00023 |
90
+ | 250 | 1.1622 | 1.162 | 0.464 | None | 0.00141 | 0.00093 |
91
+ | 500 | 0.7611 | 0.761 | 0.635 | None | 0.00059 | 0.00035 |
92
+ | 1000 | 0.4833 | 0.483 | 0.763 | None | 0.00061 | 0.00041 |
93
+ | 1500 | 0.3160 | 0.316 | 0.843 | None | 0.00078 | 0.00057 |
94
+ | 2000 | 0.2150 | 0.215 | 0.892 | None | 0.00089 | 0.00065 |
95
+ | 2500 | 0.1629 | 0.163 | 0.918 | None | 0.00086 | 0.00055 |
96
+ | 3000 | 0.1362 | 0.136 | 0.932 | None | 0.00098 | 0.00071 |
97
+ | 3500 | 0.0812 | 0.081 | 0.959 | None | 0.00088 | 0.00073 |
98
+ | 4000 | 0.0699 | 0.070 | 0.965 | None | 0.00072 | 0.00051 |
99
+ | 1 | 1.5939 | 1.594 | 0.338 | None | 0.00000 | 0.00000 |
100
+ | 50 | 1.5763 | 1.576 | 0.342 | None | 0.00030 | 0.00023 |
101
+ | 250 | 1.1622 | 1.162 | 0.464 | None | 0.00141 | 0.00093 |
102
+ | 500 | 0.7611 | 0.761 | 0.635 | None | 0.00059 | 0.00035 |
103
+ | 1000 | 0.4833 | 0.483 | 0.763 | None | 0.00061 | 0.00041 |
104
+ | 1500 | 0.3160 | 0.316 | 0.843 | None | 0.00078 | 0.00057 |
105
+ | 2000 | 0.2150 | 0.215 | 0.892 | None | 0.00089 | 0.00065 |
106
+ | 2500 | 0.1629 | 0.163 | 0.918 | None | 0.00086 | 0.00055 |
107
+ | 3000 | 0.1362 | 0.136 | 0.932 | None | 0.00098 | 0.00071 |
108
+ | 3500 | 0.0812 | 0.081 | 0.959 | None | 0.00088 | 0.00073 |
109
+ | 4000 | 0.0699 | 0.070 | 0.965 | None | 0.00072 | 0.00051 |
110
+ | 1 | 1.5939 | 1.594 | 0.338 | None | 0.00000 | 0.00000 |
111
+ | 50 | 1.5763 | 1.576 | 0.342 | None | 0.00030 | 0.00023 |
112
+ | 250 | 1.1622 | 1.162 | 0.464 | None | 0.00141 | 0.00093 |
113
+ | 500 | 0.7611 | 0.761 | 0.635 | None | 0.00059 | 0.00035 |
114
+ | 1000 | 0.4833 | 0.483 | 0.763 | None | 0.00061 | 0.00041 |
115
+ | 1500 | 0.3160 | 0.316 | 0.843 | None | 0.00078 | 0.00057 |
116
+ | 2000 | 0.2150 | 0.215 | 0.892 | None | 0.00089 | 0.00065 |
117
+ | 2500 | 0.1629 | 0.163 | 0.918 | None | 0.00086 | 0.00055 |
118
+ | 3000 | 0.1362 | 0.136 | 0.932 | None | 0.00098 | 0.00071 |
119
+ | 3500 | 0.0812 | 0.081 | 0.959 | None | 0.00088 | 0.00073 |
120
+ | 4000 | 0.0699 | 0.070 | 0.965 | None | 0.00072 | 0.00051 |
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: t5-v1_1-xxl's spelling (the sentence-piece convention; sha256 170798cc83dec7e6..., vocab 32,100; A = the text written; 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/t5/mse_solo_o0/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/t5/mse_solo_o0/s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:26542d345a09a9c04197533546046e118f397b6fce01bf73061a8bf340284adf
3
+ size 54819296
surface/t5/mse_solo_o0/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:882bcf237f58413b667e4d3ba41414b5786cd4d9f915d6a3580bf7e8994673d7
3
+ size 264309
surface/t5/mse_solo_o0/train.log ADDED
@@ -0,0 +1,160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 1.4s] THE JOB: sarm_t5_mse_solo_o0: the surface arm s11_t5 (sentencepiece spelling by google/t5-v1_1-xxl; A = the text written; 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
+ [ 18.7s] 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
+ [ 19.0s] tokenizer google/t5-v1_1-xxl (sentencepiece spelling; vocab 32100; sha256 170798cc83dec7e6): 'a taco truck' -> ['▁', 'a', '▁tac', 'o', '▁truck']; A 12 bytes, B 19 bytes; sites [(1, 1, 4), (2, 5, 10), (3, 6, 11)]
5
+ [ 19.0s] 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
+ [ 32.1s] 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 3.7 GB
7
+ [ 32.2s] streams: web text; partners none (two web rows a quiet chunk); held-out sets ['web'] x 8 batches
8
+ [ 32.3s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 136), 5,707 sites
9
+ [ 32.4s] read rows draw2: 512 captions (0 unspellable), A rows (512, 151), B rows (512, 204), 5,988 sites
10
+ [arms] attached 's11_t5' for phase 'sentencepiece_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
+ [ 58.0s] 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
+ [ 62.1s] standardization statistics from 12,096 teacher sites on 8 paired batches (a mean and a spread per block and dimension); sigma range at block 20 0.744..1.132; saved /workspace/qarm/out/anchors/sarm_t5_mse_solo_o0_stats.pt
13
+ [ 275.1s] CLOSE 0 (step 0): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
14
+ [ 275.1s] block 8: 0.735 -> 0.735 / 0.735; A 1.0000; mse 1.152 cos 0.489
15
+ [ 275.1s] block 12: 0.647 -> 0.647 / 0.647; A 1.0000; mse 1.212 cos 0.438
16
+ [ 275.1s] block 16: 0.572 -> 0.572 / 0.572; A 1.0000; mse 1.313 cos 0.388
17
+ [ 275.1s] block 18: 0.545 -> 0.545 / 0.545; A 1.0000; mse 1.334 cos 0.369
18
+ [ 275.1s] block 20: 0.519 -> 0.519 / 0.519; A 1.0000; mse 1.457 cos 0.319
19
+ [ 275.1s] block 24: 0.470 -> 0.470 / 0.470; A 1.0000; mse 1.726 cos 0.255
20
+ [ 275.1s] block 28: 0.435 -> 0.435 / 0.435; A 1.0000; mse 1.980 cos 0.179
21
+ [ 275.1s] block 31: 0.360 -> 0.360 / 0.360; A 1.0000; mse 3.369 cos 0.119
22
+ [ 275.1s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.661 / 0.586 / 0.526 / 0.489 / 0.457
23
+ [ 275.1s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.360 against the bar .85; B2: the lowest A alignment 1.0000 against .995; fineweb share +0.00000 +- 0.00000 against +.012
24
+ [ 275.1s] 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 11.8 GB
25
+ [stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
26
+ [ 284.3s] step 1/4000: task 1.5939 (mse 1.594 cos 0.338); quiet A 0.00000 off 0.00000; 1,497 sites; 9.21 s/step; the job: 4.7 min spent, about 614.0 min of steps left; card peak 13.7 GB
27
+ [ 406.0s] step 50/4000: task 1.5763 (mse 1.576 cos 0.342); quiet A 0.00030 off 0.00023; 1,469 sites; 2.62 s/step; the job: 6.8 min spent, about 172.3 min of steps left; card peak 13.8 GB
28
+ [ 529.5s] step 100/4000: task 1.5400 (mse 1.540 cos 0.352); quiet A 0.00044 off 0.00031; 1,553 sites; 2.54 s/step; the job: 8.8 min spent, about 165.4 min of steps left; card peak 13.8 GB
29
+ [ 653.1s] step 150/4000: task 1.4797 (mse 1.480 cos 0.369); quiet A 0.00046 off 0.00034; 1,485 sites; 2.52 s/step; the job: 10.9 min spent, about 161.7 min of steps left; card peak 13.8 GB
30
+ [ 776.4s] step 200/4000: task 1.3525 (mse 1.352 cos 0.407); quiet A 0.00077 off 0.00053; 1,516 sites; 2.51 s/step; the job: 12.9 min spent, about 158.7 min of steps left; card peak 13.8 GB
31
+ [ 899.5s] step 250/4000: task 1.1622 (mse 1.162 cos 0.464); quiet A 0.00141 off 0.00093; 1,498 sites; 2.50 s/step; the job: 15.0 min spent, about 156.1 min of steps left; card peak 13.8 GB
32
+ [ 1022.9s] step 300/4000: task 0.9969 (mse 0.997 cos 0.527); quiet A 0.00158 off 0.00093; 1,601 sites; 2.49 s/step; the job: 17.0 min spent, about 153.7 min of steps left; card peak 13.8 GB
33
+ [ 1147.0s] step 350/4000: task 0.9096 (mse 0.910 cos 0.567); quiet A 0.00103 off 0.00061; 1,493 sites; 2.49 s/step; the job: 19.1 min spent, about 151.5 min of steps left; card peak 13.9 GB
34
+ [transformers] Token indices sequence length is longer than the specified maximum sequence length for this model (523 > 512). Running this sequence through the model will result in indexing errors
35
+ [ 1270.2s] step 400/4000: task 0.8573 (mse 0.857 cos 0.589); quiet A 0.00093 off 0.00053; 1,486 sites; 2.49 s/step; the job: 21.2 min spent, about 149.3 min of steps left; card peak 13.9 GB
36
+ [ 1393.3s] step 450/4000: task 0.8032 (mse 0.803 cos 0.616); quiet A 0.00066 off 0.00039; 1,624 sites; 2.48 s/step; the job: 23.2 min spent, about 147.0 min of steps left; card peak 13.9 GB
37
+ [ 1516.4s] step 500/4000: task 0.7611 (mse 0.761 cos 0.635); quiet A 0.00059 off 0.00035; 1,587 sites; 2.48 s/step; the job: 25.3 min spent, about 144.8 min of steps left; card peak 13.9 GB
38
+ [ 1640.1s] step 550/4000: task 0.7219 (mse 0.722 cos 0.653); quiet A 0.00053 off 0.00039; 1,467 sites; 2.48 s/step; the job: 27.3 min spent, about 142.7 min of steps left; card peak 13.9 GB
39
+ [ 1764.1s] step 600/4000: task 0.6892 (mse 0.689 cos 0.669); quiet A 0.00055 off 0.00037; 1,546 sites; 2.48 s/step; the job: 29.4 min spent, about 140.6 min of steps left; card peak 13.9 GB
40
+ [ 1887.5s] step 650/4000: task 0.6572 (mse 0.657 cos 0.684); quiet A 0.00053 off 0.00035; 1,575 sites; 2.48 s/step; the job: 31.5 min spent, about 138.5 min of steps left; card peak 13.9 GB
41
+ [ 2010.7s] step 700/4000: task 0.6267 (mse 0.627 cos 0.697); quiet A 0.00055 off 0.00037; 1,444 sites; 2.48 s/step; the job: 33.5 min spent, about 136.4 min of steps left; card peak 13.9 GB
42
+ [ 2134.5s] step 750/4000: task 0.5992 (mse 0.599 cos 0.710); quiet A 0.00056 off 0.00037; 1,490 sites; 2.48 s/step; the job: 35.6 min spent, about 134.3 min of steps left; card peak 13.9 GB
43
+ [ 2258.2s] step 800/4000: task 0.5762 (mse 0.576 cos 0.719); quiet A 0.00061 off 0.00045; 1,559 sites; 2.48 s/step; the job: 37.6 min spent, about 132.2 min of steps left; card peak 13.9 GB
44
+ [ 2381.3s] step 850/4000: task 0.5529 (mse 0.553 cos 0.730); quiet A 0.00061 off 0.00038; 1,483 sites; 2.48 s/step; the job: 39.7 min spent, about 130.1 min of steps left; card peak 13.9 GB
45
+ [ 2504.7s] step 900/4000: task 0.5232 (mse 0.523 cos 0.745); quiet A 0.00063 off 0.00041; 1,662 sites; 2.48 s/step; the job: 41.7 min spent, about 128.0 min of steps left; card peak 13.9 GB
46
+ [ 2628.2s] step 950/4000: task 0.5019 (mse 0.502 cos 0.754); quiet A 0.00066 off 0.00042; 1,545 sites; 2.48 s/step; the job: 43.8 min spent, about 125.9 min of steps left; card peak 13.9 GB
47
+ [ 2751.6s] step 1000/4000: task 0.4833 (mse 0.483 cos 0.763); quiet A 0.00061 off 0.00041; 1,568 sites; 2.48 s/step; the job: 45.9 min spent, about 123.8 min of steps left; card peak 13.9 GB
48
+ [ 2751.8s] anchor saved: /workspace/qarm/out/anchors/sarm_t5_mse_solo_o0_close1_s11_t5.safetensors
49
+ [ 2914.1s] CLOSE 1 (step 1000): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
50
+ [ 2914.1s] block 8: 0.735 -> 0.843 / 0.843; A 0.9929; mse 0.384 cos 0.803
51
+ [ 2914.1s] block 12: 0.647 -> 0.849 / 0.851; A 0.9959; mse 0.321 cos 0.836
52
+ [ 2914.1s] block 16: 0.572 -> 0.858 / 0.859; A 0.9970; mse 0.272 cos 0.863
53
+ [ 2914.1s] block 18: 0.545 -> 0.862 / 0.864; A 0.9967; mse 0.256 cos 0.873
54
+ [ 2914.1s] block 20: 0.519 -> 0.868 / 0.868; A 0.9965; mse 0.241 cos 0.883
55
+ [ 2914.1s] block 24: 0.470 -> 0.877 / 0.878; A 0.9966; mse 0.217 cos 0.899
56
+ [ 2914.1s] block 28: 0.435 -> 0.860 / 0.861; A 0.9951; mse 0.222 cos 0.888
57
+ [ 2914.1s] block 31: 0.360 -> 0.830 / 0.830; A 0.9930; mse 0.210 cos 0.867
58
+ [ 2914.1s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.850 / 0.853 / 0.866 / 0.873 / 0.859
59
+ [ 2914.1s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.830 against the bar .85; B2: the lowest A alignment 0.9905 against .995; fineweb share +0.00031 +- 0.00013 against +.012
60
+ [ 2914.1s] gate mean 0.107 / widest 0.148; 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 15.1 GB
61
+ [ 3037.4s] step 1050/4000: task 0.4606 (mse 0.461 cos 0.773); quiet A 0.00062 off 0.00046; 1,516 sites; 2.47 s/step; the job: 50.6 min spent, about 121.7 min of steps left; card peak 13.9 GB
62
+ [ 3160.9s] step 1100/4000: task 0.4381 (mse 0.438 cos 0.783); quiet A 0.00065 off 0.00043; 1,524 sites; 2.47 s/step; the job: 52.7 min spent, about 119.7 min of steps left; card peak 13.9 GB
63
+ [ 3284.3s] step 1150/4000: task 0.4208 (mse 0.421 cos 0.791); quiet A 0.00069 off 0.00045; 1,503 sites; 2.47 s/step; the job: 54.7 min spent, about 117.6 min of steps left; card peak 13.9 GB
64
+ [ 3407.6s] step 1200/4000: task 0.4006 (mse 0.401 cos 0.801); quiet A 0.00070 off 0.00050; 1,497 sites; 2.47 s/step; the job: 56.8 min spent, about 115.5 min of steps left; card peak 13.9 GB
65
+ [ 3531.0s] step 1250/4000: task 0.3839 (mse 0.384 cos 0.810); quiet A 0.00069 off 0.00052; 1,517 sites; 2.47 s/step; the job: 58.9 min spent, about 113.4 min of steps left; card peak 13.9 GB
66
+ [ 3654.4s] step 1300/4000: task 0.3688 (mse 0.369 cos 0.816); quiet A 0.00072 off 0.00051; 1,494 sites; 2.47 s/step; the job: 60.9 min spent, about 111.3 min of steps left; card peak 13.9 GB
67
+ [ 3777.6s] step 1350/4000: task 0.3542 (mse 0.354 cos 0.824); quiet A 0.00071 off 0.00049; 1,511 sites; 2.47 s/step; the job: 63.0 min spent, about 109.3 min of steps left; card peak 13.9 GB
68
+ [ 3901.4s] step 1400/4000: task 0.3411 (mse 0.341 cos 0.830); quiet A 0.00087 off 0.00055; 1,521 sites; 2.47 s/step; the job: 65.0 min spent, about 107.2 min of steps left; card peak 13.9 GB
69
+ [ 4025.1s] step 1450/4000: task 0.3281 (mse 0.328 cos 0.837); quiet A 0.00078 off 0.00054; 1,547 sites; 2.47 s/step; the job: 67.1 min spent, about 105.1 min of steps left; card peak 13.9 GB
70
+ [ 4148.1s] step 1500/4000: task 0.3160 (mse 0.316 cos 0.843); quiet A 0.00078 off 0.00057; 1,469 sites; 2.47 s/step; the job: 69.1 min spent, about 103.1 min of steps left; card peak 13.9 GB
71
+ [ 4271.2s] step 1550/4000: task 0.3028 (mse 0.303 cos 0.849); quiet A 0.00080 off 0.00054; 1,472 sites; 2.47 s/step; the job: 71.2 min spent, about 101.0 min of steps left; card peak 13.9 GB
72
+ [ 4394.9s] step 1600/4000: task 0.2906 (mse 0.291 cos 0.855); quiet A 0.00081 off 0.00051; 1,488 sites; 2.47 s/step; the job: 73.2 min spent, about 98.9 min of steps left; card peak 13.9 GB
73
+ [ 4518.3s] step 1650/4000: task 0.2794 (mse 0.279 cos 0.860); quiet A 0.00083 off 0.00060; 1,524 sites; 2.47 s/step; the job: 75.3 min spent, about 96.9 min of steps left; card peak 13.9 GB
74
+ [ 4641.7s] step 1700/4000: task 0.2695 (mse 0.270 cos 0.865); quiet A 0.00088 off 0.00065; 1,498 sites; 2.47 s/step; the job: 77.4 min spent, about 94.8 min of steps left; card peak 13.9 GB
75
+ [ 4765.1s] step 1750/4000: task 0.2587 (mse 0.259 cos 0.871); quiet A 0.00089 off 0.00066; 1,473 sites; 2.47 s/step; the job: 79.4 min spent, about 92.7 min of steps left; card peak 13.9 GB
76
+ [ 4888.2s] step 1800/4000: task 0.2497 (mse 0.250 cos 0.875); quiet A 0.00087 off 0.00055; 1,564 sites; 2.47 s/step; the job: 81.5 min spent, about 90.7 min of steps left; card peak 13.9 GB
77
+ [ 5011.6s] step 1850/4000: task 0.2396 (mse 0.240 cos 0.880); quiet A 0.00089 off 0.00064; 1,569 sites; 2.47 s/step; the job: 83.5 min spent, about 88.6 min of steps left; card peak 13.9 GB
78
+ [ 5134.7s] step 1900/4000: task 0.2308 (mse 0.231 cos 0.885); quiet A 0.00089 off 0.00068; 1,623 sites; 2.47 s/step; the job: 85.6 min spent, about 86.5 min of steps left; card peak 13.9 GB
79
+ [ 5258.3s] step 1950/4000: task 0.2229 (mse 0.223 cos 0.888); quiet A 0.00090 off 0.00058; 1,525 sites; 2.47 s/step; the job: 87.6 min spent, about 84.5 min of steps left; card peak 13.9 GB
80
+ [ 5381.5s] step 2000/4000: task 0.2150 (mse 0.215 cos 0.892); quiet A 0.00089 off 0.00065; 1,509 sites; 2.47 s/step; the job: 89.7 min spent, about 82.4 min of steps left; card peak 13.9 GB
81
+ [ 5381.8s] anchor saved: /workspace/qarm/out/anchors/sarm_t5_mse_solo_o0_close2_s11_t5.safetensors
82
+ [ 5544.1s] CLOSE 2 (step 2000): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
83
+ [ 5544.1s] block 8: 0.735 -> 0.906 / 0.905; A 0.9915; mse 0.163 cos 0.915
84
+ [ 5544.1s] block 12: 0.647 -> 0.918 / 0.920; A 0.9904; mse 0.132 cos 0.933
85
+ [ 5544.1s] block 16: 0.572 -> 0.933 / 0.934; A 0.9933; mse 0.106 cos 0.948
86
+ [ 5544.1s] block 18: 0.545 -> 0.934 / 0.933; A 0.9945; mse 0.102 cos 0.951
87
+ [ 5544.1s] block 20: 0.519 -> 0.941 / 0.940; A 0.9940; mse 0.096 cos 0.954
88
+ [ 5544.1s] block 24: 0.470 -> 0.943 / 0.941; A 0.9935; mse 0.087 cos 0.960
89
+ [ 5544.1s] block 28: 0.435 -> 0.934 / 0.934; A 0.9939; mse 0.088 cos 0.957
90
+ [ 5544.1s] block 31: 0.360 -> 0.919 / 0.920; A 0.9894; mse 0.087 cos 0.946
91
+ [ 5544.1s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.920 / 0.928 / 0.936 / 0.938 / 0.933
92
+ [ 5544.1s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.918 against the bar .85; B2: the lowest A alignment 0.9878 against .995; fineweb share +0.00054 +- 0.00016 against +.012
93
+ [ 5544.1s] gate mean 0.142 / widest 0.287; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 64,564 (skipped {'test': 97, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 15.1 GB
94
+ [ 5667.6s] step 2050/4000: task 0.2059 (mse 0.206 cos 0.897); quiet A 0.00090 off 0.00068; 1,480 sites; 2.47 s/step; the job: 94.5 min spent, about 80.3 min of steps left; card peak 13.9 GB
95
+ [ 5791.0s] step 2100/4000: task 0.2003 (mse 0.200 cos 0.900); quiet A 0.00100 off 0.00085; 1,513 sites; 2.47 s/step; the job: 96.5 min spent, about 78.3 min of steps left; card peak 13.9 GB
96
+ [ 5913.9s] step 2150/4000: task 0.1933 (mse 0.193 cos 0.903); quiet A 0.00093 off 0.00068; 1,493 sites; 2.47 s/step; the job: 98.6 min spent, about 76.2 min of steps left; card peak 13.9 GB
97
+ [ 6037.0s] step 2200/4000: task 0.1883 (mse 0.188 cos 0.906); quiet A 0.00094 off 0.00061; 1,542 sites; 2.46 s/step; the job: 100.6 min spent, about 74.1 min of steps left; card peak 13.9 GB
98
+ [ 6160.3s] step 2250/4000: task 0.1843 (mse 0.184 cos 0.908); quiet A 0.00093 off 0.00064; 1,565 sites; 2.46 s/step; the job: 102.7 min spent, about 72.1 min of steps left; card peak 13.9 GB
99
+ [ 6283.6s] step 2300/4000: task 0.1791 (mse 0.179 cos 0.910); quiet A 0.00088 off 0.00058; 1,522 sites; 2.47 s/step; the job: 104.7 min spent, about 70.0 min of steps left; card peak 13.9 GB
100
+ [ 6407.5s] step 2350/4000: task 0.1775 (mse 0.178 cos 0.911); quiet A 0.00091 off 0.00059; 1,540 sites; 2.47 s/step; the job: 106.8 min spent, about 68.0 min of steps left; card peak 13.9 GB
101
+ [ 6530.8s] step 2400/4000: task 0.1709 (mse 0.171 cos 0.914); quiet A 0.00088 off 0.00055; 1,516 sites; 2.47 s/step; the job: 108.8 min spent, about 65.9 min of steps left; card peak 13.9 GB
102
+ [ 6654.5s] step 2450/4000: task 0.1670 (mse 0.167 cos 0.916); quiet A 0.00089 off 0.00067; 1,472 sites; 2.47 s/step; the job: 110.9 min spent, about 63.8 min of steps left; card peak 13.9 GB
103
+ [ 6777.7s] step 2500/4000: task 0.1629 (mse 0.163 cos 0.918); quiet A 0.00086 off 0.00055; 1,566 sites; 2.47 s/step; the job: 113.0 min spent, about 61.8 min of steps left; card peak 13.9 GB
104
+ [ 6900.8s] step 2550/4000: task 0.1578 (mse 0.158 cos 0.921); quiet A 0.00087 off 0.00058; 1,562 sites; 2.47 s/step; the job: 115.0 min spent, about 59.7 min of steps left; card peak 13.9 GB
105
+ [ 7023.9s] step 2600/4000: task 0.1606 (mse 0.161 cos 0.920); quiet A 0.00092 off 0.00064; 1,480 sites; 2.47 s/step; the job: 117.1 min spent, about 57.6 min of steps left; card peak 13.9 GB
106
+ [ 7147.3s] step 2650/4000: task 0.1543 (mse 0.154 cos 0.923); quiet A 0.00088 off 0.00077; 1,464 sites; 2.47 s/step; the job: 119.1 min spent, about 55.6 min of steps left; card peak 13.9 GB
107
+ [ 7270.4s] step 2700/4000: task 0.1513 (mse 0.151 cos 0.924); quiet A 0.00091 off 0.00063; 1,575 sites; 2.47 s/step; the job: 121.2 min spent, about 53.5 min of steps left; card peak 13.9 GB
108
+ [ 7393.6s] step 2750/4000: task 0.1482 (mse 0.148 cos 0.926); quiet A 0.00091 off 0.00069; 1,507 sites; 2.47 s/step; the job: 123.2 min spent, about 51.5 min of steps left; card peak 13.9 GB
109
+ [ 7516.9s] step 2800/4000: task 0.1484 (mse 0.148 cos 0.926); quiet A 0.00088 off 0.00057; 1,498 sites; 2.47 s/step; the job: 125.3 min spent, about 49.4 min of steps left; card peak 13.9 GB
110
+ [ 7641.0s] step 2850/4000: task 0.1425 (mse 0.143 cos 0.928); quiet A 0.00084 off 0.00055; 1,474 sites; 2.47 s/step; the job: 127.4 min spent, about 47.4 min of steps left; card peak 13.9 GB
111
+ [ 7764.2s] step 2900/4000: task 0.1523 (mse 0.152 cos 0.926); quiet A 0.00134 off 0.00058; 1,539 sites; 2.47 s/step; the job: 129.4 min spent, about 45.3 min of steps left; card peak 13.9 GB
112
+ [ 7887.2s] step 2950/4000: task 0.1583 (mse 0.158 cos 0.920); quiet A 0.00127 off 0.00087; 1,539 sites; 2.47 s/step; the job: 131.5 min spent, about 43.2 min of steps left; card peak 13.9 GB
113
+ [ 8010.4s] step 3000/4000: task 0.1362 (mse 0.136 cos 0.932); quiet A 0.00098 off 0.00071; 1,475 sites; 2.47 s/step; the job: 133.5 min spent, about 41.2 min of steps left; card peak 13.9 GB
114
+ [ 8010.6s] anchor saved: /workspace/qarm/out/anchors/sarm_t5_mse_solo_o0_close3_s11_t5.safetensors
115
+ [ 8173.0s] CLOSE 3 (step 3000): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
116
+ [ 8173.0s] block 8: 0.735 -> 0.923 / 0.923; A 0.9937; mse 0.106 cos 0.945
117
+ [ 8173.0s] block 12: 0.647 -> 0.937 / 0.938; A 0.9957; mse 0.089 cos 0.955
118
+ [ 8173.0s] block 16: 0.572 -> 0.949 / 0.949; A 0.9961; mse 0.073 cos 0.964
119
+ [ 8173.0s] block 18: 0.545 -> 0.948 / 0.948; A 0.9960; mse 0.071 cos 0.965
120
+ [ 8173.0s] block 20: 0.519 -> 0.956 / 0.955; A 0.9954; mse 0.068 cos 0.968
121
+ [ 8173.0s] block 24: 0.470 -> 0.956 / 0.955; A 0.9956; mse 0.062 cos 0.971
122
+ [ 8173.0s] block 28: 0.435 -> 0.950 / 0.951; A 0.9929; mse 0.063 cos 0.969
123
+ [ 8173.0s] block 31: 0.360 -> 0.937 / 0.938; A 0.9921; mse 0.064 cos 0.960
124
+ [ 8173.0s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.936 / 0.945 / 0.951 / 0.953 / 0.949
125
+ [ 8173.0s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.937 against the bar .85; B2: the lowest A alignment 0.9896 against .995; fineweb share +0.00035 +- 0.00021 against +.012
126
+ [ 8173.0s] gate mean 0.168 / widest 0.394; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 96,501 (skipped {'test': 156, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 15.2 GB
127
+ [ 8296.1s] step 3050/4000: task 0.1302 (mse 0.130 cos 0.934); quiet A 0.00092 off 0.00062; 1,509 sites; 2.46 s/step; the job: 138.3 min spent, about 39.1 min of steps left; card peak 13.9 GB
128
+ [ 8419.5s] step 3100/4000: task 0.1267 (mse 0.127 cos 0.937); quiet A 0.00092 off 0.00077; 1,579 sites; 2.46 s/step; the job: 140.3 min spent, about 37.0 min of steps left; card peak 13.9 GB
129
+ [ 8542.8s] step 3150/4000: task 0.1211 (mse 0.121 cos 0.940); quiet A 0.00093 off 0.00057; 1,509 sites; 2.47 s/step; the job: 142.4 min spent, about 35.0 min of steps left; card peak 13.9 GB
130
+ [ 8666.7s] step 3200/4000: task 0.1128 (mse 0.113 cos 0.944); quiet A 0.00090 off 0.00064; 1,560 sites; 2.47 s/step; the job: 144.4 min spent, about 32.9 min of steps left; card peak 13.9 GB
131
+ [ 8790.1s] step 3250/4000: task 0.1051 (mse 0.105 cos 0.947); quiet A 0.00088 off 0.00062; 1,518 sites; 2.47 s/step; the job: 146.5 min spent, about 30.9 min of steps left; card peak 13.9 GB
132
+ [ 8913.7s] step 3300/4000: task 0.0994 (mse 0.099 cos 0.950); quiet A 0.00093 off 0.00063; 1,511 sites; 2.47 s/step; the job: 148.6 min spent, about 28.8 min of steps left; card peak 13.9 GB
133
+ [ 9036.9s] step 3350/4000: task 0.0971 (mse 0.097 cos 0.952); quiet A 0.00090 off 0.00056; 1,594 sites; 2.47 s/step; the job: 150.6 min spent, about 26.8 min of steps left; card peak 13.9 GB
134
+ [ 9160.3s] step 3400/4000: task 0.0881 (mse 0.088 cos 0.956); quiet A 0.00087 off 0.00063; 1,592 sites; 2.47 s/step; the job: 152.7 min spent, about 24.7 min of steps left; card peak 13.9 GB
135
+ [ 9284.4s] step 3450/4000: task 0.0847 (mse 0.085 cos 0.958); quiet A 0.00086 off 0.00064; 1,464 sites; 2.47 s/step; the job: 154.7 min spent, about 22.6 min of steps left; card peak 13.9 GB
136
+ [ 9407.9s] step 3500/4000: task 0.0812 (mse 0.081 cos 0.959); quiet A 0.00088 off 0.00073; 1,539 sites; 2.47 s/step; the job: 156.8 min spent, about 20.6 min of steps left; card peak 13.9 GB
137
+ [ 9531.0s] step 3550/4000: task 0.0810 (mse 0.081 cos 0.959); quiet A 0.00090 off 0.00064; 1,530 sites; 2.47 s/step; the job: 158.8 min spent, about 18.5 min of steps left; card peak 13.9 GB
138
+ [ 9654.1s] step 3600/4000: task 0.0773 (mse 0.077 cos 0.961); quiet A 0.00082 off 0.00057; 1,585 sites; 2.47 s/step; the job: 160.9 min spent, about 16.5 min of steps left; card peak 13.9 GB
139
+ [ 9777.2s] step 3650/4000: task 0.0766 (mse 0.077 cos 0.962); quiet A 0.00081 off 0.00068; 1,463 sites; 2.47 s/step; the job: 163.0 min spent, about 14.4 min of steps left; card peak 13.9 GB
140
+ [ 9900.3s] step 3700/4000: task 0.0739 (mse 0.074 cos 0.963); quiet A 0.00079 off 0.00057; 1,516 sites; 2.47 s/step; the job: 165.0 min spent, about 12.3 min of steps left; card peak 13.9 GB
141
+ [10023.6s] step 3750/4000: task 0.0719 (mse 0.072 cos 0.964); quiet A 0.00080 off 0.00060; 1,567 sites; 2.47 s/step; the job: 167.1 min spent, about 10.3 min of steps left; card peak 13.9 GB
142
+ [10146.7s] step 3800/4000: task 0.0718 (mse 0.072 cos 0.964); quiet A 0.00076 off 0.00050; 1,506 sites; 2.47 s/step; the job: 169.1 min spent, about 8.2 min of steps left; card peak 13.9 GB
143
+ [10270.2s] step 3850/4000: task 0.0707 (mse 0.071 cos 0.965); quiet A 0.00076 off 0.00052; 1,498 sites; 2.47 s/step; the job: 171.2 min spent, about 6.2 min of steps left; card peak 13.9 GB
144
+ [10393.5s] step 3900/4000: task 0.0692 (mse 0.069 cos 0.965); quiet A 0.00074 off 0.00051; 1,433 sites; 2.47 s/step; the job: 173.2 min spent, about 4.1 min of steps left; card peak 13.9 GB
145
+ [10516.6s] step 3950/4000: task 0.0687 (mse 0.069 cos 0.965); quiet A 0.00071 off 0.00050; 1,504 sites; 2.47 s/step; the job: 175.3 min spent, about 2.1 min of steps left; card peak 13.9 GB
146
+ [10640.4s] step 4000/4000: task 0.0699 (mse 0.070 cos 0.965); quiet A 0.00072 off 0.00051; 1,516 sites; 2.47 s/step; the job: 177.3 min spent, about 0.0 min of steps left; card peak 13.9 GB
147
+ [10640.7s] anchor saved: /workspace/qarm/out/anchors/sarm_t5_mse_solo_o0_close4_s11_t5.safetensors
148
+ [10803.2s] CLOSE 4 (step 4000): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
149
+ [10803.2s] block 8: 0.735 -> 0.932 / 0.933; A 0.9925; mse 0.082 cos 0.958
150
+ [10803.2s] block 12: 0.647 -> 0.947 / 0.948; A 0.9974; mse 0.070 cos 0.965
151
+ [10803.2s] block 16: 0.572 -> 0.957 / 0.958; A 0.9959; mse 0.058 cos 0.972
152
+ [10803.2s] block 18: 0.545 -> 0.958 / 0.959; A 0.9967; mse 0.057 cos 0.972
153
+ [10803.2s] block 20: 0.519 -> 0.964 / 0.963; A 0.9966; mse 0.055 cos 0.974
154
+ [10803.2s] block 24: 0.470 -> 0.964 / 0.962; A 0.9931; mse 0.051 cos 0.977
155
+ [10803.2s] block 28: 0.435 -> 0.958 / 0.957; A 0.9938; mse 0.052 cos 0.974
156
+ [10803.2s] block 31: 0.360 -> 0.949 / 0.949; A 0.9932; mse 0.053 cos 0.967
157
+ [10803.2s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.946 / 0.952 / 0.958 / 0.961 / 0.958
158
+ [10803.2s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.947 against the bar .85; B2: the lowest A alignment 0.9917 against .995; fineweb share +0.00010 +- 0.00015 against +.012
159
+ [10803.2s] gate mean 0.189 / widest 0.482; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 128,687 (skipped {'test': 203, 'spelling': 2, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 15.2 GB
160
+ [10803.2s] DONE sarm_t5_mse_solo_o0: 4000 steps in 164.6 min of steps (2.47 s/step); 4 closes; 10803 s in all; card peak 13.9 GB; host RSS 15.2 GB; compiled chain: 5 cache entries (limit 256), 0 limit warnings