Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
1ef29cc
·
verified ·
1 Parent(s): 458db9c

surface arm qwen3/mse_solo_o0: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/qwen3/mse_solo_o0/closes/qarm_mse_solo_o0_close1_s10_qwen.safetensors +3 -0
  2. surface/qwen3/mse_solo_o0/closes/qarm_mse_solo_o0_close2_s10_qwen.safetensors +3 -0
  3. surface/qwen3/mse_solo_o0/closes/qarm_mse_solo_o0_close3_s10_qwen.safetensors +3 -0
  4. surface/qwen3/mse_solo_o0/closes/qarm_mse_solo_o0_close4_s10_qwen.safetensors +3 -0
  5. surface/qwen3/mse_solo_o0/manifest.json +20 -0
  6. surface/qwen3/mse_solo_o0/readout.md +127 -0
  7. surface/qwen3/mse_solo_o0/results.json +0 -0
  8. surface/qwen3/mse_solo_o0/s10_qwen.safetensors +3 -0
  9. surface/qwen3/mse_solo_o0/stats.pt +3 -0
  10. surface/qwen3/mse_solo_o0/train.log +159 -0
surface/qwen3/mse_solo_o0/closes/qarm_mse_solo_o0_close1_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:43212d2ff46b412f110565ed9f83b2fc80e5d035243be5bf502779cc2a4ca4cc
3
+ size 54819240
surface/qwen3/mse_solo_o0/closes/qarm_mse_solo_o0_close2_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2cd04b3f06df148c4e49bddcc39cf5887f84ef1a505bb9a5f5ed6b68324079a3
3
+ size 54819240
surface/qwen3/mse_solo_o0/closes/qarm_mse_solo_o0_close3_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d23a692de7f8a17a33eb2436a5a9a12e3cd59bb9d53b9d9d9b7b7d8af42639ed
3
+ size 54819240
surface/qwen3/mse_solo_o0/closes/qarm_mse_solo_o0_close4_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e72b3a181d3e802dbfa375490295def22976f5c32bc3f91065acd3d0bfc64ef3
3
+ size 54819240
surface/qwen3/mse_solo_o0/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "qarm_mse_solo_o0",
3
+ "run": "mse_solo_o0",
4
+ "convention": "qwen3",
5
+ "member": "s10_qwen",
6
+ "final_anchor": "qarm_mse_solo_o0_close4_s10_qwen.safetensors",
7
+ "closes": [
8
+ "qarm_mse_solo_o0_close1_s10_qwen.safetensors",
9
+ "qarm_mse_solo_o0_close2_s10_qwen.safetensors",
10
+ "qarm_mse_solo_o0_close3_s10_qwen.safetensors",
11
+ "qarm_mse_solo_o0_close4_s10_qwen.safetensors"
12
+ ],
13
+ "finished": true,
14
+ "form": "mse",
15
+ "base": "solo",
16
+ "steps": 4000,
17
+ "trunk_step": 245674,
18
+ "shipped": "2026-10-08 11:56 local",
19
+ "note": "every close's anchor ships (mids included); the final close's anchor is the run's s10_qwen.safetensors"
20
+ }
surface/qwen3/mse_solo_o0/readout.md ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # The surface arm's readout: Qwen3's spelling (the GPT-2 byte map)
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
+ - qarm_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
+ ### qarm_mse_solo_o0
12
+
13
+ - B1 the gap: MET over every served block >= 12 (the lowest 0.961 at block 31 on draw 1, 0.956 at block 31 on draw 2; bar 0.85)
14
+ - B2 silence, whitened: NOT MET (the lowest A alignment 0.9886 at block 1 on draw 1, 0.9880 on draw 2; bar 0.995)
15
+ - B2 silence, in bits: MET (web +0.00031 bpb; bar +0.012)
16
+
17
+ ## The closes
18
+
19
+ ### qarm_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 | 0.350 / 0.359 | 0.635 / 0.475 / 0.399 | 1.0000 / 1.0000 | +0.00000 | 0.047 / 0.047 |
24
+ | 1 | 1000 | 0.864 / 0.855 | 0.873 / 0.893 / 0.887 | 0.9781 / 0.9770 | +0.00012 | 0.116 / 0.167 |
25
+ | 2 | 2000 | 0.935 / 0.927 | 0.936 / 0.951 / 0.948 | 0.9886 / 0.9862 | +0.00001 | 0.135 / 0.279 |
26
+ | 3 | 3000 | 0.953 / 0.947 | 0.953 / 0.965 / 0.963 | 0.9891 / 0.9860 | +0.00025 | 0.150 / 0.390 |
27
+ | 4 | 4000 | 0.961 / 0.956 | 0.963 / 0.970 / 0.973 | 0.9886 / 0.9880 | +0.00031 | 0.162 / 0.472 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### qarm_mse_solo_o0, draw1 (512 captions, 4,920 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 | 0.899 | 0.897 | 0.898 | 0.9996 | 0.348 | 0.781 |
36
+ | 4 | 0.785 | 0.948 | 0.948 | 0.9976 | 0.091 | 0.950 |
37
+ | 8 | 0.729 | 0.953 | 0.954 | 0.9963 | 0.067 | 0.965 |
38
+ | 12 | 0.635 | 0.963 | 0.963 | 0.9972 | 0.054 | 0.974 |
39
+ | 16 | 0.524 | 0.968 | 0.968 | 0.9957 | 0.041 | 0.980 |
40
+ | 18 | 0.493 | 0.971 | 0.972 | 0.9973 | 0.039 | 0.982 |
41
+ | 20 | 0.475 | 0.970 | 0.971 | 0.9953 | 0.036 | 0.983 |
42
+ | 22 | 0.463 | 0.975 | 0.974 | 0.9959 | 0.034 | 0.985 |
43
+ | 24 | 0.446 | 0.976 | 0.976 | 0.9974 | 0.033 | 0.985 |
44
+ | 26 | 0.421 | 0.972 | 0.973 | 0.9965 | 0.033 | 0.985 |
45
+ | 28 | 0.399 | 0.973 | 0.972 | 0.9970 | 0.035 | 0.983 |
46
+ | 29 | 0.391 | 0.970 | 0.967 | 0.9945 | 0.035 | 0.982 |
47
+ | 30 | 0.372 | 0.968 | 0.966 | 0.9952 | 0.036 | 0.981 |
48
+ | 31 | 0.350 | 0.961 | 0.961 | 0.9935 | 0.038 | 0.977 |
49
+
50
+ ### qarm_mse_solo_o0, draw2 (512 captions, 5,086 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 | 0.890 | 0.886 | 0.888 | 0.9958 | 0.359 | 0.775 |
55
+ | 4 | 0.793 | 0.946 | 0.946 | 0.9968 | 0.093 | 0.949 |
56
+ | 8 | 0.738 | 0.953 | 0.953 | 0.9972 | 0.068 | 0.965 |
57
+ | 12 | 0.639 | 0.958 | 0.958 | 0.9953 | 0.055 | 0.973 |
58
+ | 16 | 0.529 | 0.966 | 0.965 | 0.9950 | 0.042 | 0.979 |
59
+ | 18 | 0.500 | 0.970 | 0.970 | 0.9955 | 0.039 | 0.981 |
60
+ | 20 | 0.483 | 0.970 | 0.969 | 0.9940 | 0.037 | 0.982 |
61
+ | 22 | 0.471 | 0.972 | 0.972 | 0.9971 | 0.035 | 0.984 |
62
+ | 24 | 0.453 | 0.975 | 0.974 | 0.9960 | 0.034 | 0.984 |
63
+ | 26 | 0.429 | 0.974 | 0.973 | 0.9967 | 0.035 | 0.984 |
64
+ | 28 | 0.410 | 0.968 | 0.968 | 0.9960 | 0.037 | 0.982 |
65
+ | 29 | 0.403 | 0.969 | 0.969 | 0.9940 | 0.038 | 0.981 |
66
+ | 30 | 0.382 | 0.961 | 0.959 | 0.9949 | 0.039 | 0.979 |
67
+ | 31 | 0.359 | 0.956 | 0.956 | 0.9900 | 0.041 | 0.976 |
68
+
69
+ The member's share in bits per byte (all on minus the member masked): web +0.00031
70
+
71
+ ## The training curves
72
+
73
+ ### qarm_mse_solo_o0
74
+
75
+ | step | task loss | mse | cos | InfoNCE | quiet on A | quiet off-domain |
76
+ |---|---|---|---|---|---|---|
77
+ | 1 | 1.9096 | 1.910 | 0.265 | None | 0.00000 | 0.00000 |
78
+ | 50 | 1.9126 | 1.913 | 0.259 | None | 0.00031 | 0.00023 |
79
+ | 250 | 1.7418 | 1.742 | 0.297 | None | 0.00166 | 0.00058 |
80
+ | 500 | 0.8752 | 0.875 | 0.604 | None | 0.00111 | 0.00050 |
81
+ | 1000 | 0.4800 | 0.480 | 0.771 | None | 0.00087 | 0.00045 |
82
+ | 1500 | 0.2948 | 0.295 | 0.856 | None | 0.00086 | 0.00058 |
83
+ | 2000 | 0.1895 | 0.189 | 0.906 | None | 0.00092 | 0.00060 |
84
+ | 2500 | 0.1274 | 0.127 | 0.936 | None | 0.00092 | 0.00060 |
85
+ | 3000 | 0.0983 | 0.098 | 0.950 | None | 0.00089 | 0.00063 |
86
+ | 3500 | 0.0769 | 0.077 | 0.960 | None | 0.00079 | 0.00250 |
87
+ | 4000 | 0.0683 | 0.068 | 0.965 | None | 0.00075 | 0.00048 |
88
+ | 1 | 1.9096 | 1.910 | 0.265 | None | 0.00000 | 0.00000 |
89
+ | 50 | 1.9126 | 1.913 | 0.259 | None | 0.00031 | 0.00023 |
90
+ | 250 | 1.7418 | 1.742 | 0.297 | None | 0.00166 | 0.00058 |
91
+ | 500 | 0.8752 | 0.875 | 0.604 | None | 0.00111 | 0.00050 |
92
+ | 1000 | 0.4800 | 0.480 | 0.771 | None | 0.00087 | 0.00045 |
93
+ | 1500 | 0.2948 | 0.295 | 0.856 | None | 0.00086 | 0.00058 |
94
+ | 2000 | 0.1895 | 0.189 | 0.906 | None | 0.00092 | 0.00060 |
95
+ | 2500 | 0.1274 | 0.127 | 0.936 | None | 0.00092 | 0.00060 |
96
+ | 3000 | 0.0983 | 0.098 | 0.950 | None | 0.00089 | 0.00063 |
97
+ | 3500 | 0.0769 | 0.077 | 0.960 | None | 0.00079 | 0.00250 |
98
+ | 4000 | 0.0683 | 0.068 | 0.965 | None | 0.00075 | 0.00048 |
99
+ | 1 | 1.9096 | 1.910 | 0.265 | None | 0.00000 | 0.00000 |
100
+ | 50 | 1.9126 | 1.913 | 0.259 | None | 0.00031 | 0.00023 |
101
+ | 250 | 1.7418 | 1.742 | 0.297 | None | 0.00166 | 0.00058 |
102
+ | 500 | 0.8752 | 0.875 | 0.604 | None | 0.00111 | 0.00050 |
103
+ | 1000 | 0.4800 | 0.480 | 0.771 | None | 0.00087 | 0.00045 |
104
+ | 1500 | 0.2948 | 0.295 | 0.856 | None | 0.00086 | 0.00058 |
105
+ | 2000 | 0.1895 | 0.189 | 0.906 | None | 0.00092 | 0.00060 |
106
+ | 2500 | 0.1274 | 0.127 | 0.936 | None | 0.00092 | 0.00060 |
107
+ | 3000 | 0.0983 | 0.098 | 0.950 | None | 0.00089 | 0.00063 |
108
+ | 3500 | 0.0769 | 0.077 | 0.960 | None | 0.00079 | 0.00250 |
109
+ | 4000 | 0.0683 | 0.068 | 0.965 | None | 0.00075 | 0.00048 |
110
+ | 1 | 1.9096 | 1.910 | 0.265 | None | 0.00000 | 0.00000 |
111
+ | 50 | 1.9126 | 1.913 | 0.259 | None | 0.00031 | 0.00023 |
112
+ | 250 | 1.7418 | 1.742 | 0.297 | None | 0.00166 | 0.00058 |
113
+ | 500 | 0.8752 | 0.875 | 0.604 | None | 0.00111 | 0.00050 |
114
+ | 1000 | 0.4800 | 0.480 | 0.771 | None | 0.00087 | 0.00045 |
115
+ | 1500 | 0.2948 | 0.295 | 0.856 | None | 0.00086 | 0.00058 |
116
+ | 2000 | 0.1895 | 0.189 | 0.906 | None | 0.00092 | 0.00060 |
117
+ | 2500 | 0.1274 | 0.127 | 0.936 | None | 0.00092 | 0.00060 |
118
+ | 3000 | 0.0983 | 0.098 | 0.950 | None | 0.00089 | 0.00063 |
119
+ | 3500 | 0.0769 | 0.077 | 0.960 | None | 0.00079 | 0.00250 |
120
+ | 4000 | 0.0683 | 0.068 | 0.965 | None | 0.00075 | 0.00048 |
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: Qwen3's byte-level BPE spelling (tokenizer.json sha256 be75606093db2094..., vocab 151,669); 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/qwen3/mse_solo_o0/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/qwen3/mse_solo_o0/s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e72b3a181d3e802dbfa375490295def22976f5c32bc3f91065acd3d0bfc64ef3
3
+ size 54819240
surface/qwen3/mse_solo_o0/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a6659d5cf3ba4e86f3aa402cd7e48497abad3412d53ee67c14b0843c60e77424
3
+ size 264285
surface/qwen3/mse_solo_o0/train.log ADDED
@@ -0,0 +1,159 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 0.3s] THE JOB: qarm_mse_solo_o0: the surface arm s10_qwen (gpt2 spelling by /workspace/qarm/dl/qwen3_tokenizer; 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
+ [ 16.0s] 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
3
+ [ 16.6s] tokenizer /workspace/qarm/dl/qwen3_tokenizer (gpt2 spelling; vocab 151669; sha256 be75606093db2094): 'a taco truck' -> ['a', 'Ġtaco', 'Ġtruck']; A 12 bytes, B 14 bytes; sites [(1, 6, 7)]
4
+ [ 16.6s] 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)
5
+ [ 29.3s] 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 3.7 GB
6
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
7
+ [ 29.5s] streams: web text; partners none (two web rows a quiet chunk); held-out sets ['web'] x 8 batches
8
+ [ 29.5s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 116), 4,920 sites
9
+ [ 29.6s] read rows draw2: 512 captions (0 unspellable), A rows (512, 151), B rows (512, 176), 5,086 sites
10
+ [arms] attached 's10_qwen' for phase 'qwen_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
+ [ 40.7s] 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
+ [ 44.7s] standardization statistics from 11,974 teacher sites on 8 paired batches (a mean and a spread per block and dimension); sigma range at block 20 0.751..1.119; saved /workspace/qarm/out/anchors/qarm_mse_solo_o0_stats.pt
13
+ [ 246.1s] CLOSE 0 (step 0): the member's alignment read on draw 1 (512 captions, 4,920 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
14
+ [ 246.1s] block 8: 0.729 -> 0.729 / 0.729; A 1.0000; mse 1.715 cos 0.367
15
+ [ 246.1s] block 12: 0.635 -> 0.635 / 0.635; A 1.0000; mse 1.729 cos 0.303
16
+ [ 246.1s] block 16: 0.524 -> 0.524 / 0.524; A 1.0000; mse 1.732 cos 0.236
17
+ [ 246.1s] block 18: 0.493 -> 0.493 / 0.493; A 1.0000; mse 1.722 cos 0.220
18
+ [ 246.1s] block 20: 0.475 -> 0.475 / 0.475; A 1.0000; mse 1.766 cos 0.204
19
+ [ 246.1s] block 24: 0.446 -> 0.446 / 0.446; A 1.0000; mse 1.950 cos 0.184
20
+ [ 246.1s] block 28: 0.399 -> 0.399 / 0.399; A 1.0000; mse 2.210 cos 0.139
21
+ [ 246.1s] block 31: 0.350 -> 0.350 / 0.350; A 1.0000; mse 3.243 cos 0.086
22
+ [ 246.1s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.639 / 0.529 / 0.483 / 0.453 / 0.410
23
+ [ 246.1s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.350 against the bar .85; B2: the lowest A alignment 1.0000 against .995; fineweb share +0.00000 +- 0.00000 against +.012
24
+ [ 246.1s] 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 10.8 GB
25
+ [stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
26
+ [ 253.9s] step 1/4000: task 1.9096 (mse 1.910 cos 0.265); quiet A 0.00000 off 0.00000; 1,503 sites; 7.72 s/step; the job: 4.2 min spent, about 514.5 min of steps left; card peak 13.7 GB
27
+ [ 373.7s] step 50/4000: task 1.9126 (mse 1.913 cos 0.259); quiet A 0.00031 off 0.00023; 1,508 sites; 2.55 s/step; the job: 6.2 min spent, about 168.0 min of steps left; card peak 13.8 GB
28
+ [ 495.9s] step 100/4000: task 1.8930 (mse 1.893 cos 0.263); quiet A 0.00046 off 0.00032; 1,506 sites; 2.50 s/step; the job: 8.3 min spent, about 162.4 min of steps left; card peak 13.8 GB
29
+ [ 618.5s] step 150/4000: task 1.8558 (mse 1.856 cos 0.271); quiet A 0.00062 off 0.00040; 1,471 sites; 2.48 s/step; the job: 10.3 min spent, about 159.3 min of steps left; card peak 13.8 GB
30
+ [ 740.6s] step 200/4000: task 1.8027 (mse 1.803 cos 0.283); quiet A 0.00103 off 0.00062; 1,483 sites; 2.47 s/step; the job: 12.3 min spent, about 156.6 min of steps left; card peak 13.8 GB
31
+ [ 863.0s] step 250/4000: task 1.7418 (mse 1.742 cos 0.297); quiet A 0.00166 off 0.00058; 1,485 sites; 2.47 s/step; the job: 14.4 min spent, about 154.2 min of steps left; card peak 13.8 GB
32
+ [ 985.4s] step 300/4000: task 1.6622 (mse 1.662 cos 0.317); quiet A 0.00185 off 0.00097; 1,497 sites; 2.46 s/step; the job: 16.4 min spent, about 152.0 min of steps left; card peak 13.8 GB
33
+ [ 1107.8s] step 350/4000: task 1.5332 (mse 1.533 cos 0.354); quiet A 0.00195 off 0.00086; 1,487 sites; 2.46 s/step; the job: 18.5 min spent, about 149.8 min of steps left; card peak 13.8 GB
34
+ [ 1230.1s] step 400/4000: task 1.2924 (mse 1.292 cos 0.427); quiet A 0.00200 off 0.00079; 1,505 sites; 2.46 s/step; the job: 20.5 min spent, about 147.6 min of steps left; card peak 13.8 GB
35
+ [ 1352.5s] step 450/4000: task 1.0151 (mse 1.015 cos 0.540); quiet A 0.00149 off 0.00052; 1,593 sites; 2.46 s/step; the job: 22.5 min spent, about 145.5 min of steps left; card peak 13.8 GB
36
+ [ 1474.8s] step 500/4000: task 0.8752 (mse 0.875 cos 0.604); quiet A 0.00111 off 0.00050; 1,535 sites; 2.46 s/step; the job: 24.6 min spent, about 143.3 min of steps left; card peak 13.8 GB
37
+ [ 1596.9s] step 550/4000: task 0.7929 (mse 0.793 cos 0.640); quiet A 0.00100 off 0.00048; 1,433 sites; 2.46 s/step; the job: 26.6 min spent, about 141.2 min of steps left; card peak 13.8 GB
38
+ [ 1719.2s] step 600/4000: task 0.7351 (mse 0.735 cos 0.664); quiet A 0.00100 off 0.00081; 1,487 sites; 2.46 s/step; the job: 28.7 min spent, about 139.1 min of steps left; card peak 13.8 GB
39
+ [ 1841.4s] step 650/4000: task 0.6810 (mse 0.681 cos 0.688); quiet A 0.00092 off 0.00042; 1,558 sites; 2.45 s/step; the job: 30.7 min spent, about 137.0 min of steps left; card peak 13.8 GB
40
+ [ 1963.7s] step 700/4000: task 0.6383 (mse 0.638 cos 0.705); quiet A 0.00089 off 0.00042; 1,429 sites; 2.45 s/step; the job: 32.7 min spent, about 135.0 min of steps left; card peak 13.8 GB
41
+ [ 2086.1s] step 750/4000: task 0.6044 (mse 0.604 cos 0.719); quiet A 0.00084 off 0.00051; 1,468 sites; 2.45 s/step; the job: 34.8 min spent, about 132.9 min of steps left; card peak 13.8 GB
42
+ [ 2208.4s] step 800/4000: task 0.5729 (mse 0.573 cos 0.731); quiet A 0.00086 off 0.00040; 1,541 sites; 2.45 s/step; the job: 36.8 min spent, about 130.8 min of steps left; card peak 13.8 GB
43
+ [ 2330.8s] step 850/4000: task 0.5489 (mse 0.549 cos 0.742); quiet A 0.00086 off 0.00040; 1,440 sites; 2.45 s/step; the job: 38.8 min spent, about 128.8 min of steps left; card peak 13.8 GB
44
+ [ 2453.1s] step 900/4000: task 0.5186 (mse 0.519 cos 0.755); quiet A 0.00094 off 0.00044; 1,652 sites; 2.45 s/step; the job: 40.9 min spent, about 126.7 min of steps left; card peak 13.8 GB
45
+ [ 2575.4s] step 950/4000: task 0.4972 (mse 0.497 cos 0.764); quiet A 0.00090 off 0.00043; 1,550 sites; 2.45 s/step; the job: 42.9 min spent, about 124.6 min of steps left; card peak 13.8 GB
46
+ [ 2697.9s] step 1000/4000: task 0.4800 (mse 0.480 cos 0.771); quiet A 0.00087 off 0.00045; 1,493 sites; 2.45 s/step; the job: 45.0 min spent, about 122.6 min of steps left; card peak 13.8 GB
47
+ [ 2698.2s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o0_close1_s10_qwen.safetensors
48
+ [ 2852.1s] CLOSE 1 (step 1000): the member's alignment read on draw 1 (512 captions, 4,920 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
49
+ [ 2852.1s] block 8: 0.729 -> 0.873 / 0.879; A 0.9896; mse 0.341 cos 0.828
50
+ [ 2852.1s] block 12: 0.635 -> 0.873 / 0.881; A 0.9875; mse 0.268 cos 0.867
51
+ [ 2852.1s] block 16: 0.524 -> 0.880 / 0.886; A 0.9877; mse 0.207 cos 0.899
52
+ [ 2852.1s] block 18: 0.493 -> 0.886 / 0.891; A 0.9902; mse 0.193 cos 0.907
53
+ [ 2852.1s] block 20: 0.475 -> 0.893 / 0.900; A 0.9881; mse 0.184 cos 0.913
54
+ [ 2852.1s] block 24: 0.446 -> 0.899 / 0.906; A 0.9872; mse 0.167 cos 0.923
55
+ [ 2852.1s] block 28: 0.399 -> 0.887 / 0.893; A 0.9861; mse 0.170 cos 0.915
56
+ [ 2852.1s] block 31: 0.350 -> 0.864 / 0.872; A 0.9781; mse 0.164 cos 0.898
57
+ [ 2852.1s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.868 / 0.878 / 0.889 / 0.895 / 0.885
58
+ [ 2852.1s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.864 against the bar .85; B2: the lowest A alignment 0.9781 against .995; fineweb share +0.00012 +- 0.00011 against +.012
59
+ [ 2852.1s] gate mean 0.116 / widest 0.167; 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 13.8 GB; host RSS 13.9 GB
60
+ [ 2974.6s] step 1050/4000: task 0.4542 (mse 0.454 cos 0.782); quiet A 0.00087 off 0.00047; 1,510 sites; 2.45 s/step; the job: 49.6 min spent, about 120.5 min of steps left; card peak 13.8 GB
61
+ [ 3097.3s] step 1100/4000: task 0.4305 (mse 0.430 cos 0.792); quiet A 0.00089 off 0.00043; 1,491 sites; 2.45 s/step; the job: 51.6 min spent, about 118.5 min of steps left; card peak 13.8 GB
62
+ [ 3219.9s] step 1150/4000: task 0.4174 (mse 0.417 cos 0.798); quiet A 0.00087 off 0.00044; 1,475 sites; 2.45 s/step; the job: 53.7 min spent, about 116.5 min of steps left; card peak 13.8 GB
63
+ [ 3342.4s] step 1200/4000: task 0.3953 (mse 0.395 cos 0.808); quiet A 0.00086 off 0.00044; 1,492 sites; 2.45 s/step; the job: 55.7 min spent, about 114.4 min of steps left; card peak 13.8 GB
64
+ [ 3465.1s] step 1250/4000: task 0.3761 (mse 0.376 cos 0.817); quiet A 0.00085 off 0.00089; 1,473 sites; 2.45 s/step; the job: 57.8 min spent, about 112.4 min of steps left; card peak 13.8 GB
65
+ [ 3587.9s] step 1300/4000: task 0.3598 (mse 0.360 cos 0.825); quiet A 0.00088 off 0.00045; 1,475 sites; 2.45 s/step; the job: 59.8 min spent, about 110.3 min of steps left; card peak 13.8 GB
66
+ [ 3710.6s] step 1350/4000: task 0.3417 (mse 0.342 cos 0.834); quiet A 0.00085 off 0.00046; 1,476 sites; 2.45 s/step; the job: 61.8 min spent, about 108.3 min of steps left; card peak 13.8 GB
67
+ [ 3833.1s] step 1400/4000: task 0.3236 (mse 0.324 cos 0.842); quiet A 0.00086 off 0.00048; 1,519 sites; 2.45 s/step; the job: 63.9 min spent, about 106.3 min of steps left; card peak 13.8 GB
68
+ [ 3955.8s] step 1450/4000: task 0.3096 (mse 0.310 cos 0.850); quiet A 0.00087 off 0.00054; 1,500 sites; 2.45 s/step; the job: 65.9 min spent, about 104.2 min of steps left; card peak 13.8 GB
69
+ [ 4078.4s] step 1500/4000: task 0.2948 (mse 0.295 cos 0.856); quiet A 0.00086 off 0.00058; 1,442 sites; 2.45 s/step; the job: 68.0 min spent, about 102.2 min of steps left; card peak 13.8 GB
70
+ [ 4200.9s] step 1550/4000: task 0.2800 (mse 0.280 cos 0.864); quiet A 0.00088 off 0.00054; 1,466 sites; 2.45 s/step; the job: 70.0 min spent, about 100.1 min of steps left; card peak 13.8 GB
71
+ [ 4323.6s] step 1600/4000: task 0.2659 (mse 0.266 cos 0.870); quiet A 0.00088 off 0.00049; 1,480 sites; 2.45 s/step; the job: 72.1 min spent, about 98.1 min of steps left; card peak 13.8 GB
72
+ [ 4446.2s] step 1650/4000: task 0.2550 (mse 0.255 cos 0.875); quiet A 0.00095 off 0.00117; 1,535 sites; 2.45 s/step; the job: 74.1 min spent, about 96.0 min of steps left; card peak 13.8 GB
73
+ [ 4568.8s] step 1700/4000: task 0.2479 (mse 0.248 cos 0.879); quiet A 0.00097 off 0.00059; 1,499 sites; 2.45 s/step; the job: 76.1 min spent, about 94.0 min of steps left; card peak 13.8 GB
74
+ [ 4691.5s] step 1750/4000: task 0.2305 (mse 0.231 cos 0.886); quiet A 0.00092 off 0.00056; 1,463 sites; 2.45 s/step; the job: 78.2 min spent, about 92.0 min of steps left; card peak 13.8 GB
75
+ [ 4814.4s] step 1800/4000: task 0.2228 (mse 0.223 cos 0.890); quiet A 0.00090 off 0.00060; 1,531 sites; 2.45 s/step; the job: 80.2 min spent, about 89.9 min of steps left; card peak 13.8 GB
76
+ [ 4937.1s] step 1850/4000: task 0.2130 (mse 0.213 cos 0.895); quiet A 0.00093 off 0.00055; 1,498 sites; 2.45 s/step; the job: 82.3 min spent, about 87.9 min of steps left; card peak 13.8 GB
77
+ [ 5059.9s] step 1900/4000: task 0.2081 (mse 0.208 cos 0.898); quiet A 0.00093 off 0.00063; 1,601 sites; 2.45 s/step; the job: 84.3 min spent, about 85.8 min of steps left; card peak 13.8 GB
78
+ [ 5182.6s] step 1950/4000: task 0.2005 (mse 0.201 cos 0.901); quiet A 0.00093 off 0.00056; 1,500 sites; 2.45 s/step; the job: 86.4 min spent, about 83.8 min of steps left; card peak 13.8 GB
79
+ [ 5305.3s] step 2000/4000: task 0.1895 (mse 0.189 cos 0.906); quiet A 0.00092 off 0.00060; 1,501 sites; 2.45 s/step; the job: 88.4 min spent, about 81.7 min of steps left; card peak 13.8 GB
80
+ [ 5305.5s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o0_close2_s10_qwen.safetensors
81
+ [ 5462.8s] CLOSE 2 (step 2000): the member's alignment read on draw 1 (512 captions, 4,920 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
82
+ [ 5462.8s] block 8: 0.729 -> 0.932 / 0.933; A 0.9953; mse 0.134 cos 0.930
83
+ [ 5462.8s] block 12: 0.635 -> 0.936 / 0.938; A 0.9956; mse 0.106 cos 0.947
84
+ [ 5462.8s] block 16: 0.524 -> 0.945 / 0.947; A 0.9946; mse 0.083 cos 0.960
85
+ [ 5462.8s] block 18: 0.493 -> 0.948 / 0.950; A 0.9959; mse 0.076 cos 0.963
86
+ [ 5462.8s] block 20: 0.475 -> 0.951 / 0.952; A 0.9951; mse 0.072 cos 0.966
87
+ [ 5462.8s] block 24: 0.446 -> 0.956 / 0.957; A 0.9955; mse 0.065 cos 0.971
88
+ [ 5462.8s] block 28: 0.399 -> 0.948 / 0.950; A 0.9950; mse 0.068 cos 0.967
89
+ [ 5462.8s] block 31: 0.350 -> 0.935 / 0.937; A 0.9886; mse 0.070 cos 0.958
90
+ [ 5462.8s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.931 / 0.940 / 0.946 / 0.952 / 0.946
91
+ [ 5462.8s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.935 against the bar .85; B2: the lowest A alignment 0.9886 against .995; fineweb share +0.00001 +- 0.00016 against +.012
92
+ [ 5462.8s] gate mean 0.135 / widest 0.279; 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 13.8 GB; host RSS 13.9 GB
93
+ [ 5585.2s] step 2050/4000: task 0.1804 (mse 0.180 cos 0.911); quiet A 0.00091 off 0.00058; 1,509 sites; 2.45 s/step; the job: 93.1 min spent, about 79.7 min of steps left; card peak 13.8 GB
94
+ [ 5707.6s] step 2100/4000: task 0.1792 (mse 0.179 cos 0.911); quiet A 0.00096 off 0.00069; 1,504 sites; 2.45 s/step; the job: 95.1 min spent, about 77.7 min of steps left; card peak 13.8 GB
95
+ [ 5830.2s] step 2150/4000: task 0.1685 (mse 0.169 cos 0.916); quiet A 0.00093 off 0.00058; 1,495 sites; 2.45 s/step; the job: 97.2 min spent, about 75.6 min of steps left; card peak 13.8 GB
96
+ [ 5952.8s] step 2200/4000: task 0.1625 (mse 0.163 cos 0.919); quiet A 0.00096 off 0.00055; 1,510 sites; 2.45 s/step; the job: 99.2 min spent, about 73.6 min of steps left; card peak 13.8 GB
97
+ [ 6075.2s] step 2250/4000: task 0.1589 (mse 0.159 cos 0.921); quiet A 0.00096 off 0.00065; 1,547 sites; 2.45 s/step; the job: 101.3 min spent, about 71.5 min of steps left; card peak 13.8 GB
98
+ [ 6197.7s] step 2300/4000: task 0.1528 (mse 0.153 cos 0.924); quiet A 0.00091 off 0.00056; 1,479 sites; 2.45 s/step; the job: 103.3 min spent, about 69.5 min of steps left; card peak 13.8 GB
99
+ [ 6320.2s] step 2350/4000: task 0.1501 (mse 0.150 cos 0.925); quiet A 0.00095 off 0.00058; 1,529 sites; 2.45 s/step; the job: 105.3 min spent, about 67.4 min of steps left; card peak 13.8 GB
100
+ [ 6442.6s] step 2400/4000: task 0.1409 (mse 0.141 cos 0.929); quiet A 0.00092 off 0.00057; 1,459 sites; 2.45 s/step; the job: 107.4 min spent, about 65.4 min of steps left; card peak 13.8 GB
101
+ [ 6565.3s] step 2450/4000: task 0.1331 (mse 0.133 cos 0.933); quiet A 0.00092 off 0.00063; 1,471 sites; 2.45 s/step; the job: 109.4 min spent, about 63.3 min of steps left; card peak 13.8 GB
102
+ [ 6687.8s] step 2500/4000: task 0.1274 (mse 0.127 cos 0.936); quiet A 0.00092 off 0.00060; 1,553 sites; 2.45 s/step; the job: 111.5 min spent, about 61.3 min of steps left; card peak 13.8 GB
103
+ [ 6810.4s] step 2550/4000: task 0.1209 (mse 0.121 cos 0.939); quiet A 0.00093 off 0.00060; 1,506 sites; 2.45 s/step; the job: 113.5 min spent, about 59.3 min of steps left; card peak 13.8 GB
104
+ [ 6932.9s] step 2600/4000: task 0.1250 (mse 0.125 cos 0.937); quiet A 0.00093 off 0.00062; 1,443 sites; 2.45 s/step; the job: 115.5 min spent, about 57.2 min of steps left; card peak 13.8 GB
105
+ [ 7055.3s] step 2650/4000: task 0.1178 (mse 0.118 cos 0.940); quiet A 0.00087 off 0.00066; 1,447 sites; 2.45 s/step; the job: 117.6 min spent, about 55.2 min of steps left; card peak 13.8 GB
106
+ [ 7177.7s] step 2700/4000: task 0.1133 (mse 0.113 cos 0.943); quiet A 0.00089 off 0.00057; 1,511 sites; 2.45 s/step; the job: 119.6 min spent, about 53.1 min of steps left; card peak 13.8 GB
107
+ [ 7300.4s] step 2750/4000: task 0.1108 (mse 0.111 cos 0.944); quiet A 0.00095 off 0.00081; 1,482 sites; 2.45 s/step; the job: 121.7 min spent, about 51.1 min of steps left; card peak 13.8 GB
108
+ [ 7422.9s] step 2800/4000: task 0.1109 (mse 0.111 cos 0.944); quiet A 0.00091 off 0.00077; 1,489 sites; 2.45 s/step; the job: 123.7 min spent, about 49.0 min of steps left; card peak 13.8 GB
109
+ [ 7545.5s] step 2850/4000: task 0.1038 (mse 0.104 cos 0.947); quiet A 0.00088 off 0.00062; 1,523 sites; 2.45 s/step; the job: 125.8 min spent, about 47.0 min of steps left; card peak 13.8 GB
110
+ [ 7668.1s] step 2900/4000: task 0.1129 (mse 0.113 cos 0.945); quiet A 0.00129 off 0.00054; 1,518 sites; 2.45 s/step; the job: 127.8 min spent, about 44.9 min of steps left; card peak 13.9 GB
111
+ [ 7790.6s] step 2950/4000: task 0.0999 (mse 0.100 cos 0.949); quiet A 0.00094 off 0.00063; 1,497 sites; 2.45 s/step; the job: 129.8 min spent, about 42.9 min of steps left; card peak 13.9 GB
112
+ [ 7913.4s] step 3000/4000: task 0.0983 (mse 0.098 cos 0.950); quiet A 0.00089 off 0.00063; 1,480 sites; 2.45 s/step; the job: 131.9 min spent, about 40.9 min of steps left; card peak 13.9 GB
113
+ [ 7913.7s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o0_close3_s10_qwen.safetensors
114
+ [ 8067.8s] CLOSE 3 (step 3000): the member's alignment read on draw 1 (512 captions, 4,920 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
115
+ [ 8067.8s] block 8: 0.729 -> 0.947 / 0.948; A 0.9953; mse 0.086 cos 0.955
116
+ [ 8067.8s] block 12: 0.635 -> 0.953 / 0.954; A 0.9967; mse 0.070 cos 0.966
117
+ [ 8067.8s] block 16: 0.524 -> 0.959 / 0.960; A 0.9957; mse 0.054 cos 0.974
118
+ [ 8067.8s] block 18: 0.493 -> 0.963 / 0.963; A 0.9964; mse 0.050 cos 0.976
119
+ [ 8067.8s] block 20: 0.475 -> 0.965 / 0.966; A 0.9953; mse 0.047 cos 0.978
120
+ [ 8067.8s] block 24: 0.446 -> 0.970 / 0.970; A 0.9961; mse 0.043 cos 0.981
121
+ [ 8067.8s] block 28: 0.399 -> 0.963 / 0.962; A 0.9954; mse 0.044 cos 0.978
122
+ [ 8067.8s] block 31: 0.350 -> 0.953 / 0.952; A 0.9918; mse 0.048 cos 0.971
123
+ [ 8067.8s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.950 / 0.957 / 0.962 / 0.967 / 0.961
124
+ [ 8067.8s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.953 against the bar .85; B2: the lowest A alignment 0.9891 against .995; fineweb share +0.00025 +- 0.00023 against +.012
125
+ [ 8067.8s] gate mean 0.150 / widest 0.390; 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 13.9 GB
126
+ [ 8190.6s] step 3050/4000: task 0.0941 (mse 0.094 cos 0.952); quiet A 0.00086 off 0.00053; 1,458 sites; 2.45 s/step; the job: 136.5 min spent, about 38.8 min of steps left; card peak 13.9 GB
127
+ [ 8313.4s] step 3100/4000: task 0.0934 (mse 0.093 cos 0.952); quiet A 0.00084 off 0.00079; 1,561 sites; 2.46 s/step; the job: 138.6 min spent, about 36.8 min of steps left; card peak 13.9 GB
128
+ [ 8436.2s] step 3150/4000: task 0.0911 (mse 0.091 cos 0.954); quiet A 0.00084 off 0.00052; 1,440 sites; 2.46 s/step; the job: 140.6 min spent, about 34.7 min of steps left; card peak 13.9 GB
129
+ [ 8558.9s] step 3200/4000: task 0.0876 (mse 0.088 cos 0.955); quiet A 0.00082 off 0.00058; 1,519 sites; 2.46 s/step; the job: 142.6 min spent, about 32.7 min of steps left; card peak 13.9 GB
130
+ [ 8681.4s] step 3250/4000: task 0.0843 (mse 0.084 cos 0.957); quiet A 0.00080 off 0.00052; 1,475 sites; 2.45 s/step; the job: 144.7 min spent, about 30.7 min of steps left; card peak 13.9 GB
131
+ [ 8804.1s] step 3300/4000: task 0.0836 (mse 0.084 cos 0.957); quiet A 0.00079 off 0.00054; 1,474 sites; 2.45 s/step; the job: 146.7 min spent, about 28.6 min of steps left; card peak 13.9 GB
132
+ [ 8926.6s] step 3350/4000: task 0.0896 (mse 0.090 cos 0.955); quiet A 0.00082 off 0.00050; 1,541 sites; 2.45 s/step; the job: 148.8 min spent, about 26.6 min of steps left; card peak 13.9 GB
133
+ [ 9048.9s] step 3400/4000: task 0.0806 (mse 0.081 cos 0.958); quiet A 0.00081 off 0.00051; 1,520 sites; 2.45 s/step; the job: 150.8 min spent, about 24.5 min of steps left; card peak 13.9 GB
134
+ [ 9171.4s] step 3450/4000: task 0.0791 (mse 0.079 cos 0.959); quiet A 0.00081 off 0.00066; 1,454 sites; 2.45 s/step; the job: 152.9 min spent, about 22.5 min of steps left; card peak 13.9 GB
135
+ [ 9293.9s] step 3500/4000: task 0.0769 (mse 0.077 cos 0.960); quiet A 0.00079 off 0.00250; 1,505 sites; 2.45 s/step; the job: 154.9 min spent, about 20.4 min of steps left; card peak 13.9 GB
136
+ [ 9416.6s] step 3550/4000: task 0.0795 (mse 0.080 cos 0.959); quiet A 0.00091 off 0.00055; 1,525 sites; 2.45 s/step; the job: 156.9 min spent, about 18.4 min of steps left; card peak 13.9 GB
137
+ [ 9539.5s] step 3600/4000: task 0.0759 (mse 0.076 cos 0.961); quiet A 0.00081 off 0.00052; 1,519 sites; 2.45 s/step; the job: 159.0 min spent, about 16.3 min of steps left; card peak 13.9 GB
138
+ [ 9662.4s] step 3650/4000: task 0.0761 (mse 0.076 cos 0.961); quiet A 0.00079 off 0.00054; 1,472 sites; 2.45 s/step; the job: 161.0 min spent, about 14.3 min of steps left; card peak 13.9 GB
139
+ [ 9785.5s] step 3700/4000: task 0.0714 (mse 0.071 cos 0.963); quiet A 0.00077 off 0.00052; 1,463 sites; 2.45 s/step; the job: 163.1 min spent, about 12.3 min of steps left; card peak 13.9 GB
140
+ [ 9908.3s] step 3750/4000: task 0.0707 (mse 0.071 cos 0.963); quiet A 0.00075 off 0.00055; 1,570 sites; 2.45 s/step; the job: 165.1 min spent, about 10.2 min of steps left; card peak 13.9 GB
141
+ [10031.1s] step 3800/4000: task 0.0706 (mse 0.071 cos 0.963); quiet A 0.00074 off 0.00050; 1,532 sites; 2.45 s/step; the job: 167.2 min spent, about 8.2 min of steps left; card peak 13.9 GB
142
+ [10154.1s] step 3850/4000: task 0.0687 (mse 0.069 cos 0.964); quiet A 0.00076 off 0.00051; 1,507 sites; 2.45 s/step; the job: 169.2 min spent, about 6.1 min of steps left; card peak 13.9 GB
143
+ [10276.9s] step 3900/4000: task 0.0680 (mse 0.068 cos 0.965); quiet A 0.00076 off 0.00050; 1,435 sites; 2.45 s/step; the job: 171.3 min spent, about 4.1 min of steps left; card peak 13.9 GB
144
+ [10399.6s] step 3950/4000: task 0.0673 (mse 0.067 cos 0.965); quiet A 0.00073 off 0.00048; 1,500 sites; 2.45 s/step; the job: 173.3 min spent, about 2.0 min of steps left; card peak 13.9 GB
145
+ [10522.2s] step 4000/4000: task 0.0683 (mse 0.068 cos 0.965); quiet A 0.00075 off 0.00048; 1,494 sites; 2.45 s/step; the job: 175.4 min spent, about 0.0 min of steps left; card peak 13.9 GB
146
+ [10522.4s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o0_close4_s10_qwen.safetensors
147
+ [10676.6s] CLOSE 4 (step 4000): the member's alignment read on draw 1 (512 captions, 4,920 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
148
+ [10676.6s] block 8: 0.729 -> 0.953 / 0.954; A 0.9963; mse 0.067 cos 0.965
149
+ [10676.6s] block 12: 0.635 -> 0.963 / 0.963; A 0.9972; mse 0.054 cos 0.974
150
+ [10676.6s] block 16: 0.524 -> 0.968 / 0.968; A 0.9957; mse 0.041 cos 0.980
151
+ [10676.6s] block 18: 0.493 -> 0.971 / 0.972; A 0.9973; mse 0.039 cos 0.982
152
+ [10676.6s] block 20: 0.475 -> 0.970 / 0.971; A 0.9953; mse 0.036 cos 0.983
153
+ [10676.6s] block 24: 0.446 -> 0.976 / 0.976; A 0.9974; mse 0.033 cos 0.985
154
+ [10676.6s] block 28: 0.399 -> 0.973 / 0.972; A 0.9970; mse 0.035 cos 0.983
155
+ [10676.6s] block 31: 0.350 -> 0.961 / 0.961; A 0.9935; mse 0.038 cos 0.977
156
+ [10676.6s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.958 / 0.966 / 0.970 / 0.975 / 0.968
157
+ [10676.6s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.961 against the bar .85; B2: the lowest A alignment 0.9886 against .995; fineweb share +0.00031 +- 0.00012 against +.012
158
+ [10676.6s] gate mean 0.162 / widest 0.472; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 128,688 (skipped {'test': 203, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 13.9 GB
159
+ [10676.6s] DONE qarm_mse_solo_o0: 4000 steps in 163.5 min of steps (2.45 s/step); 4 closes; 10677 s in all; card peak 13.9 GB; host RSS 13.9 GB; compiled chain: 5 cache entries (limit 256), 0 limit warnings