Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
04da644
·
verified ·
1 Parent(s): cf185d4

surface arm qwen3/mse_nce_gXA_o0: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/qwen3/mse_nce_gXA_o0/closes/qarm_mse_nce_gXA_o0_close1_s10_qwen.safetensors +3 -0
  2. surface/qwen3/mse_nce_gXA_o0/closes/qarm_mse_nce_gXA_o0_close2_s10_qwen.safetensors +3 -0
  3. surface/qwen3/mse_nce_gXA_o0/closes/qarm_mse_nce_gXA_o0_close3_s10_qwen.safetensors +3 -0
  4. surface/qwen3/mse_nce_gXA_o0/closes/qarm_mse_nce_gXA_o0_close4_s10_qwen.safetensors +3 -0
  5. surface/qwen3/mse_nce_gXA_o0/manifest.json +20 -0
  6. surface/qwen3/mse_nce_gXA_o0/readout.md +127 -0
  7. surface/qwen3/mse_nce_gXA_o0/results.json +0 -0
  8. surface/qwen3/mse_nce_gXA_o0/s10_qwen.safetensors +3 -0
  9. surface/qwen3/mse_nce_gXA_o0/stats.pt +3 -0
  10. surface/qwen3/mse_nce_gXA_o0/train.log +174 -0
surface/qwen3/mse_nce_gXA_o0/closes/qarm_mse_nce_gXA_o0_close1_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b4c070cdf7e48cd54657a8798f2100715a5e7712e11b17c41697c41542dffb33
3
+ size 54819216
surface/qwen3/mse_nce_gXA_o0/closes/qarm_mse_nce_gXA_o0_close2_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:42070865ec3aa6d2bee7500788d2f8d486b71a36a6aab8447053d3906f8b5b3d
3
+ size 54819216
surface/qwen3/mse_nce_gXA_o0/closes/qarm_mse_nce_gXA_o0_close3_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5214fbecdb5ba363bcb63e608f2f5e004c8d4b5c1928959d0dc1bf43b076db70
3
+ size 54819216
surface/qwen3/mse_nce_gXA_o0/closes/qarm_mse_nce_gXA_o0_close4_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac6428d5efbdefce546b646db632ff12e5f5f27214ff7326ea5f45734d237527
3
+ size 54819216
surface/qwen3/mse_nce_gXA_o0/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "qarm_mse_nce_gXA_o0",
3
+ "run": "mse_nce_gXA_o0",
4
+ "convention": "qwen3",
5
+ "member": "s10_qwen",
6
+ "final_anchor": "qarm_mse_nce_gXA_o0_close4_s10_qwen.safetensors",
7
+ "closes": [
8
+ "qarm_mse_nce_gXA_o0_close1_s10_qwen.safetensors",
9
+ "qarm_mse_nce_gXA_o0_close2_s10_qwen.safetensors",
10
+ "qarm_mse_nce_gXA_o0_close3_s10_qwen.safetensors",
11
+ "qarm_mse_nce_gXA_o0_close4_s10_qwen.safetensors"
12
+ ],
13
+ "finished": true,
14
+ "form": "mse_nce",
15
+ "base": "gXA",
16
+ "steps": 4000,
17
+ "trunk_step": 245674,
18
+ "shipped": "2026-10-08 04:49 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_nce_gXA_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_nce_gXA_o0: per-site MSE + 0.5 x state-table InfoNCE (tau 0.1); over the frozen eight of gXA (seed A); blocks 0-31; 4000 of 4000 steps; seed offset 0
8
+
9
+ ## The verdicts (the final close of each run)
10
+
11
+ ### qarm_mse_nce_gXA_o0
12
+
13
+ - B1 the gap: MET over every served block >= 12 (the lowest 0.890 at block 31 on draw 1, 0.884 at block 31 on draw 2; bar 0.85)
14
+ - B2 silence, whitened: NOT MET (the lowest A alignment 0.9766 at block 31 on draw 1, 0.9771 on draw 2; bar 0.995)
15
+ - B2 silence, in bits: MET (web +0.00017 bpb; the worst partner s3_rules +0.00044; bar +0.012)
16
+
17
+ ## The closes
18
+
19
+ ### qarm_mse_nce_gXA_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 | 0.322 / 0.319 | 0.640 / 0.486 / 0.407 | 1.0000 / 1.0000 | +0.00000 | 0.047 / 0.047 |
24
+ | 1 | 1000 | 0.701 / 0.694 | 0.860 / 0.872 / 0.848 | 0.9236 / 0.9286 | +0.00014 | 0.115 / 0.164 |
25
+ | 2 | 2000 | 0.813 / 0.812 | 0.920 / 0.931 / 0.919 | 0.9534 / 0.9510 | +0.00061 | 0.137 / 0.272 |
26
+ | 3 | 3000 | 0.865 / 0.858 | 0.941 / 0.954 / 0.945 | 0.9645 / 0.9653 | +0.00046 | 0.160 / 0.378 |
27
+ | 4 | 4000 | 0.890 / 0.884 | 0.954 / 0.964 / 0.954 | 0.9766 / 0.9771 | +0.00017 | 0.191 / 0.453 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### qarm_mse_nce_gXA_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 | 0.900 | 0.897 | 0.897 | 0.9992 | 0.498 | 0.728 |
36
+ | 4 | 0.790 | 0.946 | 0.947 | 0.9970 | 0.105 | 0.944 |
37
+ | 8 | 0.733 | 0.953 | 0.953 | 0.9964 | 0.081 | 0.959 |
38
+ | 12 | 0.640 | 0.954 | 0.955 | 0.9953 | 0.067 | 0.967 |
39
+ | 16 | 0.532 | 0.960 | 0.959 | 0.9951 | 0.056 | 0.973 |
40
+ | 18 | 0.502 | 0.964 | 0.963 | 0.9964 | 0.053 | 0.975 |
41
+ | 20 | 0.486 | 0.964 | 0.965 | 0.9937 | 0.051 | 0.976 |
42
+ | 22 | 0.472 | 0.968 | 0.967 | 0.9945 | 0.050 | 0.979 |
43
+ | 24 | 0.457 | 0.962 | 0.963 | 0.9941 | 0.050 | 0.978 |
44
+ | 26 | 0.430 | 0.964 | 0.962 | 0.9940 | 0.054 | 0.976 |
45
+ | 28 | 0.407 | 0.954 | 0.955 | 0.9938 | 0.065 | 0.972 |
46
+ | 29 | 0.398 | 0.951 | 0.951 | 0.9915 | 0.075 | 0.970 |
47
+ | 30 | 0.376 | 0.938 | 0.939 | 0.9898 | 0.121 | 0.967 |
48
+ | 31 | 0.322 | 0.890 | 0.894 | 0.9766 | 0.332 | 0.956 |
49
+
50
+ ### qarm_mse_nce_gXA_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 | 0.887 | 0.886 | 0.9986 | 0.513 | 0.722 |
55
+ | 4 | 0.795 | 0.944 | 0.945 | 0.9966 | 0.107 | 0.942 |
56
+ | 8 | 0.738 | 0.950 | 0.951 | 0.9960 | 0.082 | 0.958 |
57
+ | 12 | 0.644 | 0.952 | 0.952 | 0.9964 | 0.068 | 0.966 |
58
+ | 16 | 0.540 | 0.956 | 0.957 | 0.9957 | 0.057 | 0.972 |
59
+ | 18 | 0.505 | 0.958 | 0.959 | 0.9937 | 0.055 | 0.974 |
60
+ | 20 | 0.492 | 0.959 | 0.961 | 0.9956 | 0.053 | 0.975 |
61
+ | 22 | 0.481 | 0.960 | 0.960 | 0.9948 | 0.052 | 0.977 |
62
+ | 24 | 0.465 | 0.962 | 0.963 | 0.9954 | 0.052 | 0.977 |
63
+ | 26 | 0.440 | 0.958 | 0.958 | 0.9933 | 0.057 | 0.975 |
64
+ | 28 | 0.419 | 0.952 | 0.954 | 0.9930 | 0.068 | 0.970 |
65
+ | 29 | 0.409 | 0.947 | 0.948 | 0.9889 | 0.079 | 0.968 |
66
+ | 30 | 0.388 | 0.934 | 0.935 | 0.9894 | 0.128 | 0.966 |
67
+ | 31 | 0.319 | 0.884 | 0.889 | 0.9771 | 0.343 | 0.954 |
68
+
69
+ The member's share in bits per byte (all on minus the member masked): s1_perspective +0.00011, s2_concept +0.00005, s3_rules +0.00044, s4_arith +0.00027, s5_causal +0.00011, s6_tryfail +0.00006, s7_mixed +0.00020, s8_register +0.00008, web +0.00017
70
+
71
+ ## The training curves
72
+
73
+ ### qarm_mse_nce_gXA_o0
74
+
75
+ | step | task loss | mse | cos | InfoNCE | quiet on A | quiet off-domain |
76
+ |---|---|---|---|---|---|---|
77
+ | 1 | 4.5308 | 1.923 | 0.260 | 5.216 | 0.00025 | 0.00012 |
78
+ | 50 | 4.5626 | 1.929 | 0.255 | 5.268 | 0.00034 | 0.00016 |
79
+ | 250 | 4.1714 | 1.747 | 0.303 | 4.849 | 0.00309 | 0.00124 |
80
+ | 500 | 2.0935 | 0.863 | 0.619 | 2.461 | 0.00218 | 0.00062 |
81
+ | 1000 | 1.2437 | 0.496 | 0.771 | 1.496 | 0.00180 | 0.00055 |
82
+ | 1500 | 0.8099 | 0.307 | 0.854 | 1.005 | 0.00153 | 0.00059 |
83
+ | 2000 | 0.6165 | 0.208 | 0.900 | 0.818 | 0.00156 | 0.00065 |
84
+ | 2500 | 0.4933 | 0.146 | 0.929 | 0.695 | 0.00161 | 0.00124 |
85
+ | 3000 | 0.4448 | 0.117 | 0.943 | 0.656 | 0.00144 | 0.00058 |
86
+ | 3500 | 0.3913 | 0.093 | 0.953 | 0.597 | 0.00126 | 0.00051 |
87
+ | 4000 | 0.3844 | 0.086 | 0.957 | 0.598 | 0.00123 | 0.00047 |
88
+ | 1 | 4.5308 | 1.923 | 0.260 | 5.216 | 0.00025 | 0.00012 |
89
+ | 50 | 4.5626 | 1.929 | 0.255 | 5.268 | 0.00034 | 0.00016 |
90
+ | 250 | 4.1714 | 1.747 | 0.303 | 4.849 | 0.00309 | 0.00124 |
91
+ | 500 | 2.0935 | 0.863 | 0.619 | 2.461 | 0.00218 | 0.00062 |
92
+ | 1000 | 1.2437 | 0.496 | 0.771 | 1.496 | 0.00180 | 0.00055 |
93
+ | 1500 | 0.8099 | 0.307 | 0.854 | 1.005 | 0.00153 | 0.00059 |
94
+ | 2000 | 0.6165 | 0.208 | 0.900 | 0.818 | 0.00156 | 0.00065 |
95
+ | 2500 | 0.4933 | 0.146 | 0.929 | 0.695 | 0.00161 | 0.00124 |
96
+ | 3000 | 0.4448 | 0.117 | 0.943 | 0.656 | 0.00144 | 0.00058 |
97
+ | 3500 | 0.3913 | 0.093 | 0.953 | 0.597 | 0.00126 | 0.00051 |
98
+ | 4000 | 0.3844 | 0.086 | 0.957 | 0.598 | 0.00123 | 0.00047 |
99
+ | 1 | 4.5308 | 1.923 | 0.260 | 5.216 | 0.00025 | 0.00012 |
100
+ | 50 | 4.5626 | 1.929 | 0.255 | 5.268 | 0.00034 | 0.00016 |
101
+ | 250 | 4.1714 | 1.747 | 0.303 | 4.849 | 0.00309 | 0.00124 |
102
+ | 500 | 2.0935 | 0.863 | 0.619 | 2.461 | 0.00218 | 0.00062 |
103
+ | 1000 | 1.2437 | 0.496 | 0.771 | 1.496 | 0.00180 | 0.00055 |
104
+ | 1500 | 0.8099 | 0.307 | 0.854 | 1.005 | 0.00153 | 0.00059 |
105
+ | 2000 | 0.6165 | 0.208 | 0.900 | 0.818 | 0.00156 | 0.00065 |
106
+ | 2500 | 0.4933 | 0.146 | 0.929 | 0.695 | 0.00161 | 0.00124 |
107
+ | 3000 | 0.4448 | 0.117 | 0.943 | 0.656 | 0.00144 | 0.00058 |
108
+ | 3500 | 0.3913 | 0.093 | 0.953 | 0.597 | 0.00126 | 0.00051 |
109
+ | 4000 | 0.3844 | 0.086 | 0.957 | 0.598 | 0.00123 | 0.00047 |
110
+ | 1 | 4.5308 | 1.923 | 0.260 | 5.216 | 0.00025 | 0.00012 |
111
+ | 50 | 4.5626 | 1.929 | 0.255 | 5.268 | 0.00034 | 0.00016 |
112
+ | 250 | 4.1714 | 1.747 | 0.303 | 4.849 | 0.00309 | 0.00124 |
113
+ | 500 | 2.0935 | 0.863 | 0.619 | 2.461 | 0.00218 | 0.00062 |
114
+ | 1000 | 1.2437 | 0.496 | 0.771 | 1.496 | 0.00180 | 0.00055 |
115
+ | 1500 | 0.8099 | 0.307 | 0.854 | 1.005 | 0.00153 | 0.00059 |
116
+ | 2000 | 0.6165 | 0.208 | 0.900 | 0.818 | 0.00156 | 0.00065 |
117
+ | 2500 | 0.4933 | 0.146 | 0.929 | 0.695 | 0.00161 | 0.00124 |
118
+ | 3000 | 0.4448 | 0.117 | 0.943 | 0.656 | 0.00144 | 0.00058 |
119
+ | 3500 | 0.3913 | 0.093 | 0.953 | 0.597 | 0.00126 | 0.00051 |
120
+ | 4000 | 0.3844 | 0.086 | 0.957 | 0.598 | 0.00123 | 0.00047 |
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 + 0.5 x the state-table InfoNCE over the batch's sites; 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_nce_gXA_o0/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/qwen3/mse_nce_gXA_o0/s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac6428d5efbdefce546b646db632ff12e5f5f27214ff7326ea5f45734d237527
3
+ size 54819216
surface/qwen3/mse_nce_gXA_o0/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e3be7cc2b31738cac48db2765e218b57fb3341ba00fb3b470699c46a40d4d0c
3
+ size 264309
surface/qwen3/mse_nce_gXA_o0/train.log ADDED
@@ -0,0 +1,174 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 0.4s] THE JOB: qarm_mse_nce_gXA_o0: the Qwen tokenizer arm (s10_qwen) over the frozen eight of gXA; form mse_nce (tau 0.1, weight 0.5); 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
+ [ 28.8s] 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
3
+ [ 29.4s] Qwen3 tokenizer from /workspace/qarm/dl/qwen3_tokenizer (vocab 151669, tokenizer.json sha256 be75606093db2094): 'a taco truck' -> ['a', 'Ġtaco', 'Ġtruck']; A 12 bytes, B 14 bytes; sites [(1, 6, 7)]
4
+ [arms] attached 's1_perspective' for phase 'curriculum_s1': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
5
+ [arms] attached 's2_concept' for phase 'curriculum_s2': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
6
+ [arms] attached 's3_rules' for phase 'curriculum_s3': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
7
+ [arms] attached 's4_arith' for phase 'curriculum_s4': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
8
+ [arms] attached 's5_causal' for phase 'curriculum_s5': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
9
+ [arms] attached 's6_tryfail' for phase 'curriculum_s6': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
10
+ [arms] attached 's7_mixed' for phase 'curriculum_s7': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
11
+ [arms] attached 's8_register' for phase 'curriculum_s8': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
12
+ [ 41.8s] the eight of gXA (close 8) loaded in training order; they freeze when the member attaches
13
+ [ 41.8s] 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)
14
+ [ 54.9s] 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 4.5 GB
15
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
16
+ [stream] opening 'fineweb-good' (HuggingFaceFW/fineweb-edu) …
17
+ [stream] opening 'simple-wiki' (wikimedia/wikipedia) …
18
+ [stream] opening 'babi-prose' (Muennighoff/babi) …
19
+ Repo card metadata block was not found. Setting CardData to empty.
20
+ [stream] opening 'tinystories' (roneneldan/TinyStories) …
21
+ [stream] opening 'proofwriter-prose' (tasksource/proofwriter) …
22
+ [ 67.9s] streams: web text; partners ['s1_perspective', 's2_concept', 's3_rules', 's4_arith', 's5_causal', 's6_tryfail', 's8_register']; held-out sets ['s1_perspective', 's2_concept', 's3_rules', 's4_arith', 's5_causal', 's6_tryfail', 's7_mixed', 's8_register', 'web'] x 8 batches
23
+ [ 68.0s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 116), 4,920 sites
24
+ [ 68.0s] read rows draw2: 512 captions (0 unspellable), A rows (512, 151), B rows (512, 176), 5,086 sites
25
+ [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
26
+ [ 81.2s] BIRTH: the member attached over ['s1_perspective', 's2_concept', 's3_rules', 's4_arith', 's5_causal', 's6_tryfail', 's7_mixed', 's8_register']: 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)
27
+ [ 87.0s] 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 0.749..1.110; saved /workspace/qarm/out/anchors/qarm_mse_nce_gXA_o0_stats.pt
28
+ [ 349.3s] 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
29
+ [ 349.3s] block 8: 0.733 -> 0.733 / 0.733; A 1.0000; mse 1.718 cos 0.366
30
+ [ 349.3s] block 12: 0.640 -> 0.640 / 0.640; A 1.0000; mse 1.718 cos 0.307
31
+ [ 349.3s] block 16: 0.532 -> 0.532 / 0.532; A 1.0000; mse 1.699 cos 0.250
32
+ [ 349.3s] block 18: 0.502 -> 0.502 / 0.502; A 1.0000; mse 1.682 cos 0.236
33
+ [ 349.3s] block 20: 0.486 -> 0.486 / 0.486; A 1.0000; mse 1.699 cos 0.233
34
+ [ 349.3s] block 24: 0.457 -> 0.457 / 0.457; A 1.0000; mse 1.858 cos 0.231
35
+ [ 349.3s] block 28: 0.407 -> 0.407 / 0.407; A 1.0000; mse 2.189 cos 0.194
36
+ [ 349.3s] block 31: 0.322 -> 0.322 / 0.322; A 1.0000; mse 5.319 cos 0.105
37
+ [ 349.3s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.644 / 0.540 / 0.492 / 0.465 / 0.419
38
+ [ 349.3s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.322 against the bar .85; B2: the lowest A alignment 1.0000 against .995; fineweb share +0.00000 +- 0.00000 against +.012; partners' shares s1 +0.0000, s2 +0.0000, s3 +0.0000, s4 +0.0000, s5 +0.0000, s6 +0.0000, s7 +0.0000, s8 +0.0000
39
+ [ 349.3s] gate mean 0.047 / widest 0.047; compiled chain: 13 cache entries (limit 256), 0 limit warnings; docs drawn 256 (skipped {'test': 2, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 2.7 GB; host RSS 12.7 GB
40
+ [stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
41
+ [ 357.6s] step 1/4000: task 4.5308 (mse 1.923 cos 0.260 nce 5.216); quiet A 0.00025 off 0.00012; 1,503 sites; 8.20 s/step; the job: 6.0 min spent, about 546.5 min of steps left; card peak 15.1 GB
42
+ [ 528.6s] step 50/4000: task 4.5626 (mse 1.929 cos 0.255 nce 5.268); quiet A 0.00034 off 0.00016; 1,508 sites; 3.58 s/step; the job: 8.8 min spent, about 235.9 min of steps left; card peak 15.4 GB
43
+ [ 703.2s] step 100/4000: task 4.5154 (mse 1.907 cos 0.260 nce 5.217); quiet A 0.00061 off 0.00025; 1,506 sites; 3.54 s/step; the job: 11.7 min spent, about 230.0 min of steps left; card peak 15.4 GB
44
+ [ 878.3s] step 150/4000: task 4.4340 (mse 1.869 cos 0.270 nce 5.131); quiet A 0.00090 off 0.00037; 1,471 sites; 3.53 s/step; the job: 14.6 min spent, about 226.3 min of steps left; card peak 15.4 GB
45
+ [ 1053.5s] step 200/4000: task 4.3216 (mse 1.813 cos 0.285 nce 5.016); quiet A 0.00172 off 0.00074; 1,483 sites; 3.52 s/step; the job: 17.6 min spent, about 223.0 min of steps left; card peak 15.4 GB
46
+ [ 1228.3s] step 250/4000: task 4.1714 (mse 1.747 cos 0.303 nce 4.849); quiet A 0.00309 off 0.00124; 1,485 sites; 3.52 s/step; the job: 20.5 min spent, about 219.7 min of steps left; card peak 15.4 GB
47
+ [ 1403.1s] step 300/4000: task 3.9517 (mse 1.651 cos 0.332 nce 4.601); quiet A 0.00463 off 0.00174; 1,497 sites; 3.51 s/step; the job: 23.4 min spent, about 216.6 min of steps left; card peak 15.4 GB
48
+ [ 1578.0s] step 350/4000: task 3.5429 (mse 1.477 cos 0.386 nce 4.133); quiet A 0.00516 off 0.00175; 1,487 sites; 3.51 s/step; the job: 26.3 min spent, about 213.5 min of steps left; card peak 15.4 GB
49
+ [ 1752.9s] step 400/4000: task 2.8361 (mse 1.170 cos 0.487 nce 3.331); quiet A 0.00488 off 0.00131; 1,505 sites; 3.51 s/step; the job: 29.2 min spent, about 210.5 min of steps left; card peak 15.4 GB
50
+ [ 1927.7s] step 450/4000: task 2.3445 (mse 0.970 cos 0.572 nce 2.749); quiet A 0.00291 off 0.00075; 1,593 sites; 3.51 s/step; the job: 32.1 min spent, about 207.5 min of steps left; card peak 15.4 GB
51
+ [ 2102.6s] step 500/4000: task 2.0935 (mse 0.863 cos 0.619 nce 2.461); quiet A 0.00218 off 0.00062; 1,535 sites; 3.51 s/step; the job: 35.0 min spent, about 204.5 min of steps left; card peak 15.4 GB
52
+ [ 2277.5s] step 550/4000: task 1.9346 (mse 0.795 cos 0.647 nce 2.279); quiet A 0.00187 off 0.00051; 1,433 sites; 3.51 s/step; the job: 38.0 min spent, about 201.6 min of steps left; card peak 15.4 GB
53
+ [ 2452.8s] step 600/4000: task 1.8173 (mse 0.744 cos 0.668 nce 2.146); quiet A 0.00177 off 0.00054; 1,487 sites; 3.51 s/step; the job: 40.9 min spent, about 198.7 min of steps left; card peak 15.4 GB
54
+ [ 2628.2s] step 650/4000: task 1.6993 (mse 0.696 cos 0.689 nce 2.007); quiet A 0.00171 off 0.00051; 1,558 sites; 3.51 s/step; the job: 43.8 min spent, about 195.7 min of steps left; card peak 15.4 GB
55
+ [ 2803.1s] step 700/4000: task 1.6240 (mse 0.657 cos 0.704 nce 1.934); quiet A 0.00165 off 0.00053; 1,429 sites; 3.51 s/step; the job: 46.7 min spent, about 192.8 min of steps left; card peak 15.4 GB
56
+ [ 2977.9s] step 750/4000: task 1.5370 (mse 0.624 cos 0.718 nce 1.826); quiet A 0.00162 off 0.00048; 1,468 sites; 3.50 s/step; the job: 49.6 min spent, about 189.8 min of steps left; card peak 15.4 GB
57
+ [ 3152.4s] step 800/4000: task 1.4833 (mse 0.593 cos 0.729 nce 1.781); quiet A 0.00169 off 0.00049; 1,541 sites; 3.50 s/step; the job: 52.5 min spent, about 186.9 min of steps left; card peak 15.4 GB
58
+ [ 3327.2s] step 850/4000: task 1.4231 (mse 0.569 cos 0.740 nce 1.709); quiet A 0.00167 off 0.00050; 1,440 sites; 3.50 s/step; the job: 55.5 min spent, about 183.9 min of steps left; card peak 15.4 GB
59
+ [ 3501.7s] step 900/4000: task 1.3584 (mse 0.537 cos 0.753 nce 1.642); quiet A 0.00187 off 0.00052; 1,652 sites; 3.50 s/step; the job: 58.4 min spent, about 181.0 min of steps left; card peak 15.5 GB
60
+ [ 3675.9s] step 950/4000: task 1.2943 (mse 0.513 cos 0.763 nce 1.563); quiet A 0.00184 off 0.00056; 1,550 sites; 3.50 s/step; the job: 61.3 min spent, about 178.0 min of steps left; card peak 15.5 GB
61
+ [ 3850.8s] step 1000/4000: task 1.2437 (mse 0.496 cos 0.771 nce 1.496); quiet A 0.00180 off 0.00055; 1,493 sites; 3.50 s/step; the job: 64.2 min spent, about 175.1 min of steps left; card peak 15.5 GB
62
+ [ 3851.0s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_nce_gXA_o0_close1_s10_qwen.safetensors
63
+ [ 4056.6s] 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
64
+ [ 4056.6s] block 8: 0.733 -> 0.867 / 0.870; A 0.9864; mse 0.392 cos 0.811
65
+ [ 4056.6s] block 12: 0.640 -> 0.860 / 0.868; A 0.9834; mse 0.308 cos 0.851
66
+ [ 4056.6s] block 16: 0.532 -> 0.864 / 0.870; A 0.9828; mse 0.249 cos 0.880
67
+ [ 4056.6s] block 18: 0.502 -> 0.866 / 0.873; A 0.9832; mse 0.234 cos 0.888
68
+ [ 4056.6s] block 20: 0.486 -> 0.872 / 0.877; A 0.9805; mse 0.225 cos 0.894
69
+ [ 4056.6s] block 24: 0.457 -> 0.875 / 0.883; A 0.9780; mse 0.218 cos 0.903
70
+ [ 4056.6s] block 28: 0.407 -> 0.848 / 0.857; A 0.9756; mse 0.266 cos 0.883
71
+ [ 4056.6s] block 31: 0.322 -> 0.701 / 0.720; A 0.9236; mse 1.153 cos 0.837
72
+ [ 4056.6s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.856 / 0.857 / 0.866 / 0.872 / 0.842
73
+ [ 4056.6s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.701 against the bar .85; B2: the lowest A alignment 0.9236 against .995; fineweb share +0.00014 +- 0.00016 against +.012; partners' shares s1 +0.0002, s2 +0.0004, s3 +0.0008, s4 +0.0007, s5 +0.0002, s6 +0.0000, s7 -0.0001, s8 +0.0001
74
+ [ 4056.6s] gate mean 0.115 / widest 0.164; compiled chain: 15 cache entries (limit 256), 0 limit warnings; docs drawn 32,523 (skipped {'test': 44, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 15.5 GB; host RSS 13.4 GB
75
+ [ 4231.3s] step 1050/4000: task 1.1912 (mse 0.468 cos 0.782 nce 1.446); quiet A 0.00175 off 0.00057; 1,510 sites; 3.49 s/step; the job: 70.5 min spent, about 172.1 min of steps left; card peak 15.5 GB
76
+ [ 4406.0s] step 1100/4000: task 1.1430 (mse 0.442 cos 0.792 nce 1.401); quiet A 0.00161 off 0.00057; 1,491 sites; 3.49 s/step; the job: 73.4 min spent, about 169.2 min of steps left; card peak 15.5 GB
77
+ [ 4580.4s] step 1150/4000: task 1.1047 (mse 0.428 cos 0.799 nce 1.353); quiet A 0.00160 off 0.00063; 1,475 sites; 3.49 s/step; the job: 76.3 min spent, about 166.3 min of steps left; card peak 15.5 GB
78
+ [ 4755.1s] step 1200/4000: task 1.0483 (mse 0.406 cos 0.809 nce 1.285); quiet A 0.00172 off 0.00136; 1,492 sites; 3.49 s/step; the job: 79.3 min spent, about 163.3 min of steps left; card peak 15.5 GB
79
+ [ 4929.9s] step 1250/4000: task 0.9931 (mse 0.385 cos 0.819 nce 1.217); quiet A 0.00159 off 0.00054; 1,473 sites; 3.49 s/step; the job: 82.2 min spent, about 160.4 min of steps left; card peak 15.5 GB
80
+ [ 5104.5s] step 1300/4000: task 0.9541 (mse 0.369 cos 0.826 nce 1.171); quiet A 0.00161 off 0.00055; 1,475 sites; 3.49 s/step; the job: 85.1 min spent, about 157.5 min of steps left; card peak 15.5 GB
81
+ [ 5279.5s] step 1350/4000: task 0.9208 (mse 0.351 cos 0.834 nce 1.139); quiet A 0.00157 off 0.00060; 1,476 sites; 3.49 s/step; the job: 88.0 min spent, about 154.6 min of steps left; card peak 15.5 GB
82
+ [ 5454.6s] step 1400/4000: task 0.8906 (mse 0.335 cos 0.841 nce 1.111); quiet A 0.00163 off 0.00056; 1,519 sites; 3.49 s/step; the job: 90.9 min spent, about 151.6 min of steps left; card peak 15.5 GB
83
+ [ 5629.9s] step 1450/4000: task 0.8441 (mse 0.322 cos 0.848 nce 1.045); quiet A 0.00154 off 0.00077; 1,500 sites; 3.50 s/step; the job: 93.8 min spent, about 148.7 min of steps left; card peak 15.5 GB
84
+ [ 5805.1s] step 1500/4000: task 0.8099 (mse 0.307 cos 0.854 nce 1.005); quiet A 0.00153 off 0.00059; 1,442 sites; 3.50 s/step; the job: 96.8 min spent, about 145.8 min of steps left; card peak 15.5 GB
85
+ [ 5979.9s] step 1550/4000: task 0.8035 (mse 0.295 cos 0.860 nce 1.017); quiet A 0.00158 off 0.00064; 1,466 sites; 3.50 s/step; the job: 99.7 min spent, about 142.9 min of steps left; card peak 15.5 GB
86
+ [ 6154.9s] step 1600/4000: task 0.7685 (mse 0.282 cos 0.866 nce 0.973); quiet A 0.00156 off 0.00058; 1,480 sites; 3.50 s/step; the job: 102.6 min spent, about 140.0 min of steps left; card peak 15.5 GB
87
+ [ 6329.8s] step 1650/4000: task 0.7516 (mse 0.271 cos 0.871 nce 0.962); quiet A 0.00157 off 0.00057; 1,535 sites; 3.50 s/step; the job: 105.5 min spent, about 137.1 min of steps left; card peak 15.5 GB
88
+ [ 6505.7s] step 1700/4000: task 0.7317 (mse 0.265 cos 0.874 nce 0.933); quiet A 0.00166 off 0.00062; 1,499 sites; 3.50 s/step; the job: 108.4 min spent, about 134.2 min of steps left; card peak 15.5 GB
89
+ [ 6680.9s] step 1750/4000: task 0.6961 (mse 0.247 cos 0.881 nce 0.898); quiet A 0.00160 off 0.00059; 1,463 sites; 3.50 s/step; the job: 111.3 min spent, about 131.3 min of steps left; card peak 15.5 GB
90
+ [ 6856.0s] step 1800/4000: task 0.6761 (mse 0.239 cos 0.886 nce 0.874); quiet A 0.00164 off 0.00078; 1,531 sites; 3.50 s/step; the job: 114.3 min spent, about 128.3 min of steps left; card peak 15.5 GB
91
+ [ 7031.6s] step 1850/4000: task 0.6666 (mse 0.231 cos 0.889 nce 0.872); quiet A 0.00176 off 0.00070; 1,498 sites; 3.50 s/step; the job: 117.2 min spent, about 125.4 min of steps left; card peak 15.5 GB
92
+ [ 7206.9s] step 1900/4000: task 0.6515 (mse 0.225 cos 0.892 nce 0.852); quiet A 0.00169 off 0.00068; 1,601 sites; 3.50 s/step; the job: 120.1 min spent, about 122.5 min of steps left; card peak 15.5 GB
93
+ [ 7382.4s] step 1950/4000: task 0.6462 (mse 0.219 cos 0.895 nce 0.855); quiet A 0.00159 off 0.00067; 1,500 sites; 3.50 s/step; the job: 123.0 min spent, about 119.6 min of steps left; card peak 15.5 GB
94
+ [ 7557.7s] step 2000/4000: task 0.6165 (mse 0.208 cos 0.900 nce 0.818); quiet A 0.00156 off 0.00065; 1,501 sites; 3.50 s/step; the job: 126.0 min spent, about 116.7 min of steps left; card peak 15.5 GB
95
+ [ 7558.0s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_nce_gXA_o0_close2_s10_qwen.safetensors
96
+ [ 7763.7s] 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
97
+ [ 7763.7s] block 8: 0.733 -> 0.923 / 0.924; A 0.9910; mse 0.166 cos 0.918
98
+ [ 7763.7s] block 12: 0.640 -> 0.920 / 0.922; A 0.9919; mse 0.138 cos 0.933
99
+ [ 7763.7s] block 16: 0.532 -> 0.926 / 0.929; A 0.9886; mse 0.115 cos 0.945
100
+ [ 7763.7s] block 18: 0.502 -> 0.930 / 0.931; A 0.9919; mse 0.108 cos 0.949
101
+ [ 7763.7s] block 20: 0.486 -> 0.931 / 0.933; A 0.9919; mse 0.104 cos 0.951
102
+ [ 7763.7s] block 24: 0.457 -> 0.936 / 0.937; A 0.9884; mse 0.102 cos 0.955
103
+ [ 7763.7s] block 28: 0.407 -> 0.919 / 0.921; A 0.9868; mse 0.129 cos 0.944
104
+ [ 7763.7s] block 31: 0.322 -> 0.813 / 0.823; A 0.9534; mse 0.656 cos 0.917
105
+ [ 7763.7s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.919 / 0.922 / 0.928 / 0.931 / 0.916
106
+ [ 7763.7s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.813 against the bar .85; B2: the lowest A alignment 0.9534 against .995; fineweb share +0.00061 +- 0.00028 against +.012; partners' shares s1 +0.0003, s2 +0.0001, s3 +0.0005, s4 +0.0010, s5 +0.0002, s6 +0.0001, s7 -0.0001, s8 +0.0002
107
+ [ 7763.7s] gate mean 0.137 / widest 0.272; compiled chain: 15 cache entries (limit 256), 0 limit warnings; docs drawn 64,564 (skipped {'test': 97, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 15.5 GB; host RSS 13.4 GB
108
+ [ 7952.6s] step 2050/4000: task 0.6031 (mse 0.199 cos 0.904 nce 0.808); quiet A 0.00166 off 0.00065; 1,509 sites; 3.78 s/step; the job: 132.5 min spent, about 114.0 min of steps left; card peak 15.5 GB
109
+ [ 8128.8s] step 2100/4000: task 0.6002 (mse 0.197 cos 0.905 nce 0.806); quiet A 0.00163 off 0.00083; 1,504 sites; 3.65 s/step; the job: 135.5 min spent, about 111.1 min of steps left; card peak 15.5 GB
110
+ [ 8303.8s] step 2150/4000: task 0.5680 (mse 0.187 cos 0.910 nce 0.763); quiet A 0.00151 off 0.00060; 1,495 sites; 3.60 s/step; the job: 138.4 min spent, about 108.2 min of steps left; card peak 15.5 GB
111
+ [ 8479.0s] step 2200/4000: task 0.5619 (mse 0.181 cos 0.912 nce 0.762); quiet A 0.00160 off 0.00066; 1,510 sites; 3.58 s/step; the job: 141.3 min spent, about 105.2 min of steps left; card peak 15.5 GB
112
+ [ 8654.1s] step 2250/4000: task 0.5596 (mse 0.177 cos 0.915 nce 0.765); quiet A 0.00156 off 0.00061; 1,547 sites; 3.56 s/step; the job: 144.2 min spent, about 102.3 min of steps left; card peak 15.5 GB
113
+ [ 8828.7s] step 2300/4000: task 0.5360 (mse 0.169 cos 0.918 nce 0.735); quiet A 0.00151 off 0.00062; 1,479 sites; 3.55 s/step; the job: 147.1 min spent, about 99.4 min of steps left; card peak 15.5 GB
114
+ [ 9003.7s] step 2350/4000: task 0.5318 (mse 0.166 cos 0.920 nce 0.732); quiet A 0.00159 off 0.00069; 1,529 sites; 3.54 s/step; the job: 150.1 min spent, about 96.5 min of steps left; card peak 15.5 GB
115
+ [ 9178.4s] step 2400/4000: task 0.5120 (mse 0.158 cos 0.924 nce 0.709); quiet A 0.00157 off 0.00065; 1,459 sites; 3.54 s/step; the job: 153.0 min spent, about 93.5 min of steps left; card peak 15.5 GB
116
+ [ 9353.6s] step 2450/4000: task 0.5082 (mse 0.151 cos 0.926 nce 0.714); quiet A 0.00160 off 0.00063; 1,471 sites; 3.53 s/step; the job: 155.9 min spent, about 90.6 min of steps left; card peak 15.5 GB
117
+ [ 9528.6s] step 2500/4000: task 0.4933 (mse 0.146 cos 0.929 nce 0.695); quiet A 0.00161 off 0.00124; 1,553 sites; 3.53 s/step; the job: 158.8 min spent, about 87.7 min of steps left; card peak 15.5 GB
118
+ [ 9703.5s] step 2550/4000: task 0.4825 (mse 0.139 cos 0.932 nce 0.688); quiet A 0.00170 off 0.00071; 1,506 sites; 3.53 s/step; the job: 161.7 min spent, about 84.7 min of steps left; card peak 15.5 GB
119
+ [ 9878.6s] step 2600/4000: task 0.4940 (mse 0.143 cos 0.931 nce 0.702); quiet A 0.00155 off 0.00060; 1,443 sites; 3.52 s/step; the job: 164.6 min spent, about 81.8 min of steps left; card peak 15.5 GB
120
+ [10053.6s] step 2650/4000: task 0.4704 (mse 0.136 cos 0.933 nce 0.668); quiet A 0.00147 off 0.00060; 1,447 sites; 3.52 s/step; the job: 167.6 min spent, about 78.9 min of steps left; card peak 15.5 GB
121
+ [10228.9s] step 2700/4000: task 0.4709 (mse 0.131 cos 0.936 nce 0.679); quiet A 0.00146 off 0.00060; 1,511 sites; 3.52 s/step; the job: 170.5 min spent, about 76.0 min of steps left; card peak 15.5 GB
122
+ [10403.8s] step 2750/4000: task 0.4575 (mse 0.127 cos 0.937 nce 0.660); quiet A 0.00147 off 0.00068; 1,482 sites; 3.52 s/step; the job: 173.4 min spent, about 73.1 min of steps left; card peak 15.5 GB
123
+ [10578.3s] step 2800/4000: task 0.4615 (mse 0.129 cos 0.937 nce 0.665); quiet A 0.00146 off 0.00055; 1,489 sites; 3.52 s/step; the job: 176.3 min spent, about 70.1 min of steps left; card peak 15.5 GB
124
+ [10753.4s] step 2850/4000: task 0.4548 (mse 0.122 cos 0.940 nce 0.665); quiet A 0.00146 off 0.00063; 1,523 sites; 3.52 s/step; the job: 179.2 min spent, about 67.2 min of steps left; card peak 15.5 GB
125
+ [10928.1s] step 2900/4000: task 0.4749 (mse 0.131 cos 0.938 nce 0.689); quiet A 0.00198 off 0.00088; 1,518 sites; 3.52 s/step; the job: 182.1 min spent, about 64.3 min of steps left; card peak 15.8 GB
126
+ [11103.0s] step 2950/4000: task 0.4445 (mse 0.116 cos 0.942 nce 0.656); quiet A 0.00147 off 0.00081; 1,497 sites; 3.52 s/step; the job: 185.1 min spent, about 61.4 min of steps left; card peak 15.8 GB
127
+ [11277.9s] step 3000/4000: task 0.4448 (mse 0.117 cos 0.943 nce 0.656); quiet A 0.00144 off 0.00058; 1,480 sites; 3.51 s/step; the job: 188.0 min spent, about 58.4 min of steps left; card peak 15.8 GB
128
+ [11278.2s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_nce_gXA_o0_close3_s10_qwen.safetensors
129
+ [11483.6s] 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
130
+ [11483.6s] block 8: 0.733 -> 0.945 / 0.945; A 0.9936; mse 0.103 cos 0.948
131
+ [11483.6s] block 12: 0.640 -> 0.941 / 0.943; A 0.9920; mse 0.087 cos 0.958
132
+ [11483.6s] block 16: 0.532 -> 0.950 / 0.951; A 0.9922; mse 0.073 cos 0.965
133
+ [11483.6s] block 18: 0.502 -> 0.954 / 0.954; A 0.9948; mse 0.069 cos 0.967
134
+ [11483.6s] block 20: 0.486 -> 0.954 / 0.954; A 0.9941; mse 0.066 cos 0.969
135
+ [11483.6s] block 24: 0.457 -> 0.956 / 0.955; A 0.9919; mse 0.065 cos 0.972
136
+ [11483.6s] block 28: 0.407 -> 0.945 / 0.945; A 0.9920; mse 0.083 cos 0.964
137
+ [11483.6s] block 31: 0.322 -> 0.865 / 0.869; A 0.9645; mse 0.432 cos 0.945
138
+ [11483.6s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.942 / 0.944 / 0.950 / 0.953 / 0.941
139
+ [11483.6s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.865 against the bar .85; B2: the lowest A alignment 0.9645 against .995; fineweb share +0.00046 +- 0.00028 against +.012; partners' shares s1 +0.0002, s2 -0.0001, s3 +0.0006, s4 +0.0006, s5 +0.0001, s6 +0.0001, s7 +0.0001, s8 +0.0001
140
+ [11483.6s] gate mean 0.160 / widest 0.378; compiled chain: 15 cache entries (limit 256), 0 limit warnings; docs drawn 96,501 (skipped {'test': 156, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 15.8 GB; host RSS 13.4 GB
141
+ [11658.9s] step 3050/4000: task 0.4340 (mse 0.113 cos 0.944 nce 0.642); quiet A 0.00141 off 0.00064; 1,458 sites; 3.50 s/step; the job: 194.3 min spent, about 55.5 min of steps left; card peak 15.8 GB
142
+ [11833.6s] step 3100/4000: task 0.4289 (mse 0.112 cos 0.945 nce 0.634); quiet A 0.00138 off 0.00094; 1,561 sites; 3.50 s/step; the job: 197.2 min spent, about 52.6 min of steps left; card peak 15.8 GB
143
+ [12008.2s] step 3150/4000: task 0.4256 (mse 0.109 cos 0.946 nce 0.633); quiet A 0.00137 off 0.00054; 1,440 sites; 3.50 s/step; the job: 200.1 min spent, about 49.7 min of steps left; card peak 15.8 GB
144
+ [12183.2s] step 3200/4000: task 0.4248 (mse 0.106 cos 0.947 nce 0.638); quiet A 0.00136 off 0.00052; 1,519 sites; 3.50 s/step; the job: 203.1 min spent, about 46.7 min of steps left; card peak 15.8 GB
145
+ [12358.3s] step 3250/4000: task 0.4136 (mse 0.102 cos 0.949 nce 0.622); quiet A 0.00131 off 0.00059; 1,475 sites; 3.50 s/step; the job: 206.0 min spent, about 43.8 min of steps left; card peak 15.8 GB
146
+ [12533.0s] step 3300/4000: task 0.4143 (mse 0.103 cos 0.949 nce 0.622); quiet A 0.00140 off 0.00314; 1,474 sites; 3.50 s/step; the job: 208.9 min spent, about 40.9 min of steps left; card peak 15.8 GB
147
+ [12708.2s] step 3350/4000: task 0.4197 (mse 0.108 cos 0.947 nce 0.623); quiet A 0.00140 off 0.00063; 1,541 sites; 3.50 s/step; the job: 211.8 min spent, about 38.0 min of steps left; card peak 15.8 GB
148
+ [12883.5s] step 3400/4000: task 0.4066 (mse 0.098 cos 0.951 nce 0.616); quiet A 0.00134 off 0.00056; 1,520 sites; 3.50 s/step; the job: 214.7 min spent, about 35.0 min of steps left; card peak 15.8 GB
149
+ [13058.8s] step 3450/4000: task 0.4056 (mse 0.097 cos 0.952 nce 0.617); quiet A 0.00131 off 0.00056; 1,454 sites; 3.50 s/step; the job: 217.6 min spent, about 32.1 min of steps left; card peak 15.8 GB
150
+ [13233.8s] step 3500/4000: task 0.3913 (mse 0.093 cos 0.953 nce 0.597); quiet A 0.00126 off 0.00051; 1,505 sites; 3.50 s/step; the job: 220.6 min spent, about 29.2 min of steps left; card peak 15.8 GB
151
+ [13409.2s] step 3550/4000: task 0.3966 (mse 0.094 cos 0.953 nce 0.605); quiet A 0.00126 off 0.00128; 1,525 sites; 3.50 s/step; the job: 223.5 min spent, about 26.3 min of steps left; card peak 15.8 GB
152
+ [13584.3s] step 3600/4000: task 0.3940 (mse 0.093 cos 0.954 nce 0.603); quiet A 0.00130 off 0.00056; 1,519 sites; 3.50 s/step; the job: 226.4 min spent, about 23.4 min of steps left; card peak 15.8 GB
153
+ [13760.1s] step 3650/4000: task 0.4022 (mse 0.093 cos 0.953 nce 0.618); quiet A 0.00128 off 0.00052; 1,472 sites; 3.50 s/step; the job: 229.3 min spent, about 20.4 min of steps left; card peak 15.8 GB
154
+ [13936.0s] step 3700/4000: task 0.3904 (mse 0.089 cos 0.955 nce 0.604); quiet A 0.00129 off 0.00052; 1,463 sites; 3.50 s/step; the job: 232.3 min spent, about 17.5 min of steps left; card peak 15.8 GB
155
+ [14111.9s] step 3750/4000: task 0.3899 (mse 0.088 cos 0.956 nce 0.603); quiet A 0.00126 off 0.00069; 1,570 sites; 3.50 s/step; the job: 235.2 min spent, about 14.6 min of steps left; card peak 15.8 GB
156
+ [14287.7s] step 3800/4000: task 0.3890 (mse 0.088 cos 0.956 nce 0.602); quiet A 0.00124 off 0.00048; 1,532 sites; 3.51 s/step; the job: 238.1 min spent, about 11.7 min of steps left; card peak 15.8 GB
157
+ [14463.3s] step 3850/4000: task 0.3878 (mse 0.085 cos 0.957 nce 0.605); quiet A 0.00129 off 0.00054; 1,507 sites; 3.51 s/step; the job: 241.1 min spent, about 8.8 min of steps left; card peak 15.8 GB
158
+ [14638.5s] step 3900/4000: task 0.3728 (mse 0.084 cos 0.958 nce 0.577); quiet A 0.00123 off 0.00048; 1,435 sites; 3.51 s/step; the job: 244.0 min spent, about 5.8 min of steps left; card peak 15.8 GB
159
+ [14814.1s] step 3950/4000: task 0.3711 (mse 0.083 cos 0.958 nce 0.575); quiet A 0.00118 off 0.00055; 1,500 sites; 3.51 s/step; the job: 246.9 min spent, about 2.9 min of steps left; card peak 15.8 GB
160
+ [14989.7s] step 4000/4000: task 0.3844 (mse 0.086 cos 0.957 nce 0.598); quiet A 0.00123 off 0.00047; 1,494 sites; 3.51 s/step; the job: 249.8 min spent, about 0.0 min of steps left; card peak 15.8 GB
161
+ [14989.9s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_nce_gXA_o0_close4_s10_qwen.safetensors
162
+ [15198.2s] 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
163
+ [15198.2s] block 8: 0.733 -> 0.953 / 0.953; A 0.9964; mse 0.081 cos 0.959
164
+ [15198.2s] block 12: 0.640 -> 0.954 / 0.955; A 0.9953; mse 0.067 cos 0.967
165
+ [15198.2s] block 16: 0.532 -> 0.960 / 0.959; A 0.9951; mse 0.056 cos 0.973
166
+ [15198.2s] block 18: 0.502 -> 0.964 / 0.963; A 0.9964; mse 0.053 cos 0.975
167
+ [15198.2s] block 20: 0.486 -> 0.964 / 0.965; A 0.9937; mse 0.051 cos 0.976
168
+ [15198.2s] block 24: 0.457 -> 0.962 / 0.963; A 0.9941; mse 0.050 cos 0.978
169
+ [15198.2s] block 28: 0.407 -> 0.954 / 0.955; A 0.9938; mse 0.065 cos 0.972
170
+ [15198.2s] block 31: 0.322 -> 0.890 / 0.894; A 0.9766; mse 0.332 cos 0.956
171
+ [15198.2s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.952 / 0.956 / 0.959 / 0.962 / 0.952
172
+ [15198.2s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.890 against the bar .85; B2: the lowest A alignment 0.9766 against .995; fineweb share +0.00017 +- 0.00029 against +.012; partners' shares s1 +0.0001, s2 +0.0000, s3 +0.0004, s4 +0.0003, s5 +0.0001, s6 +0.0001, s7 +0.0002, s8 +0.0001
173
+ [15198.2s] gate mean 0.191 / widest 0.453; compiled chain: 15 cache entries (limit 256), 0 limit warnings; docs drawn 128,688 (skipped {'test': 203, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 15.8 GB; host RSS 13.4 GB
174
+ [15198.3s] DONE qarm_mse_nce_gXA_o0: 4000 steps in 233.7 min of steps (3.51 s/step); 4 closes; 15198 s in all; card peak 15.8 GB; host RSS 13.4 GB; compiled chain: 15 cache entries (limit 256), 0 limit warnings