Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
bc7435f
·
verified ·
1 Parent(s): 2de9523

surface arm qwen3/mse_gXA_o0_lamA8: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/qwen3/mse_gXA_o0_lamA8/closes/qarm_mse_gXA_o0_lamA8_close1_s10_qwen.safetensors +3 -0
  2. surface/qwen3/mse_gXA_o0_lamA8/closes/qarm_mse_gXA_o0_lamA8_close2_s10_qwen.safetensors +3 -0
  3. surface/qwen3/mse_gXA_o0_lamA8/closes/qarm_mse_gXA_o0_lamA8_close3_s10_qwen.safetensors +3 -0
  4. surface/qwen3/mse_gXA_o0_lamA8/closes/qarm_mse_gXA_o0_lamA8_close4_s10_qwen.safetensors +3 -0
  5. surface/qwen3/mse_gXA_o0_lamA8/manifest.json +20 -0
  6. surface/qwen3/mse_gXA_o0_lamA8/readout.md +127 -0
  7. surface/qwen3/mse_gXA_o0_lamA8/results.json +0 -0
  8. surface/qwen3/mse_gXA_o0_lamA8/s10_qwen.safetensors +3 -0
  9. surface/qwen3/mse_gXA_o0_lamA8/stats.pt +3 -0
  10. surface/qwen3/mse_gXA_o0_lamA8/train.log +174 -0
surface/qwen3/mse_gXA_o0_lamA8/closes/qarm_mse_gXA_o0_lamA8_close1_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:73838b4025c8645d45bb6be840117d20a882614dfac96b45d3d34f06fd937cb1
3
+ size 54819352
surface/qwen3/mse_gXA_o0_lamA8/closes/qarm_mse_gXA_o0_lamA8_close2_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:10bda42d0890e198b0bcc2699dae7ef3561047b1ad64cd277d2ab98495a82fe7
3
+ size 54819352
surface/qwen3/mse_gXA_o0_lamA8/closes/qarm_mse_gXA_o0_lamA8_close3_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0613a869bf2a1eed2124602655d7243b1816c03fd883b8b94dc6472849e2d213
3
+ size 54819352
surface/qwen3/mse_gXA_o0_lamA8/closes/qarm_mse_gXA_o0_lamA8_close4_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f22d3212a08807aab84ebb3b3c39c55707d018f52b4286fee557e7302b7d3446
3
+ size 54819352
surface/qwen3/mse_gXA_o0_lamA8/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "qarm_mse_gXA_o0_lamA8",
3
+ "run": "mse_gXA_o0_lamA8",
4
+ "convention": "qwen3",
5
+ "member": "s10_qwen",
6
+ "final_anchor": "qarm_mse_gXA_o0_lamA8_close4_s10_qwen.safetensors",
7
+ "closes": [
8
+ "qarm_mse_gXA_o0_lamA8_close1_s10_qwen.safetensors",
9
+ "qarm_mse_gXA_o0_lamA8_close2_s10_qwen.safetensors",
10
+ "qarm_mse_gXA_o0_lamA8_close3_s10_qwen.safetensors",
11
+ "qarm_mse_gXA_o0_lamA8_close4_s10_qwen.safetensors"
12
+ ],
13
+ "finished": true,
14
+ "form": "mse",
15
+ "base": "gXA",
16
+ "steps": 4000,
17
+ "trunk_step": 245674,
18
+ "shipped": "2026-10-08 16:07 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_gXA_o0_lamA8/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_gXA_o0_lamA8: per-site MSE; 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_gXA_o0_lamA8
12
+
13
+ - B1 the gap: MET over every served block >= 12 (the lowest 0.883 at block 31 on draw 1, 0.881 at block 31 on draw 2; bar 0.85)
14
+ - B2 silence, whitened: NOT MET (the lowest A alignment 0.9899 at block 31 on draw 1, 0.9907 on draw 2; bar 0.995)
15
+ - B2 silence, in bits: MET (web -0.00005 bpb; the worst partner s4_arith +0.00037; bar +0.012)
16
+
17
+ ## The closes
18
+
19
+ ### qarm_mse_gXA_o0_lamA8
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 | 0.705 / 0.692 | 0.861 / 0.874 / 0.852 | 0.9815 / 0.9865 | -0.00001 | 0.116 / 0.168 |
25
+ | 2 | 2000 | 0.823 / 0.820 | 0.924 / 0.938 / 0.925 | 0.9805 / 0.9813 | +0.00035 | 0.136 / 0.275 |
26
+ | 3 | 3000 | 0.864 / 0.857 | 0.941 / 0.955 / 0.945 | 0.9847 / 0.9907 | -0.00007 | 0.151 / 0.382 |
27
+ | 4 | 4000 | 0.883 / 0.881 | 0.951 / 0.964 / 0.954 | 0.9899 / 0.9907 | -0.00005 | 0.164 / 0.475 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### qarm_mse_gXA_o0_lamA8, 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 | 0.9998 | 0.366 | 0.773 |
36
+ | 4 | 0.790 | 0.946 | 0.946 | 0.9992 | 0.098 | 0.947 |
37
+ | 8 | 0.733 | 0.952 | 0.952 | 0.9991 | 0.076 | 0.961 |
38
+ | 12 | 0.640 | 0.951 | 0.952 | 0.9986 | 0.064 | 0.969 |
39
+ | 16 | 0.532 | 0.961 | 0.960 | 0.9981 | 0.054 | 0.974 |
40
+ | 18 | 0.502 | 0.963 | 0.963 | 0.9986 | 0.051 | 0.976 |
41
+ | 20 | 0.486 | 0.964 | 0.963 | 0.9968 | 0.050 | 0.976 |
42
+ | 22 | 0.472 | 0.966 | 0.966 | 0.9950 | 0.050 | 0.978 |
43
+ | 24 | 0.457 | 0.963 | 0.963 | 0.9985 | 0.050 | 0.978 |
44
+ | 26 | 0.430 | 0.962 | 0.963 | 0.9962 | 0.055 | 0.976 |
45
+ | 28 | 0.407 | 0.954 | 0.956 | 0.9972 | 0.066 | 0.971 |
46
+ | 29 | 0.398 | 0.950 | 0.950 | 0.9972 | 0.076 | 0.969 |
47
+ | 30 | 0.376 | 0.937 | 0.938 | 0.9957 | 0.128 | 0.965 |
48
+ | 31 | 0.322 | 0.883 | 0.884 | 0.9899 | 0.362 | 0.955 |
49
+
50
+ ### qarm_mse_gXA_o0_lamA8, 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 | 0.885 | 0.885 | 0.9993 | 0.377 | 0.767 |
55
+ | 4 | 0.795 | 0.946 | 0.945 | 0.9990 | 0.099 | 0.946 |
56
+ | 8 | 0.738 | 0.951 | 0.951 | 0.9984 | 0.077 | 0.961 |
57
+ | 12 | 0.644 | 0.952 | 0.952 | 0.9985 | 0.064 | 0.968 |
58
+ | 16 | 0.540 | 0.957 | 0.957 | 0.9983 | 0.054 | 0.974 |
59
+ | 18 | 0.505 | 0.962 | 0.962 | 0.9986 | 0.052 | 0.975 |
60
+ | 20 | 0.492 | 0.961 | 0.962 | 0.9984 | 0.051 | 0.976 |
61
+ | 22 | 0.481 | 0.961 | 0.963 | 0.9970 | 0.051 | 0.978 |
62
+ | 24 | 0.465 | 0.964 | 0.964 | 0.9985 | 0.051 | 0.977 |
63
+ | 26 | 0.440 | 0.958 | 0.959 | 0.9978 | 0.056 | 0.975 |
64
+ | 28 | 0.419 | 0.953 | 0.953 | 0.9967 | 0.067 | 0.971 |
65
+ | 29 | 0.409 | 0.948 | 0.949 | 0.9965 | 0.078 | 0.968 |
66
+ | 30 | 0.388 | 0.934 | 0.935 | 0.9965 | 0.130 | 0.965 |
67
+ | 31 | 0.319 | 0.881 | 0.881 | 0.9907 | 0.344 | 0.954 |
68
+
69
+ The member's share in bits per byte (all on minus the member masked): s1_perspective -0.00002, s2_concept +0.00011, s3_rules +0.00024, s4_arith +0.00037, s5_causal +0.00007, s6_tryfail +0.00007, s7_mixed +0.00014, s8_register +0.00003, web -0.00005
70
+
71
+ ## The training curves
72
+
73
+ ### qarm_mse_gXA_o0_lamA8
74
+
75
+ | step | task loss | mse | cos | InfoNCE | quiet on A | quiet off-domain |
76
+ |---|---|---|---|---|---|---|
77
+ | 1 | 1.9230 | 1.923 | 0.260 | None | 0.00025 | 0.00012 |
78
+ | 50 | 1.9287 | 1.929 | 0.254 | None | 0.00032 | 0.00015 |
79
+ | 250 | 1.7628 | 1.763 | 0.291 | None | 0.00061 | 0.00028 |
80
+ | 500 | 0.8965 | 0.896 | 0.594 | None | 0.00061 | 0.00029 |
81
+ | 1000 | 0.4965 | 0.496 | 0.764 | None | 0.00052 | 0.00023 |
82
+ | 1500 | 0.3130 | 0.313 | 0.847 | None | 0.00051 | 0.00025 |
83
+ | 2000 | 0.2039 | 0.204 | 0.899 | None | 0.00054 | 0.00029 |
84
+ | 2500 | 0.1431 | 0.143 | 0.928 | None | 0.00055 | 0.00088 |
85
+ | 3000 | 0.1127 | 0.113 | 0.943 | None | 0.00056 | 0.00028 |
86
+ | 3500 | 0.0843 | 0.084 | 0.956 | None | 0.00049 | 0.00027 |
87
+ | 4000 | 0.0750 | 0.075 | 0.961 | None | 0.00048 | 0.00025 |
88
+ | 1 | 1.9230 | 1.923 | 0.260 | None | 0.00025 | 0.00012 |
89
+ | 50 | 1.9287 | 1.929 | 0.254 | None | 0.00032 | 0.00015 |
90
+ | 250 | 1.7628 | 1.763 | 0.291 | None | 0.00061 | 0.00028 |
91
+ | 500 | 0.8965 | 0.896 | 0.594 | None | 0.00061 | 0.00029 |
92
+ | 1000 | 0.4965 | 0.496 | 0.764 | None | 0.00052 | 0.00023 |
93
+ | 1500 | 0.3130 | 0.313 | 0.847 | None | 0.00051 | 0.00025 |
94
+ | 2000 | 0.2039 | 0.204 | 0.899 | None | 0.00054 | 0.00029 |
95
+ | 2500 | 0.1431 | 0.143 | 0.928 | None | 0.00055 | 0.00088 |
96
+ | 3000 | 0.1127 | 0.113 | 0.943 | None | 0.00056 | 0.00028 |
97
+ | 3500 | 0.0843 | 0.084 | 0.956 | None | 0.00049 | 0.00027 |
98
+ | 4000 | 0.0750 | 0.075 | 0.961 | None | 0.00048 | 0.00025 |
99
+ | 1 | 1.9230 | 1.923 | 0.260 | None | 0.00025 | 0.00012 |
100
+ | 50 | 1.9287 | 1.929 | 0.254 | None | 0.00032 | 0.00015 |
101
+ | 250 | 1.7628 | 1.763 | 0.291 | None | 0.00061 | 0.00028 |
102
+ | 500 | 0.8965 | 0.896 | 0.594 | None | 0.00061 | 0.00029 |
103
+ | 1000 | 0.4965 | 0.496 | 0.764 | None | 0.00052 | 0.00023 |
104
+ | 1500 | 0.3130 | 0.313 | 0.847 | None | 0.00051 | 0.00025 |
105
+ | 2000 | 0.2039 | 0.204 | 0.899 | None | 0.00054 | 0.00029 |
106
+ | 2500 | 0.1431 | 0.143 | 0.928 | None | 0.00055 | 0.00088 |
107
+ | 3000 | 0.1127 | 0.113 | 0.943 | None | 0.00056 | 0.00028 |
108
+ | 3500 | 0.0843 | 0.084 | 0.956 | None | 0.00049 | 0.00027 |
109
+ | 4000 | 0.0750 | 0.075 | 0.961 | None | 0.00048 | 0.00025 |
110
+ | 1 | 1.9230 | 1.923 | 0.260 | None | 0.00025 | 0.00012 |
111
+ | 50 | 1.9287 | 1.929 | 0.254 | None | 0.00032 | 0.00015 |
112
+ | 250 | 1.7628 | 1.763 | 0.291 | None | 0.00061 | 0.00028 |
113
+ | 500 | 0.8965 | 0.896 | 0.594 | None | 0.00061 | 0.00029 |
114
+ | 1000 | 0.4965 | 0.496 | 0.764 | None | 0.00052 | 0.00023 |
115
+ | 1500 | 0.3130 | 0.313 | 0.847 | None | 0.00051 | 0.00025 |
116
+ | 2000 | 0.2039 | 0.204 | 0.899 | None | 0.00054 | 0.00029 |
117
+ | 2500 | 0.1431 | 0.143 | 0.928 | None | 0.00055 | 0.00088 |
118
+ | 3000 | 0.1127 | 0.113 | 0.943 | None | 0.00056 | 0.00028 |
119
+ | 3500 | 0.0843 | 0.084 | 0.956 | None | 0.00049 | 0.00027 |
120
+ | 4000 | 0.0750 | 0.075 | 0.961 | None | 0.00048 | 0.00025 |
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_gXA_o0_lamA8/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/qwen3/mse_gXA_o0_lamA8/s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f22d3212a08807aab84ebb3b3c39c55707d018f52b4286fee557e7302b7d3446
3
+ size 54819352
surface/qwen3/mse_gXA_o0_lamA8/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bcaeef8e9613d6dc9aacbd8016ca20df4a79256616e446d9b81f20820d1c8a3d
3
+ size 264325
surface/qwen3/mse_gXA_o0_lamA8/train.log ADDED
@@ -0,0 +1,174 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 0.3s] THE JOB: qarm_mse_gXA_o0_lamA8: the surface arm s10_qwen (gpt2 spelling by /workspace/qarm/dl/qwen3_tokenizer; A = the text written; sites = tokens) over the frozen eight of gXA; form mse; blocks all; 4000 steps, a close every 1000; lambda 2.0 (on the A rows 8.0); pure Adam lr 0.001; the compiled chain with recompute; card 30.5 GB with 30.4 GB free; host RSS 0.6 GB
2
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
3
+ [ 16.7s] 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
+ [ 17.3s] 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)]
5
+ [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
6
+ [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
7
+ [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
8
+ [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
9
+ [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
10
+ [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
11
+ [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
12
+ [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
13
+ [ 26.4s] the eight of gXA (close 8) loaded in training order; they freeze when the member attaches
14
+ [ 26.4s] 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)
15
+ [ 39.9s] 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
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
+ [ 57.8s] 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
+ [ 57.9s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 116), 4,920 sites
24
+ [ 57.9s] 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
+ [ 71.4s] 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
+ [ 76.6s] 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_gXA_o0_lamA8_stats.pt
28
+ [ 335.7s] 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
+ [ 335.7s] block 8: 0.733 -> 0.733 / 0.733; A 1.0000; mse 1.718 cos 0.366
30
+ [ 335.7s] block 12: 0.640 -> 0.640 / 0.640; A 1.0000; mse 1.718 cos 0.307
31
+ [ 335.7s] block 16: 0.532 -> 0.532 / 0.532; A 1.0000; mse 1.699 cos 0.250
32
+ [ 335.7s] block 18: 0.502 -> 0.502 / 0.502; A 1.0000; mse 1.682 cos 0.236
33
+ [ 335.7s] block 20: 0.486 -> 0.486 / 0.486; A 1.0000; mse 1.699 cos 0.233
34
+ [ 335.7s] block 24: 0.457 -> 0.457 / 0.457; A 1.0000; mse 1.858 cos 0.231
35
+ [ 335.7s] block 28: 0.407 -> 0.407 / 0.407; A 1.0000; mse 2.189 cos 0.194
36
+ [ 335.7s] block 31: 0.322 -> 0.322 / 0.322; A 1.0000; mse 5.319 cos 0.105
37
+ [ 335.7s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.644 / 0.540 / 0.492 / 0.465 / 0.419
38
+ [ 335.7s] 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
+ [ 335.7s] 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 11.7 GB
40
+ [stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
41
+ [ 344.4s] step 1/4000: task 1.9230 (mse 1.923 cos 0.260); quiet A 0.00025 off 0.00012; 1,503 sites; 8.65 s/step; the job: 5.7 min spent, about 576.3 min of steps left; card peak 14.2 GB
42
+ [ 514.0s] step 50/4000: task 1.9287 (mse 1.929 cos 0.254); quiet A 0.00032 off 0.00015; 1,508 sites; 3.57 s/step; the job: 8.6 min spent, about 234.8 min of steps left; card peak 14.3 GB
43
+ [ 687.3s] step 100/4000: task 1.9080 (mse 1.908 cos 0.258); quiet A 0.00040 off 0.00018; 1,506 sites; 3.52 s/step; the job: 11.5 min spent, about 228.5 min of steps left; card peak 14.3 GB
44
+ [ 860.4s] step 150/4000: task 1.8717 (mse 1.872 cos 0.266); quiet A 0.00047 off 0.00022; 1,471 sites; 3.50 s/step; the job: 14.3 min spent, about 224.4 min of steps left; card peak 14.3 GB
45
+ [ 1033.8s] step 200/4000: task 1.8211 (mse 1.821 cos 0.278); quiet A 0.00054 off 0.00024; 1,483 sites; 3.49 s/step; the job: 17.2 min spent, about 221.1 min of steps left; card peak 14.3 GB
46
+ [ 1207.1s] step 250/4000: task 1.7628 (mse 1.763 cos 0.291); quiet A 0.00061 off 0.00028; 1,485 sites; 3.49 s/step; the job: 20.1 min spent, about 217.8 min of steps left; card peak 14.3 GB
47
+ [ 1379.7s] step 300/4000: task 1.6841 (mse 1.684 cos 0.311); quiet A 0.00067 off 0.00030; 1,497 sites; 3.48 s/step; the job: 23.0 min spent, about 214.6 min of steps left; card peak 14.4 GB
48
+ [ 1553.1s] step 350/4000: task 1.5562 (mse 1.556 cos 0.347); quiet A 0.00071 off 0.00031; 1,487 sites; 3.48 s/step; the job: 25.9 min spent, about 211.6 min of steps left; card peak 14.4 GB
49
+ [ 1725.9s] step 400/4000: task 1.3246 (mse 1.325 cos 0.414); quiet A 0.00075 off 0.00036; 1,505 sites; 3.48 s/step; the job: 28.8 min spent, about 208.5 min of steps left; card peak 14.4 GB
50
+ [ 1898.4s] step 450/4000: task 1.0434 (mse 1.043 cos 0.527); quiet A 0.00067 off 0.00033; 1,593 sites; 3.47 s/step; the job: 31.6 min spent, about 205.5 min of steps left; card peak 14.4 GB
51
+ [ 2071.4s] step 500/4000: task 0.8965 (mse 0.896 cos 0.594); quiet A 0.00061 off 0.00029; 1,535 sites; 3.47 s/step; the job: 34.5 min spent, about 202.5 min of steps left; card peak 14.4 GB
52
+ [ 2244.8s] step 550/4000: task 0.8106 (mse 0.811 cos 0.632); quiet A 0.00058 off 0.00025; 1,433 sites; 3.47 s/step; the job: 37.4 min spent, about 199.6 min of steps left; card peak 14.4 GB
53
+ [ 2417.6s] step 600/4000: task 0.7504 (mse 0.750 cos 0.658); quiet A 0.00057 off 0.00027; 1,487 sites; 3.47 s/step; the job: 40.3 min spent, about 196.6 min of steps left; card peak 14.4 GB
54
+ [ 2590.8s] step 650/4000: task 0.6953 (mse 0.695 cos 0.682); quiet A 0.00053 off 0.00025; 1,558 sites; 3.47 s/step; the job: 43.2 min spent, about 193.7 min of steps left; card peak 14.4 GB
55
+ [ 2763.9s] step 700/4000: task 0.6514 (mse 0.651 cos 0.699); quiet A 0.00049 off 0.00027; 1,429 sites; 3.47 s/step; the job: 46.1 min spent, about 190.8 min of steps left; card peak 14.4 GB
56
+ [ 2936.7s] step 750/4000: task 0.6164 (mse 0.616 cos 0.713); quiet A 0.00049 off 0.00023; 1,468 sites; 3.47 s/step; the job: 48.9 min spent, about 187.8 min of steps left; card peak 14.4 GB
57
+ [ 3109.4s] step 800/4000: task 0.5854 (mse 0.585 cos 0.725); quiet A 0.00048 off 0.00022; 1,541 sites; 3.47 s/step; the job: 51.8 min spent, about 184.9 min of steps left; card peak 14.4 GB
58
+ [ 3282.4s] step 850/4000: task 0.5619 (mse 0.562 cos 0.736); quiet A 0.00047 off 0.00026; 1,440 sites; 3.47 s/step; the job: 54.7 min spent, about 182.0 min of steps left; card peak 14.4 GB
59
+ [ 3456.1s] step 900/4000: task 0.5321 (mse 0.532 cos 0.748); quiet A 0.00055 off 0.00025; 1,652 sites; 3.47 s/step; the job: 57.6 min spent, about 179.1 min of steps left; card peak 14.4 GB
60
+ [ 3629.2s] step 950/4000: task 0.5149 (mse 0.515 cos 0.755); quiet A 0.00053 off 0.00030; 1,550 sites; 3.47 s/step; the job: 60.5 min spent, about 176.2 min of steps left; card peak 14.4 GB
61
+ [ 3802.3s] step 1000/4000: task 0.4965 (mse 0.496 cos 0.764); quiet A 0.00052 off 0.00023; 1,493 sites; 3.47 s/step; the job: 63.4 min spent, about 173.3 min of steps left; card peak 14.4 GB
62
+ [ 3802.5s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXA_o0_lamA8_close1_s10_qwen.safetensors
63
+ [ 4007.7s] 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
+ [ 4007.7s] block 8: 0.733 -> 0.869 / 0.870; A 0.9982; mse 0.382 cos 0.810
65
+ [ 4007.7s] block 12: 0.640 -> 0.861 / 0.863; A 0.9967; mse 0.304 cos 0.850
66
+ [ 4007.7s] block 16: 0.532 -> 0.865 / 0.867; A 0.9964; mse 0.244 cos 0.881
67
+ [ 4007.7s] block 18: 0.502 -> 0.868 / 0.869; A 0.9975; mse 0.229 cos 0.889
68
+ [ 4007.7s] block 20: 0.486 -> 0.874 / 0.876; A 0.9971; mse 0.221 cos 0.895
69
+ [ 4007.7s] block 24: 0.457 -> 0.878 / 0.880; A 0.9943; mse 0.215 cos 0.903
70
+ [ 4007.7s] block 28: 0.407 -> 0.852 / 0.855; A 0.9956; mse 0.264 cos 0.883
71
+ [ 4007.7s] block 31: 0.322 -> 0.705 / 0.713; A 0.9815; mse 1.136 cos 0.835
72
+ [ 4007.7s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.859 / 0.860 / 0.868 / 0.874 / 0.846
73
+ [ 4007.7s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.705 against the bar .85; B2: the lowest A alignment 0.9815 against .995; fineweb share -0.00001 +- 0.00013 against +.012; partners' shares s1 +0.0001, s2 -0.0000, s3 +0.0002, s4 +0.0002, s5 +0.0001, s6 -0.0000, s7 -0.0002, s8 +0.0000
74
+ [ 4007.7s] gate mean 0.116 / widest 0.168; 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 14.4 GB; host RSS 14.1 GB
75
+ [ 4180.4s] step 1050/4000: task 0.4715 (mse 0.471 cos 0.773); quiet A 0.00049 off 0.00024; 1,510 sites; 3.45 s/step; the job: 69.7 min spent, about 170.4 min of steps left; card peak 14.4 GB
76
+ [ 4353.3s] step 1100/4000: task 0.4481 (mse 0.448 cos 0.783); quiet A 0.00047 off 0.00024; 1,491 sites; 3.46 s/step; the job: 72.6 min spent, about 167.5 min of steps left; card peak 14.4 GB
77
+ [ 4526.1s] step 1150/4000: task 0.4357 (mse 0.436 cos 0.789); quiet A 0.00048 off 0.00030; 1,475 sites; 3.46 s/step; the job: 75.4 min spent, about 164.6 min of steps left; card peak 14.4 GB
78
+ [ 4699.4s] step 1200/4000: task 0.4168 (mse 0.417 cos 0.798); quiet A 0.00053 off 0.00087; 1,492 sites; 3.46 s/step; the job: 78.3 min spent, about 161.7 min of steps left; card peak 14.4 GB
79
+ [ 4874.4s] step 1250/4000: task 0.3953 (mse 0.395 cos 0.808); quiet A 0.00050 off 0.00023; 1,473 sites; 3.47 s/step; the job: 81.2 min spent, about 158.9 min of steps left; card peak 14.4 GB
80
+ [ 5047.2s] step 1300/4000: task 0.3784 (mse 0.378 cos 0.816); quiet A 0.00049 off 0.00023; 1,475 sites; 3.47 s/step; the job: 84.1 min spent, about 156.0 min of steps left; card peak 14.4 GB
81
+ [ 5220.0s] step 1350/4000: task 0.3607 (mse 0.361 cos 0.824); quiet A 0.00050 off 0.00025; 1,476 sites; 3.46 s/step; the job: 87.0 min spent, about 153.1 min of steps left; card peak 14.4 GB
82
+ [ 5392.5s] step 1400/4000: task 0.3430 (mse 0.343 cos 0.833); quiet A 0.00052 off 0.00024; 1,519 sites; 3.46 s/step; the job: 89.9 min spent, about 150.2 min of steps left; card peak 14.4 GB
83
+ [ 5565.0s] step 1450/4000: task 0.3282 (mse 0.328 cos 0.840); quiet A 0.00050 off 0.00041; 1,500 sites; 3.46 s/step; the job: 92.7 min spent, about 147.2 min of steps left; card peak 14.4 GB
84
+ [ 5738.4s] step 1500/4000: task 0.3130 (mse 0.313 cos 0.847); quiet A 0.00051 off 0.00025; 1,442 sites; 3.46 s/step; the job: 95.6 min spent, about 144.4 min of steps left; card peak 14.4 GB
85
+ [ 5911.1s] step 1550/4000: task 0.2979 (mse 0.298 cos 0.855); quiet A 0.00052 off 0.00027; 1,466 sites; 3.46 s/step; the job: 98.5 min spent, about 141.5 min of steps left; card peak 14.4 GB
86
+ [ 6083.7s] step 1600/4000: task 0.2836 (mse 0.284 cos 0.861); quiet A 0.00053 off 0.00026; 1,480 sites; 3.46 s/step; the job: 101.4 min spent, about 138.6 min of steps left; card peak 14.4 GB
87
+ [ 6255.9s] step 1650/4000: task 0.2711 (mse 0.271 cos 0.867); quiet A 0.00054 off 0.00025; 1,535 sites; 3.46 s/step; the job: 104.3 min spent, about 135.7 min of steps left; card peak 14.4 GB
88
+ [ 6428.6s] step 1700/4000: task 0.2632 (mse 0.263 cos 0.871); quiet A 0.00057 off 0.00027; 1,499 sites; 3.46 s/step; the job: 107.1 min spent, about 132.8 min of steps left; card peak 14.4 GB
89
+ [ 6601.7s] step 1750/4000: task 0.2460 (mse 0.246 cos 0.879); quiet A 0.00055 off 0.00026; 1,463 sites; 3.46 s/step; the job: 110.0 min spent, about 129.9 min of steps left; card peak 14.4 GB
90
+ [ 6774.8s] step 1800/4000: task 0.2375 (mse 0.237 cos 0.883); quiet A 0.00055 off 0.00033; 1,531 sites; 3.46 s/step; the job: 112.9 min spent, about 127.0 min of steps left; card peak 14.4 GB
91
+ [ 6948.0s] step 1850/4000: task 0.2270 (mse 0.227 cos 0.888); quiet A 0.00055 off 0.00027; 1,498 sites; 3.46 s/step; the job: 115.8 min spent, about 124.1 min of steps left; card peak 14.4 GB
92
+ [ 7121.0s] step 1900/4000: task 0.2223 (mse 0.222 cos 0.891); quiet A 0.00056 off 0.00029; 1,601 sites; 3.46 s/step; the job: 118.7 min spent, about 121.2 min of steps left; card peak 14.4 GB
93
+ [ 7293.8s] step 1950/4000: task 0.2146 (mse 0.215 cos 0.894); quiet A 0.00056 off 0.00030; 1,500 sites; 3.46 s/step; the job: 121.6 min spent, about 118.3 min of steps left; card peak 14.4 GB
94
+ [ 7467.0s] step 2000/4000: task 0.2039 (mse 0.204 cos 0.899); quiet A 0.00054 off 0.00029; 1,501 sites; 3.46 s/step; the job: 124.5 min spent, about 115.4 min of steps left; card peak 14.4 GB
95
+ [ 7467.2s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXA_o0_lamA8_close2_s10_qwen.safetensors
96
+ [ 7672.5s] 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
+ [ 7672.5s] block 8: 0.733 -> 0.929 / 0.930; A 0.9977; mse 0.149 cos 0.924
98
+ [ 7672.5s] block 12: 0.640 -> 0.924 / 0.925; A 0.9977; mse 0.124 cos 0.939
99
+ [ 7672.5s] block 16: 0.532 -> 0.932 / 0.932; A 0.9970; mse 0.103 cos 0.950
100
+ [ 7672.5s] block 18: 0.502 -> 0.936 / 0.936; A 0.9978; mse 0.096 cos 0.954
101
+ [ 7672.5s] block 20: 0.486 -> 0.938 / 0.939; A 0.9973; mse 0.093 cos 0.956
102
+ [ 7672.5s] block 24: 0.457 -> 0.940 / 0.941; A 0.9974; mse 0.092 cos 0.959
103
+ [ 7672.5s] block 28: 0.407 -> 0.925 / 0.927; A 0.9948; mse 0.117 cos 0.949
104
+ [ 7672.5s] block 31: 0.322 -> 0.823 / 0.830; A 0.9805; mse 0.621 cos 0.923
105
+ [ 7672.5s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.925 / 0.929 / 0.935 / 0.938 / 0.922
106
+ [ 7672.5s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.823 against the bar .85; B2: the lowest A alignment 0.9805 against .995; fineweb share +0.00035 +- 0.00029 against +.012; partners' shares s1 +0.0001, s2 +0.0003, s3 +0.0003, s4 +0.0005, s5 +0.0001, s6 +0.0002, s7 -0.0002, s8 +0.0000
107
+ [ 7672.5s] gate mean 0.136 / widest 0.275; 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 14.4 GB; host RSS 14.1 GB
108
+ [ 7845.3s] step 2050/4000: task 0.1953 (mse 0.195 cos 0.903); quiet A 0.00055 off 0.00027; 1,509 sites; 3.45 s/step; the job: 130.8 min spent, about 112.5 min of steps left; card peak 14.4 GB
109
+ [ 8018.3s] step 2100/4000: task 0.1935 (mse 0.193 cos 0.904); quiet A 0.00059 off 0.00040; 1,504 sites; 3.46 s/step; the job: 133.6 min spent, about 109.7 min of steps left; card peak 14.4 GB
110
+ [ 8191.3s] step 2150/4000: task 0.1820 (mse 0.182 cos 0.909); quiet A 0.00055 off 0.00026; 1,495 sites; 3.46 s/step; the job: 136.5 min spent, about 106.8 min of steps left; card peak 14.4 GB
111
+ [ 8364.4s] step 2200/4000: task 0.1755 (mse 0.175 cos 0.912); quiet A 0.00055 off 0.00028; 1,510 sites; 3.46 s/step; the job: 139.4 min spent, about 103.9 min of steps left; card peak 14.4 GB
112
+ [ 8537.5s] step 2250/4000: task 0.1726 (mse 0.173 cos 0.914); quiet A 0.00055 off 0.00027; 1,547 sites; 3.46 s/step; the job: 142.3 min spent, about 101.0 min of steps left; card peak 14.4 GB
113
+ [ 8709.7s] step 2300/4000: task 0.1656 (mse 0.166 cos 0.917); quiet A 0.00053 off 0.00026; 1,479 sites; 3.46 s/step; the job: 145.2 min spent, about 98.1 min of steps left; card peak 14.4 GB
114
+ [ 8882.4s] step 2350/4000: task 0.1633 (mse 0.163 cos 0.919); quiet A 0.00054 off 0.00029; 1,529 sites; 3.46 s/step; the job: 148.0 min spent, about 95.2 min of steps left; card peak 14.4 GB
115
+ [ 9054.7s] step 2400/4000: task 0.1568 (mse 0.157 cos 0.921); quiet A 0.00055 off 0.00029; 1,459 sites; 3.46 s/step; the job: 150.9 min spent, about 92.3 min of steps left; card peak 14.4 GB
116
+ [ 9227.7s] step 2450/4000: task 0.1502 (mse 0.150 cos 0.924); quiet A 0.00055 off 0.00029; 1,471 sites; 3.46 s/step; the job: 153.8 min spent, about 89.4 min of steps left; card peak 14.4 GB
117
+ [ 9401.1s] step 2500/4000: task 0.1431 (mse 0.143 cos 0.928); quiet A 0.00055 off 0.00088; 1,553 sites; 3.46 s/step; the job: 156.7 min spent, about 86.5 min of steps left; card peak 14.4 GB
118
+ [ 9573.4s] step 2550/4000: task 0.1344 (mse 0.134 cos 0.932); quiet A 0.00055 off 0.00029; 1,506 sites; 3.46 s/step; the job: 159.6 min spent, about 83.7 min of steps left; card peak 14.4 GB
119
+ [ 9745.5s] step 2600/4000: task 0.1378 (mse 0.138 cos 0.931); quiet A 0.00053 off 0.00026; 1,443 sites; 3.45 s/step; the job: 162.4 min spent, about 80.8 min of steps left; card peak 14.4 GB
120
+ [ 9919.2s] step 2650/4000: task 0.1301 (mse 0.130 cos 0.934); quiet A 0.00051 off 0.00026; 1,447 sites; 3.46 s/step; the job: 165.3 min spent, about 77.9 min of steps left; card peak 14.4 GB
121
+ [10092.4s] step 2700/4000: task 0.1256 (mse 0.126 cos 0.936); quiet A 0.00051 off 0.00026; 1,511 sites; 3.46 s/step; the job: 168.2 min spent, about 75.0 min of steps left; card peak 14.4 GB
122
+ [10265.8s] step 2750/4000: task 0.1222 (mse 0.122 cos 0.938); quiet A 0.00051 off 0.00031; 1,482 sites; 3.46 s/step; the job: 171.1 min spent, about 72.1 min of steps left; card peak 14.4 GB
123
+ [10439.0s] step 2800/4000: task 0.1229 (mse 0.123 cos 0.938); quiet A 0.00053 off 0.00025; 1,489 sites; 3.46 s/step; the job: 174.0 min spent, about 69.2 min of steps left; card peak 14.4 GB
124
+ [10612.7s] step 2850/4000: task 0.1161 (mse 0.116 cos 0.941); quiet A 0.00051 off 0.00031; 1,523 sites; 3.46 s/step; the job: 176.9 min spent, about 66.4 min of steps left; card peak 14.4 GB
125
+ [10786.2s] step 2900/4000: task 0.1267 (mse 0.127 cos 0.938); quiet A 0.00106 off 0.00032; 1,518 sites; 3.46 s/step; the job: 179.8 min spent, about 63.5 min of steps left; card peak 14.5 GB
126
+ [10960.0s] step 2950/4000: task 0.1199 (mse 0.120 cos 0.939); quiet A 0.00069 off 0.00050; 1,497 sites; 3.46 s/step; the job: 182.7 min spent, about 60.6 min of steps left; card peak 14.5 GB
127
+ [11133.1s] step 3000/4000: task 0.1127 (mse 0.113 cos 0.943); quiet A 0.00056 off 0.00028; 1,480 sites; 3.46 s/step; the job: 185.6 min spent, about 57.7 min of steps left; card peak 14.5 GB
128
+ [11133.3s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXA_o0_lamA8_close3_s10_qwen.safetensors
129
+ [11338.4s] 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
+ [11338.4s] block 8: 0.733 -> 0.944 / 0.945; A 0.9989; mse 0.098 cos 0.950
131
+ [11338.4s] block 12: 0.640 -> 0.941 / 0.942; A 0.9985; mse 0.084 cos 0.959
132
+ [11338.4s] block 16: 0.532 -> 0.949 / 0.950; A 0.9961; mse 0.071 cos 0.966
133
+ [11338.4s] block 18: 0.502 -> 0.954 / 0.955; A 0.9985; mse 0.066 cos 0.968
134
+ [11338.4s] block 20: 0.486 -> 0.955 / 0.956; A 0.9983; mse 0.065 cos 0.970
135
+ [11338.4s] block 24: 0.457 -> 0.954 / 0.954; A 0.9983; mse 0.063 cos 0.972
136
+ [11338.4s] block 28: 0.407 -> 0.945 / 0.945; A 0.9965; mse 0.082 cos 0.964
137
+ [11338.4s] block 31: 0.322 -> 0.864 / 0.868; A 0.9847; mse 0.432 cos 0.946
138
+ [11338.4s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.943 / 0.947 / 0.951 / 0.955 / 0.942
139
+ [11338.4s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.864 against the bar .85; B2: the lowest A alignment 0.9847 against .995; fineweb share -0.00007 +- 0.00020 against +.012; partners' shares s1 +0.0001, s2 +0.0000, s3 +0.0002, s4 +0.0003, s5 +0.0001, s6 +0.0001, s7 +0.0001, s8 +0.0000
140
+ [11338.4s] gate mean 0.151 / widest 0.382; 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 14.5 GB; host RSS 14.1 GB
141
+ [11512.0s] step 3050/4000: task 0.1068 (mse 0.107 cos 0.945); quiet A 0.00054 off 0.00037; 1,458 sites; 3.47 s/step; the job: 191.9 min spent, about 54.8 min of steps left; card peak 14.5 GB
142
+ [11684.8s] step 3100/4000: task 0.1059 (mse 0.106 cos 0.946); quiet A 0.00053 off 0.00058; 1,561 sites; 3.46 s/step; the job: 194.7 min spent, about 51.9 min of steps left; card peak 14.5 GB
143
+ [11858.3s] step 3150/4000: task 0.1024 (mse 0.102 cos 0.948); quiet A 0.00052 off 0.00026; 1,440 sites; 3.47 s/step; the job: 197.6 min spent, about 49.0 min of steps left; card peak 14.5 GB
144
+ [12031.8s] step 3200/4000: task 0.0986 (mse 0.099 cos 0.949); quiet A 0.00053 off 0.00025; 1,519 sites; 3.47 s/step; the job: 200.5 min spent, about 46.2 min of steps left; card peak 14.5 GB
145
+ [12205.5s] step 3250/4000: task 0.0949 (mse 0.095 cos 0.951); quiet A 0.00049 off 0.00032; 1,475 sites; 3.47 s/step; the job: 203.4 min spent, about 43.3 min of steps left; card peak 14.5 GB
146
+ [12378.6s] step 3300/4000: task 0.0980 (mse 0.098 cos 0.950); quiet A 0.00053 off 0.00231; 1,474 sites; 3.47 s/step; the job: 206.3 min spent, about 40.4 min of steps left; card peak 14.5 GB
147
+ [12551.8s] step 3350/4000: task 0.0999 (mse 0.100 cos 0.949); quiet A 0.00051 off 0.00029; 1,541 sites; 3.47 s/step; the job: 209.2 min spent, about 37.5 min of steps left; card peak 14.5 GB
148
+ [12724.4s] step 3400/4000: task 0.0899 (mse 0.090 cos 0.954); quiet A 0.00050 off 0.00025; 1,520 sites; 3.46 s/step; the job: 212.1 min spent, about 34.6 min of steps left; card peak 14.5 GB
149
+ [12897.8s] step 3450/4000: task 0.0885 (mse 0.089 cos 0.954); quiet A 0.00049 off 0.00026; 1,454 sites; 3.47 s/step; the job: 215.0 min spent, about 31.7 min of steps left; card peak 14.5 GB
150
+ [13070.4s] step 3500/4000: task 0.0843 (mse 0.084 cos 0.956); quiet A 0.00049 off 0.00027; 1,505 sites; 3.46 s/step; the job: 217.8 min spent, about 28.9 min of steps left; card peak 14.5 GB
151
+ [13243.2s] step 3550/4000: task 0.0860 (mse 0.086 cos 0.956); quiet A 0.00050 off 0.00093; 1,525 sites; 3.46 s/step; the job: 220.7 min spent, about 26.0 min of steps left; card peak 14.5 GB
152
+ [13416.4s] step 3600/4000: task 0.0839 (mse 0.084 cos 0.957); quiet A 0.00049 off 0.00031; 1,519 sites; 3.46 s/step; the job: 223.6 min spent, about 23.1 min of steps left; card peak 14.5 GB
153
+ [13589.8s] step 3650/4000: task 0.0846 (mse 0.085 cos 0.957); quiet A 0.00050 off 0.00026; 1,472 sites; 3.46 s/step; the job: 226.5 min spent, about 20.2 min of steps left; card peak 14.5 GB
154
+ [13762.5s] step 3700/4000: task 0.0795 (mse 0.080 cos 0.959); quiet A 0.00048 off 0.00026; 1,463 sites; 3.46 s/step; the job: 229.4 min spent, about 17.3 min of steps left; card peak 14.5 GB
155
+ [13935.1s] step 3750/4000: task 0.0788 (mse 0.079 cos 0.959); quiet A 0.00048 off 0.00033; 1,570 sites; 3.46 s/step; the job: 232.3 min spent, about 14.4 min of steps left; card peak 14.5 GB
156
+ [14108.5s] step 3800/4000: task 0.0786 (mse 0.079 cos 0.959); quiet A 0.00048 off 0.00025; 1,532 sites; 3.46 s/step; the job: 235.1 min spent, about 11.5 min of steps left; card peak 14.5 GB
157
+ [14281.1s] step 3850/4000: task 0.0765 (mse 0.076 cos 0.960); quiet A 0.00050 off 0.00027; 1,507 sites; 3.46 s/step; the job: 238.0 min spent, about 8.7 min of steps left; card peak 14.5 GB
158
+ [14453.8s] step 3900/4000: task 0.0754 (mse 0.075 cos 0.961); quiet A 0.00050 off 0.00025; 1,435 sites; 3.46 s/step; the job: 240.9 min spent, about 5.8 min of steps left; card peak 14.5 GB
159
+ [14627.6s] step 3950/4000: task 0.0744 (mse 0.074 cos 0.961); quiet A 0.00048 off 0.00028; 1,500 sites; 3.46 s/step; the job: 243.8 min spent, about 2.9 min of steps left; card peak 14.5 GB
160
+ [14800.8s] step 4000/4000: task 0.0750 (mse 0.075 cos 0.961); quiet A 0.00048 off 0.00025; 1,494 sites; 3.46 s/step; the job: 246.7 min spent, about 0.0 min of steps left; card peak 14.5 GB
161
+ [14801.1s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXA_o0_lamA8_close4_s10_qwen.safetensors
162
+ [15006.7s] 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
+ [15006.7s] block 8: 0.733 -> 0.952 / 0.952; A 0.9991; mse 0.076 cos 0.961
164
+ [15006.7s] block 12: 0.640 -> 0.951 / 0.952; A 0.9986; mse 0.064 cos 0.969
165
+ [15006.7s] block 16: 0.532 -> 0.961 / 0.960; A 0.9981; mse 0.054 cos 0.974
166
+ [15006.7s] block 18: 0.502 -> 0.963 / 0.963; A 0.9986; mse 0.051 cos 0.976
167
+ [15006.7s] block 20: 0.486 -> 0.964 / 0.963; A 0.9968; mse 0.050 cos 0.976
168
+ [15006.7s] block 24: 0.457 -> 0.963 / 0.963; A 0.9985; mse 0.050 cos 0.978
169
+ [15006.7s] block 28: 0.407 -> 0.954 / 0.956; A 0.9972; mse 0.066 cos 0.971
170
+ [15006.7s] block 31: 0.322 -> 0.883 / 0.884; A 0.9899; mse 0.362 cos 0.955
171
+ [15006.7s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.952 / 0.957 / 0.961 / 0.964 / 0.953
172
+ [15006.7s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.883 against the bar .85; B2: the lowest A alignment 0.9899 against .995; fineweb share -0.00005 +- 0.00018 against +.012; partners' shares s1 -0.0000, s2 +0.0001, s3 +0.0002, s4 +0.0004, s5 +0.0001, s6 +0.0001, s7 +0.0001, s8 +0.0000
173
+ [15006.7s] gate mean 0.164 / widest 0.475; 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 14.5 GB; host RSS 14.1 GB
174
+ [15006.7s] DONE qarm_mse_gXA_o0_lamA8: 4000 steps in 230.8 min of steps (3.46 s/step); 4 closes; 15007 s in all; card peak 14.5 GB; host RSS 14.1 GB; compiled chain: 15 cache entries (limit 256), 0 limit warnings