Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
95682eb
·
verified ·
1 Parent(s): f34d5a7

surface arm qwen3/mse_gXB_o1: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/qwen3/mse_gXB_o1/closes/qarm_mse_gXB_o1_close1_s10_qwen.safetensors +3 -0
  2. surface/qwen3/mse_gXB_o1/closes/qarm_mse_gXB_o1_close2_s10_qwen.safetensors +3 -0
  3. surface/qwen3/mse_gXB_o1/closes/qarm_mse_gXB_o1_close3_s10_qwen.safetensors +3 -0
  4. surface/qwen3/mse_gXB_o1/closes/qarm_mse_gXB_o1_close4_s10_qwen.safetensors +3 -0
  5. surface/qwen3/mse_gXB_o1/manifest.json +20 -0
  6. surface/qwen3/mse_gXB_o1/readout.md +127 -0
  7. surface/qwen3/mse_gXB_o1/results.json +0 -0
  8. surface/qwen3/mse_gXB_o1/s10_qwen.safetensors +3 -0
  9. surface/qwen3/mse_gXB_o1/stats.pt +3 -0
  10. surface/qwen3/mse_gXB_o1/train.log +174 -0
surface/qwen3/mse_gXB_o1/closes/qarm_mse_gXB_o1_close1_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:84b88c0c6bff9fe43b51c5d156452bba49ef1b797cafe2f619d8fe2b9aec4746
3
+ size 54819328
surface/qwen3/mse_gXB_o1/closes/qarm_mse_gXB_o1_close2_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:874ff3338605d6e34dd2434220a2ffe70bdee5b23f2231bed5023c320e1d264f
3
+ size 54819328
surface/qwen3/mse_gXB_o1/closes/qarm_mse_gXB_o1_close3_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e103e74d9b86d3fd55cf8151c712becc2914df53f4f7f90fdc1c26a35786e733
3
+ size 54819328
surface/qwen3/mse_gXB_o1/closes/qarm_mse_gXB_o1_close4_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:67be7d15bb81c20e3eac9909292a90c7c3aa10d08cb738616a65e8b5e718b137
3
+ size 54819328
surface/qwen3/mse_gXB_o1/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "qarm_mse_gXB_o1",
3
+ "run": "mse_gXB_o1",
4
+ "convention": "qwen3",
5
+ "member": "s10_qwen",
6
+ "final_anchor": "qarm_mse_gXB_o1_close4_s10_qwen.safetensors",
7
+ "closes": [
8
+ "qarm_mse_gXB_o1_close1_s10_qwen.safetensors",
9
+ "qarm_mse_gXB_o1_close2_s10_qwen.safetensors",
10
+ "qarm_mse_gXB_o1_close3_s10_qwen.safetensors",
11
+ "qarm_mse_gXB_o1_close4_s10_qwen.safetensors"
12
+ ],
13
+ "finished": true,
14
+ "form": "mse",
15
+ "base": "gXB",
16
+ "steps": 4000,
17
+ "trunk_step": 245674,
18
+ "shipped": "2026-10-08 08:59 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_gXB_o1/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_gXB_o1: per-site MSE; over the frozen eight of gXB (seed B); blocks 0-31; 4000 of 4000 steps; seed offset 1
8
+
9
+ ## The verdicts (the final close of each run)
10
+
11
+ ### qarm_mse_gXB_o1
12
+
13
+ - B1 the gap: MET over every served block >= 12 (the lowest 0.915 at block 31 on draw 1, 0.909 at block 31 on draw 2; bar 0.85)
14
+ - B2 silence, whitened: NOT MET (the lowest A alignment 0.9847 at block 31 on draw 1, 0.9812 on draw 2; bar 0.995)
15
+ - B2 silence, in bits: MET (web +0.00032 bpb; the worst partner s3_rules +0.00025; bar +0.012)
16
+
17
+ ## The closes
18
+
19
+ ### qarm_mse_gXB_o1
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 | 0.343 / 0.343 | 0.640 / 0.477 / 0.401 | 1.0000 / 1.0000 | +0.00000 | 0.047 / 0.047 |
24
+ | 1 | 1000 | 0.790 / 0.775 | 0.862 / 0.876 / 0.859 | 0.9705 / 0.9573 | +0.00042 | 0.117 / 0.169 |
25
+ | 2 | 2000 | 0.882 / 0.872 | 0.924 / 0.939 / 0.932 | 0.9739 / 0.9604 | +0.00083 | 0.137 / 0.268 |
26
+ | 3 | 3000 | 0.906 / 0.903 | 0.944 / 0.958 / 0.953 | 0.9821 / 0.9715 | +0.00031 | 0.156 / 0.385 |
27
+ | 4 | 4000 | 0.915 / 0.909 | 0.953 / 0.963 / 0.957 | 0.9847 / 0.9812 | +0.00032 | 0.169 / 0.485 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### qarm_mse_gXB_o1, 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 | 0.897 | 0.9997 | 0.402 | 0.759 |
36
+ | 4 | 0.788 | 0.943 | 0.942 | 0.9978 | 0.096 | 0.948 |
37
+ | 8 | 0.731 | 0.953 | 0.952 | 0.9971 | 0.075 | 0.962 |
38
+ | 12 | 0.640 | 0.953 | 0.953 | 0.9969 | 0.064 | 0.969 |
39
+ | 16 | 0.526 | 0.960 | 0.957 | 0.9958 | 0.054 | 0.974 |
40
+ | 18 | 0.493 | 0.964 | 0.963 | 0.9969 | 0.051 | 0.976 |
41
+ | 20 | 0.477 | 0.963 | 0.963 | 0.9936 | 0.050 | 0.977 |
42
+ | 22 | 0.463 | 0.967 | 0.966 | 0.9968 | 0.049 | 0.979 |
43
+ | 24 | 0.450 | 0.963 | 0.963 | 0.9949 | 0.049 | 0.978 |
44
+ | 26 | 0.423 | 0.964 | 0.963 | 0.9960 | 0.051 | 0.977 |
45
+ | 28 | 0.401 | 0.957 | 0.956 | 0.9945 | 0.059 | 0.973 |
46
+ | 29 | 0.393 | 0.953 | 0.951 | 0.9950 | 0.065 | 0.971 |
47
+ | 30 | 0.377 | 0.944 | 0.941 | 0.9927 | 0.089 | 0.967 |
48
+ | 31 | 0.343 | 0.915 | 0.914 | 0.9847 | 0.127 | 0.961 |
49
+
50
+ ### qarm_mse_gXB_o1, 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 | 0.887 | 0.9996 | 0.414 | 0.753 |
55
+ | 4 | 0.794 | 0.941 | 0.940 | 0.9965 | 0.098 | 0.947 |
56
+ | 8 | 0.739 | 0.952 | 0.952 | 0.9969 | 0.076 | 0.961 |
57
+ | 12 | 0.644 | 0.951 | 0.950 | 0.9960 | 0.065 | 0.968 |
58
+ | 16 | 0.533 | 0.955 | 0.954 | 0.9954 | 0.055 | 0.973 |
59
+ | 18 | 0.499 | 0.957 | 0.954 | 0.9943 | 0.053 | 0.974 |
60
+ | 20 | 0.482 | 0.959 | 0.957 | 0.9936 | 0.052 | 0.975 |
61
+ | 22 | 0.474 | 0.962 | 0.960 | 0.9960 | 0.052 | 0.977 |
62
+ | 24 | 0.456 | 0.960 | 0.959 | 0.9961 | 0.052 | 0.976 |
63
+ | 26 | 0.430 | 0.959 | 0.957 | 0.9951 | 0.055 | 0.975 |
64
+ | 28 | 0.411 | 0.952 | 0.951 | 0.9941 | 0.064 | 0.971 |
65
+ | 29 | 0.405 | 0.948 | 0.946 | 0.9938 | 0.070 | 0.968 |
66
+ | 30 | 0.388 | 0.939 | 0.935 | 0.9870 | 0.095 | 0.965 |
67
+ | 31 | 0.343 | 0.909 | 0.909 | 0.9812 | 0.124 | 0.959 |
68
+
69
+ The member's share in bits per byte (all on minus the member masked): s1_perspective +0.00014, s2_concept +0.00024, s3_rules +0.00025, s4_arith +0.00019, s5_causal +0.00006, s6_tryfail +0.00009, s7_mixed +0.00019, s8_register +0.00005, web +0.00032
70
+
71
+ ## The training curves
72
+
73
+ ### qarm_mse_gXB_o1
74
+
75
+ | step | task loss | mse | cos | InfoNCE | quiet on A | quiet off-domain |
76
+ |---|---|---|---|---|---|---|
77
+ | 1 | 1.9375 | 1.938 | 0.238 | None | 0.00026 | 0.00016 |
78
+ | 50 | 1.9201 | 1.920 | 0.248 | None | 0.00032 | 0.00015 |
79
+ | 250 | 1.7441 | 1.744 | 0.290 | None | 0.00233 | 0.00027 |
80
+ | 500 | 0.8910 | 0.891 | 0.593 | None | 0.00127 | 0.00036 |
81
+ | 1000 | 0.4771 | 0.477 | 0.769 | None | 0.00096 | 0.00039 |
82
+ | 1500 | 0.3018 | 0.302 | 0.851 | None | 0.00103 | 0.00042 |
83
+ | 2000 | 0.2010 | 0.201 | 0.900 | None | 0.00120 | 0.00053 |
84
+ | 2500 | 0.1409 | 0.141 | 0.929 | None | 0.00131 | 0.00057 |
85
+ | 3000 | 0.1083 | 0.108 | 0.945 | None | 0.00103 | 0.00049 |
86
+ | 3500 | 0.0842 | 0.084 | 0.956 | None | 0.00089 | 0.00041 |
87
+ | 4000 | 0.0732 | 0.073 | 0.962 | None | 0.00083 | 0.00039 |
88
+ | 1 | 1.9375 | 1.938 | 0.238 | None | 0.00026 | 0.00016 |
89
+ | 50 | 1.9201 | 1.920 | 0.248 | None | 0.00032 | 0.00015 |
90
+ | 250 | 1.7441 | 1.744 | 0.290 | None | 0.00233 | 0.00027 |
91
+ | 500 | 0.8910 | 0.891 | 0.593 | None | 0.00127 | 0.00036 |
92
+ | 1000 | 0.4771 | 0.477 | 0.769 | None | 0.00096 | 0.00039 |
93
+ | 1500 | 0.3018 | 0.302 | 0.851 | None | 0.00103 | 0.00042 |
94
+ | 2000 | 0.2010 | 0.201 | 0.900 | None | 0.00120 | 0.00053 |
95
+ | 2500 | 0.1409 | 0.141 | 0.929 | None | 0.00131 | 0.00057 |
96
+ | 3000 | 0.1083 | 0.108 | 0.945 | None | 0.00103 | 0.00049 |
97
+ | 3500 | 0.0842 | 0.084 | 0.956 | None | 0.00089 | 0.00041 |
98
+ | 4000 | 0.0732 | 0.073 | 0.962 | None | 0.00083 | 0.00039 |
99
+ | 1 | 1.9375 | 1.938 | 0.238 | None | 0.00026 | 0.00016 |
100
+ | 50 | 1.9201 | 1.920 | 0.248 | None | 0.00032 | 0.00015 |
101
+ | 250 | 1.7441 | 1.744 | 0.290 | None | 0.00233 | 0.00027 |
102
+ | 500 | 0.8910 | 0.891 | 0.593 | None | 0.00127 | 0.00036 |
103
+ | 1000 | 0.4771 | 0.477 | 0.769 | None | 0.00096 | 0.00039 |
104
+ | 1500 | 0.3018 | 0.302 | 0.851 | None | 0.00103 | 0.00042 |
105
+ | 2000 | 0.2010 | 0.201 | 0.900 | None | 0.00120 | 0.00053 |
106
+ | 2500 | 0.1409 | 0.141 | 0.929 | None | 0.00131 | 0.00057 |
107
+ | 3000 | 0.1083 | 0.108 | 0.945 | None | 0.00103 | 0.00049 |
108
+ | 3500 | 0.0842 | 0.084 | 0.956 | None | 0.00089 | 0.00041 |
109
+ | 4000 | 0.0732 | 0.073 | 0.962 | None | 0.00083 | 0.00039 |
110
+ | 1 | 1.9375 | 1.938 | 0.238 | None | 0.00026 | 0.00016 |
111
+ | 50 | 1.9201 | 1.920 | 0.248 | None | 0.00032 | 0.00015 |
112
+ | 250 | 1.7441 | 1.744 | 0.290 | None | 0.00233 | 0.00027 |
113
+ | 500 | 0.8910 | 0.891 | 0.593 | None | 0.00127 | 0.00036 |
114
+ | 1000 | 0.4771 | 0.477 | 0.769 | None | 0.00096 | 0.00039 |
115
+ | 1500 | 0.3018 | 0.302 | 0.851 | None | 0.00103 | 0.00042 |
116
+ | 2000 | 0.2010 | 0.201 | 0.900 | None | 0.00120 | 0.00053 |
117
+ | 2500 | 0.1409 | 0.141 | 0.929 | None | 0.00131 | 0.00057 |
118
+ | 3000 | 0.1083 | 0.108 | 0.945 | None | 0.00103 | 0.00049 |
119
+ | 3500 | 0.0842 | 0.084 | 0.956 | None | 0.00089 | 0.00041 |
120
+ | 4000 | 0.0732 | 0.073 | 0.962 | None | 0.00083 | 0.00039 |
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_gXB_o1/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/qwen3/mse_gXB_o1/s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:67be7d15bb81c20e3eac9909292a90c7c3aa10d08cb738616a65e8b5e718b137
3
+ size 54819328
surface/qwen3/mse_gXB_o1/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cfe1ee99d7f614d7d9170978faa3b54f86d24d98cd1f4dd3ba6b49039b6f6382
3
+ size 264213
surface/qwen3/mse_gXB_o1/train.log ADDED
@@ -0,0 +1,174 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 0.3s] THE JOB: qarm_mse_gXB_o1: the surface arm s10_qwen (gpt2 spelling by /workspace/qarm/dl/qwen3_tokenizer; A = the text written; sites = tokens) over the frozen eight of gXB; form mse; blocks all; 4000 steps, a close every 1000; lambda 2.0; pure Adam lr 0.001; the compiled chain with recompute; card 30.5 GB with 30.4 GB free; host RSS 0.6 GB
2
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
3
+ [ 16.7s] 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
+ [ 17.2s] 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
+ [ 26.1s] the eight of gXB (close 8) loaded in training order; they freeze when the member attaches
14
+ [ 26.1s] 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
+ [ 39.7s] 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
+ [ 54.3s] 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
+ [ 54.4s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 116), 4,920 sites
24
+ [ 54.4s] 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
+ [ 67.9s] 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
+ [ 73.3s] standardization statistics from 12,013 teacher sites on 8 paired batches (a mean and a spread per block and dimension); sigma range at block 20 0.754..1.107; saved /workspace/qarm/out/anchors/qarm_mse_gXB_o1_stats.pt
28
+ [ 332.6s] 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
+ [ 332.6s] block 8: 0.731 -> 0.731 / 0.731; A 1.0000; mse 1.686 cos 0.373
30
+ [ 332.6s] block 12: 0.640 -> 0.640 / 0.640; A 1.0000; mse 1.708 cos 0.307
31
+ [ 332.6s] block 16: 0.526 -> 0.526 / 0.526; A 1.0000; mse 1.710 cos 0.241
32
+ [ 332.6s] block 18: 0.493 -> 0.493 / 0.493; A 1.0000; mse 1.700 cos 0.226
33
+ [ 332.6s] block 20: 0.477 -> 0.477 / 0.477; A 1.0000; mse 1.726 cos 0.217
34
+ [ 332.6s] block 24: 0.450 -> 0.450 / 0.450; A 1.0000; mse 1.876 cos 0.213
35
+ [ 332.6s] block 28: 0.401 -> 0.401 / 0.401; A 1.0000; mse 2.165 cos 0.173
36
+ [ 332.6s] block 31: 0.343 -> 0.343 / 0.343; A 1.0000; mse 3.582 cos 0.114
37
+ [ 332.6s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.644 / 0.533 / 0.482 / 0.456 / 0.411
38
+ [ 332.6s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.343 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
+ [ 332.6s] gate mean 0.047 / widest 0.047; compiled chain: 13 cache entries (limit 256), 0 limit warnings; docs drawn 274 (skipped {'test': 3, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 2.7 GB; host RSS 12.4 GB
40
+ [stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
41
+ [ 341.3s] step 1/4000: task 1.9375 (mse 1.938 cos 0.238); quiet A 0.00026 off 0.00016; 1,502 sites; 8.66 s/step; the job: 5.7 min spent, about 577.0 min of steps left; card peak 14.2 GB
42
+ [ 511.2s] step 50/4000: task 1.9201 (mse 1.920 cos 0.248); quiet A 0.00032 off 0.00015; 1,490 sites; 3.57 s/step; the job: 8.5 min spent, about 235.0 min of steps left; card peak 14.3 GB
43
+ [ 684.0s] step 100/4000: task 1.8791 (mse 1.879 cos 0.258); quiet A 0.00060 off 0.00024; 1,550 sites; 3.51 s/step; the job: 11.4 min spent, about 228.4 min of steps left; card peak 14.3 GB
44
+ [ 857.8s] step 150/4000: task 1.8398 (mse 1.840 cos 0.267); quiet A 0.00096 off 0.00041; 1,510 sites; 3.50 s/step; the job: 14.3 min spent, about 224.6 min of steps left; card peak 14.3 GB
45
+ [ 1030.6s] step 200/4000: task 1.7908 (mse 1.791 cos 0.279); quiet A 0.00132 off 0.00031; 1,507 sites; 3.49 s/step; the job: 17.2 min spent, about 221.0 min of steps left; card peak 14.3 GB
46
+ [ 1204.5s] step 250/4000: task 1.7441 (mse 1.744 cos 0.290); quiet A 0.00233 off 0.00027; 1,458 sites; 3.49 s/step; the job: 20.1 min spent, about 218.0 min of steps left; card peak 14.3 GB
47
+ [ 1377.3s] step 300/4000: task 1.6704 (mse 1.670 cos 0.306); quiet A 0.00215 off 0.00065; 1,479 sites; 3.48 s/step; the job: 23.0 min spent, about 214.7 min of steps left; card peak 14.3 GB
48
+ [ 1551.2s] step 350/4000: task 1.5378 (mse 1.538 cos 0.345); quiet A 0.00229 off 0.00080; 1,505 sites; 3.48 s/step; the job: 25.9 min spent, about 211.8 min of steps left; card peak 14.3 GB
49
+ [ 1724.8s] step 400/4000: task 1.3084 (mse 1.308 cos 0.415); quiet A 0.00240 off 0.00056; 1,505 sites; 3.48 s/step; the job: 28.7 min spent, about 208.8 min of steps left; card peak 14.3 GB
50
+ [ 1897.9s] step 450/4000: task 1.0406 (mse 1.041 cos 0.523); quiet A 0.00162 off 0.00041; 1,467 sites; 3.48 s/step; the job: 31.6 min spent, about 205.8 min of steps left; card peak 14.3 GB
51
+ [ 2071.6s] step 500/4000: task 0.8910 (mse 0.891 cos 0.593); quiet A 0.00127 off 0.00036; 1,486 sites; 3.48 s/step; the job: 34.5 min spent, about 202.9 min of steps left; card peak 14.3 GB
52
+ [ 2244.9s] step 550/4000: task 0.8018 (mse 0.802 cos 0.631); quiet A 0.00112 off 0.00040; 1,453 sites; 3.48 s/step; the job: 37.4 min spent, about 199.9 min of steps left; card peak 14.3 GB
53
+ [ 2417.5s] step 600/4000: task 0.7380 (mse 0.738 cos 0.660); quiet A 0.00109 off 0.00040; 1,473 sites; 3.47 s/step; the job: 40.3 min spent, about 196.9 min of steps left; card peak 14.3 GB
54
+ [ 2591.0s] step 650/4000: task 0.6865 (mse 0.686 cos 0.681); quiet A 0.00104 off 0.00039; 1,492 sites; 3.47 s/step; the job: 43.2 min spent, about 194.0 min of steps left; card peak 14.3 GB
55
+ [ 2764.0s] step 700/4000: task 0.6475 (mse 0.648 cos 0.698); quiet A 0.00100 off 0.00044; 1,497 sites; 3.47 s/step; the job: 46.1 min spent, about 191.0 min of steps left; card peak 14.4 GB
56
+ [ 2937.9s] step 750/4000: task 0.6072 (mse 0.607 cos 0.715); quiet A 0.00099 off 0.00033; 1,494 sites; 3.47 s/step; the job: 49.0 min spent, about 188.2 min of steps left; card peak 14.4 GB
57
+ [ 3111.5s] step 800/4000: task 0.5780 (mse 0.578 cos 0.726); quiet A 0.00095 off 0.00037; 1,579 sites; 3.47 s/step; the job: 51.9 min spent, about 185.3 min of steps left; card peak 14.4 GB
58
+ [ 3285.8s] step 850/4000: task 0.5477 (mse 0.548 cos 0.739); quiet A 0.00094 off 0.00034; 1,547 sites; 3.47 s/step; the job: 54.8 min spent, about 182.4 min of steps left; card peak 14.4 GB
59
+ [ 3459.8s] step 900/4000: task 0.5225 (mse 0.522 cos 0.750); quiet A 0.00095 off 0.00036; 1,447 sites; 3.47 s/step; the job: 57.7 min spent, about 179.5 min of steps left; card peak 14.4 GB
60
+ [ 3633.7s] step 950/4000: task 0.5006 (mse 0.501 cos 0.759); quiet A 0.00092 off 0.00038; 1,449 sites; 3.47 s/step; the job: 60.6 min spent, about 176.6 min of steps left; card peak 14.4 GB
61
+ [ 3807.7s] step 1000/4000: task 0.4771 (mse 0.477 cos 0.769); quiet A 0.00096 off 0.00039; 1,488 sites; 3.47 s/step; the job: 63.5 min spent, about 173.7 min of steps left; card peak 14.4 GB
62
+ [ 3807.9s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXB_o1_close1_s10_qwen.safetensors
63
+ [ 4013.8s] 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
+ [ 4013.8s] block 8: 0.731 -> 0.867 / 0.871; A 0.9917; mse 0.368 cos 0.816
65
+ [ 4013.8s] block 12: 0.640 -> 0.862 / 0.866; A 0.9889; mse 0.299 cos 0.853
66
+ [ 4013.8s] block 16: 0.526 -> 0.865 / 0.869; A 0.9894; mse 0.240 cos 0.884
67
+ [ 4013.8s] block 18: 0.493 -> 0.868 / 0.874; A 0.9886; mse 0.224 cos 0.892
68
+ [ 4013.8s] block 20: 0.477 -> 0.876 / 0.881; A 0.9896; mse 0.215 cos 0.898
69
+ [ 4013.8s] block 24: 0.450 -> 0.879 / 0.884; A 0.9872; mse 0.204 cos 0.907
70
+ [ 4013.8s] block 28: 0.401 -> 0.859 / 0.863; A 0.9855; mse 0.235 cos 0.890
71
+ [ 4013.8s] block 31: 0.343 -> 0.790 / 0.799; A 0.9705; mse 0.401 cos 0.857
72
+ [ 4013.8s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.855 / 0.861 / 0.871 / 0.877 / 0.854
73
+ [ 4013.8s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.790 against the bar .85; B2: the lowest A alignment 0.9705 against .995; fineweb share +0.00042 +- 0.00026 against +.012; partners' shares s1 +0.0002, s2 +0.0004, s3 +0.0002, s4 +0.0003, s5 +0.0002, s6 +0.0001, s7 +0.0005, s8 +0.0000
74
+ [ 4013.8s] gate mean 0.117 / widest 0.169; compiled chain: 15 cache entries (limit 256), 0 limit warnings; docs drawn 32,659 (skipped {'test': 51, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 14.4 GB; host RSS 13.6 GB
75
+ [ 4187.2s] step 1050/4000: task 0.4599 (mse 0.460 cos 0.777); quiet A 0.00094 off 0.00036; 1,496 sites; 3.47 s/step; the job: 69.8 min spent, about 170.8 min of steps left; card peak 14.4 GB
76
+ [ 4360.4s] step 1100/4000: task 0.4401 (mse 0.440 cos 0.786); quiet A 0.00094 off 0.00035; 1,521 sites; 3.47 s/step; the job: 72.7 min spent, about 167.9 min of steps left; card peak 14.4 GB
77
+ [ 4534.4s] step 1150/4000: task 0.4193 (mse 0.419 cos 0.795); quiet A 0.00093 off 0.00042; 1,500 sites; 3.47 s/step; the job: 75.6 min spent, about 165.0 min of steps left; card peak 14.4 GB
78
+ [ 4708.2s] step 1200/4000: task 0.4001 (mse 0.400 cos 0.803); quiet A 0.00100 off 0.00038; 1,558 sites; 3.47 s/step; the job: 78.5 min spent, about 162.1 min of steps left; card peak 14.4 GB
79
+ [ 4881.9s] step 1250/4000: task 0.3863 (mse 0.386 cos 0.810); quiet A 0.00095 off 0.00036; 1,475 sites; 3.47 s/step; the job: 81.4 min spent, about 159.2 min of steps left; card peak 14.4 GB
80
+ [ 5055.7s] step 1300/4000: task 0.3668 (mse 0.367 cos 0.820); quiet A 0.00096 off 0.00038; 1,482 sites; 3.47 s/step; the job: 84.3 min spent, about 156.4 min of steps left; card peak 14.4 GB
81
+ [ 5229.3s] step 1350/4000: task 0.3471 (mse 0.347 cos 0.829); quiet A 0.00103 off 0.00107; 1,508 sites; 3.47 s/step; the job: 87.2 min spent, about 153.5 min of steps left; card peak 14.4 GB
82
+ [ 5402.9s] step 1400/4000: task 0.3333 (mse 0.333 cos 0.836); quiet A 0.00095 off 0.00041; 1,436 sites; 3.47 s/step; the job: 90.0 min spent, about 150.6 min of steps left; card peak 14.4 GB
83
+ [ 5576.5s] step 1450/4000: task 0.3152 (mse 0.315 cos 0.845); quiet A 0.00101 off 0.00043; 1,499 sites; 3.47 s/step; the job: 92.9 min spent, about 147.7 min of steps left; card peak 14.4 GB
84
+ [ 5749.5s] step 1500/4000: task 0.3018 (mse 0.302 cos 0.851); quiet A 0.00103 off 0.00042; 1,463 sites; 3.47 s/step; the job: 95.8 min spent, about 144.7 min of steps left; card peak 14.4 GB
85
+ [ 5922.9s] step 1550/4000: task 0.2881 (mse 0.288 cos 0.858); quiet A 0.00111 off 0.00043; 1,494 sites; 3.47 s/step; the job: 98.7 min spent, about 141.8 min of steps left; card peak 14.4 GB
86
+ [ 6097.3s] step 1600/4000: task 0.2780 (mse 0.278 cos 0.863); quiet A 0.00114 off 0.00054; 1,469 sites; 3.47 s/step; the job: 101.6 min spent, about 139.0 min of steps left; card peak 14.4 GB
87
+ [ 6271.0s] step 1650/4000: task 0.2688 (mse 0.269 cos 0.868); quiet A 0.00137 off 0.00051; 1,493 sites; 3.47 s/step; the job: 104.5 min spent, about 136.1 min of steps left; card peak 14.4 GB
88
+ [ 6444.6s] step 1700/4000: task 0.2559 (mse 0.256 cos 0.874); quiet A 0.00134 off 0.00053; 1,534 sites; 3.47 s/step; the job: 107.4 min spent, about 133.2 min of steps left; card peak 14.4 GB
89
+ [ 6617.9s] step 1750/4000: task 0.2420 (mse 0.242 cos 0.880); quiet A 0.00112 off 0.00047; 1,511 sites; 3.47 s/step; the job: 110.3 min spent, about 130.3 min of steps left; card peak 14.4 GB
90
+ [ 6791.0s] step 1800/4000: task 0.2342 (mse 0.234 cos 0.884); quiet A 0.00118 off 0.00049; 1,512 sites; 3.47 s/step; the job: 113.2 min spent, about 127.4 min of steps left; card peak 14.4 GB
91
+ [ 6964.9s] step 1850/4000: task 0.2247 (mse 0.225 cos 0.888); quiet A 0.00109 off 0.00051; 1,457 sites; 3.47 s/step; the job: 116.1 min spent, about 124.5 min of steps left; card peak 14.4 GB
92
+ [ 7139.0s] step 1900/4000: task 0.2196 (mse 0.220 cos 0.891); quiet A 0.00137 off 0.00061; 1,504 sites; 3.47 s/step; the job: 119.0 min spent, about 121.6 min of steps left; card peak 14.4 GB
93
+ [ 7312.1s] step 1950/4000: task 0.2128 (mse 0.213 cos 0.895); quiet A 0.00130 off 0.00060; 1,434 sites; 3.47 s/step; the job: 121.9 min spent, about 118.7 min of steps left; card peak 14.4 GB
94
+ [ 7485.8s] step 2000/4000: task 0.2010 (mse 0.201 cos 0.900); quiet A 0.00120 off 0.00053; 1,477 sites; 3.47 s/step; the job: 124.8 min spent, about 115.8 min of steps left; card peak 14.4 GB
95
+ [ 7486.1s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXB_o1_close2_s10_qwen.safetensors
96
+ [ 7691.7s] 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
+ [ 7691.7s] block 8: 0.731 -> 0.926 / 0.928; A 0.9899; mse 0.151 cos 0.923
98
+ [ 7691.7s] block 12: 0.640 -> 0.924 / 0.924; A 0.9897; mse 0.123 cos 0.940
99
+ [ 7691.7s] block 16: 0.526 -> 0.935 / 0.935; A 0.9913; mse 0.100 cos 0.952
100
+ [ 7691.7s] block 18: 0.493 -> 0.937 / 0.939; A 0.9925; mse 0.093 cos 0.956
101
+ [ 7691.7s] block 20: 0.477 -> 0.939 / 0.940; A 0.9908; mse 0.089 cos 0.958
102
+ [ 7691.7s] block 24: 0.450 -> 0.942 / 0.942; A 0.9897; mse 0.085 cos 0.962
103
+ [ 7691.7s] block 28: 0.401 -> 0.932 / 0.931; A 0.9894; mse 0.101 cos 0.953
104
+ [ 7691.7s] block 31: 0.343 -> 0.882 / 0.887; A 0.9739; mse 0.198 cos 0.935
105
+ [ 7691.7s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.925 / 0.930 / 0.937 / 0.941 / 0.925
106
+ [ 7691.7s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.882 against the bar .85; B2: the lowest A alignment 0.9739 against .995; fineweb share +0.00083 +- 0.00027 against +.012; partners' shares s1 +0.0004, s2 +0.0002, s3 +0.0004, s4 +0.0006, s5 +0.0003, s6 +0.0001, s7 +0.0007, s8 +0.0001
107
+ [ 7691.7s] gate mean 0.137 / widest 0.268; compiled chain: 15 cache entries (limit 256), 0 limit warnings; docs drawn 64,676 (skipped {'test': 88, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 14.4 GB; host RSS 13.6 GB
108
+ [ 7864.7s] step 2050/4000: task 0.1904 (mse 0.190 cos 0.905); quiet A 0.00126 off 0.00052; 1,464 sites; 3.46 s/step; the job: 131.1 min spent, about 112.9 min of steps left; card peak 14.4 GB
109
+ [ 8038.9s] step 2100/4000: task 0.1866 (mse 0.187 cos 0.907); quiet A 0.00117 off 0.00057; 1,530 sites; 3.47 s/step; the job: 134.0 min spent, about 110.0 min of steps left; card peak 14.4 GB
110
+ [ 8212.3s] step 2150/4000: task 0.1804 (mse 0.180 cos 0.910); quiet A 0.00112 off 0.00055; 1,488 sites; 3.47 s/step; the job: 136.9 min spent, about 107.1 min of steps left; card peak 14.4 GB
111
+ [ 8385.6s] step 2200/4000: task 0.1755 (mse 0.175 cos 0.912); quiet A 0.00109 off 0.00054; 1,547 sites; 3.47 s/step; the job: 139.8 min spent, about 104.2 min of steps left; card peak 14.4 GB
112
+ [ 8559.4s] step 2250/4000: task 0.1700 (mse 0.170 cos 0.915); quiet A 0.00119 off 0.00079; 1,415 sites; 3.47 s/step; the job: 142.7 min spent, about 101.3 min of steps left; card peak 14.4 GB
113
+ [ 8732.8s] step 2300/4000: task 0.1655 (mse 0.166 cos 0.917); quiet A 0.00129 off 0.00049; 1,486 sites; 3.47 s/step; the job: 145.5 min spent, about 98.4 min of steps left; card peak 14.4 GB
114
+ [ 8906.1s] step 2350/4000: task 0.1577 (mse 0.158 cos 0.920); quiet A 0.00115 off 0.00055; 1,514 sites; 3.47 s/step; the job: 148.4 min spent, about 95.5 min of steps left; card peak 14.4 GB
115
+ [ 9079.3s] step 2400/4000: task 0.1554 (mse 0.155 cos 0.922); quiet A 0.00113 off 0.00049; 1,542 sites; 3.47 s/step; the job: 151.3 min spent, about 92.6 min of steps left; card peak 14.4 GB
116
+ [ 9253.1s] step 2450/4000: task 0.1494 (mse 0.149 cos 0.925); quiet A 0.00108 off 0.00057; 1,478 sites; 3.47 s/step; the job: 154.2 min spent, about 89.7 min of steps left; card peak 14.4 GB
117
+ [ 9426.4s] step 2500/4000: task 0.1409 (mse 0.141 cos 0.929); quiet A 0.00131 off 0.00057; 1,499 sites; 3.47 s/step; the job: 157.1 min spent, about 86.8 min of steps left; card peak 14.4 GB
118
+ [ 9601.0s] step 2550/4000: task 0.1402 (mse 0.140 cos 0.930); quiet A 0.00115 off 0.00050; 1,455 sites; 3.47 s/step; the job: 160.0 min spent, about 83.9 min of steps left; card peak 14.4 GB
119
+ [ 9773.7s] step 2600/4000: task 0.1395 (mse 0.140 cos 0.930); quiet A 0.00107 off 0.00054; 1,482 sites; 3.47 s/step; the job: 162.9 min spent, about 81.0 min of steps left; card peak 14.4 GB
120
+ [ 9946.4s] step 2650/4000: task 0.1253 (mse 0.125 cos 0.936); quiet A 0.00101 off 0.00049; 1,459 sites; 3.47 s/step; the job: 165.8 min spent, about 78.1 min of steps left; card peak 14.4 GB
121
+ [10119.4s] step 2700/4000: task 0.1243 (mse 0.124 cos 0.937); quiet A 0.00103 off 0.00055; 1,451 sites; 3.47 s/step; the job: 168.7 min spent, about 75.2 min of steps left; card peak 14.4 GB
122
+ [10293.7s] step 2750/4000: task 0.1200 (mse 0.120 cos 0.939); quiet A 0.00106 off 0.00047; 1,499 sites; 3.47 s/step; the job: 171.6 min spent, about 72.3 min of steps left; card peak 14.4 GB
123
+ [10467.5s] step 2800/4000: task 0.1207 (mse 0.121 cos 0.939); quiet A 0.00098 off 0.00042; 1,488 sites; 3.47 s/step; the job: 174.5 min spent, about 69.4 min of steps left; card peak 14.4 GB
124
+ [10640.7s] step 2850/4000: task 0.1165 (mse 0.117 cos 0.941); quiet A 0.00098 off 0.00045; 1,453 sites; 3.47 s/step; the job: 177.3 min spent, about 66.6 min of steps left; card peak 14.4 GB
125
+ [10813.0s] step 2900/4000: task 0.1097 (mse 0.110 cos 0.944); quiet A 0.00095 off 0.00040; 1,500 sites; 3.47 s/step; the job: 180.2 min spent, about 63.7 min of steps left; card peak 14.4 GB
126
+ [10986.8s] step 2950/4000: task 0.1102 (mse 0.110 cos 0.944); quiet A 0.00098 off 0.00044; 1,466 sites; 3.47 s/step; the job: 183.1 min spent, about 60.8 min of steps left; card peak 14.4 GB
127
+ [11160.3s] step 3000/4000: task 0.1083 (mse 0.108 cos 0.945); quiet A 0.00103 off 0.00049; 1,472 sites; 3.47 s/step; the job: 186.0 min spent, about 57.9 min of steps left; card peak 14.4 GB
128
+ [11160.5s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXB_o1_close3_s10_qwen.safetensors
129
+ [11366.1s] 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
+ [11366.1s] block 8: 0.731 -> 0.944 / 0.945; A 0.9941; mse 0.097 cos 0.951
131
+ [11366.1s] block 12: 0.640 -> 0.944 / 0.944; A 0.9952; mse 0.080 cos 0.961
132
+ [11366.1s] block 16: 0.526 -> 0.953 / 0.952; A 0.9951; mse 0.066 cos 0.969
133
+ [11366.1s] block 18: 0.493 -> 0.956 / 0.957; A 0.9958; mse 0.061 cos 0.971
134
+ [11366.1s] block 20: 0.477 -> 0.958 / 0.957; A 0.9945; mse 0.059 cos 0.973
135
+ [11366.1s] block 24: 0.450 -> 0.961 / 0.960; A 0.9934; mse 0.056 cos 0.975
136
+ [11366.1s] block 28: 0.401 -> 0.953 / 0.953; A 0.9914; mse 0.067 cos 0.969
137
+ [11366.1s] block 31: 0.343 -> 0.906 / 0.910; A 0.9821; mse 0.145 cos 0.956
138
+ [11366.1s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.946 / 0.949 / 0.953 / 0.957 / 0.946
139
+ [11366.1s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.906 against the bar .85; B2: the lowest A alignment 0.9821 against .995; fineweb share +0.00031 +- 0.00021 against +.012; partners' shares s1 +0.0002, s2 +0.0004, s3 +0.0002, s4 +0.0002, s5 +0.0001, s6 +0.0000, s7 +0.0004, s8 +0.0001
140
+ [11366.1s] gate mean 0.156 / widest 0.385; compiled chain: 15 cache entries (limit 256), 0 limit warnings; docs drawn 96,834 (skipped {'test': 137, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 14.4 GB; host RSS 13.6 GB
141
+ [11538.4s] step 3050/4000: task 0.1121 (mse 0.112 cos 0.943); quiet A 0.00096 off 0.00050; 1,481 sites; 3.45 s/step; the job: 192.3 min spent, about 55.0 min of steps left; card peak 14.4 GB
142
+ [11710.6s] step 3100/4000: task 0.1008 (mse 0.101 cos 0.948); quiet A 0.00096 off 0.00042; 1,458 sites; 3.44 s/step; the job: 195.2 min spent, about 52.1 min of steps left; card peak 14.4 GB
143
+ [11884.1s] step 3150/4000: task 0.1019 (mse 0.102 cos 0.948); quiet A 0.00098 off 0.00067; 1,473 sites; 3.45 s/step; the job: 198.1 min spent, about 49.2 min of steps left; card peak 14.4 GB
144
+ [12058.2s] step 3200/4000: task 0.1002 (mse 0.100 cos 0.949); quiet A 0.00097 off 0.00040; 1,496 sites; 3.46 s/step; the job: 201.0 min spent, about 46.3 min of steps left; card peak 14.4 GB
145
+ [12231.8s] step 3250/4000: task 0.0946 (mse 0.095 cos 0.951); quiet A 0.00096 off 0.00044; 1,489 sites; 3.46 s/step; the job: 203.9 min spent, about 43.4 min of steps left; card peak 14.4 GB
146
+ [12405.1s] step 3300/4000: task 0.0962 (mse 0.096 cos 0.951); quiet A 0.00093 off 0.00048; 1,502 sites; 3.46 s/step; the job: 206.8 min spent, about 40.5 min of steps left; card peak 14.4 GB
147
+ [12578.0s] step 3350/4000: task 0.0910 (mse 0.091 cos 0.953); quiet A 0.00091 off 0.00066; 1,459 sites; 3.46 s/step; the job: 209.6 min spent, about 37.6 min of steps left; card peak 14.4 GB
148
+ [12751.6s] step 3400/4000: task 0.0909 (mse 0.091 cos 0.953); quiet A 0.00089 off 0.00044; 1,475 sites; 3.46 s/step; the job: 212.5 min spent, about 34.7 min of steps left; card peak 14.4 GB
149
+ [12924.6s] step 3450/4000: task 0.0856 (mse 0.086 cos 0.956); quiet A 0.00090 off 0.00040; 1,508 sites; 3.46 s/step; the job: 215.4 min spent, about 31.8 min of steps left; card peak 14.4 GB
150
+ [13097.9s] step 3500/4000: task 0.0842 (mse 0.084 cos 0.956); quiet A 0.00089 off 0.00041; 1,556 sites; 3.46 s/step; the job: 218.3 min spent, about 28.9 min of steps left; card peak 14.4 GB
151
+ [13271.2s] step 3550/4000: task 0.0838 (mse 0.084 cos 0.957); quiet A 0.00086 off 0.00038; 1,469 sites; 3.46 s/step; the job: 221.2 min spent, about 26.0 min of steps left; card peak 14.4 GB
152
+ [13444.1s] step 3600/4000: task 0.0849 (mse 0.085 cos 0.956); quiet A 0.00088 off 0.00039; 1,488 sites; 3.46 s/step; the job: 224.1 min spent, about 23.1 min of steps left; card peak 14.4 GB
153
+ [13616.6s] step 3650/4000: task 0.0794 (mse 0.079 cos 0.959); quiet A 0.00085 off 0.00047; 1,434 sites; 3.46 s/step; the job: 226.9 min spent, about 20.2 min of steps left; card peak 14.4 GB
154
+ [13790.3s] step 3700/4000: task 0.0785 (mse 0.078 cos 0.959); quiet A 0.00083 off 0.00039; 1,473 sites; 3.46 s/step; the job: 229.8 min spent, about 17.4 min of steps left; card peak 14.4 GB
155
+ [13963.7s] step 3750/4000: task 0.0756 (mse 0.076 cos 0.960); quiet A 0.00085 off 0.00037; 1,473 sites; 3.46 s/step; the job: 232.7 min spent, about 14.5 min of steps left; card peak 14.4 GB
156
+ [14136.8s] step 3800/4000: task 0.0776 (mse 0.078 cos 0.960); quiet A 0.00083 off 0.00037; 1,441 sites; 3.46 s/step; the job: 235.6 min spent, about 11.6 min of steps left; card peak 14.4 GB
157
+ [14310.3s] step 3850/4000: task 0.0781 (mse 0.078 cos 0.960); quiet A 0.00084 off 0.00037; 1,474 sites; 3.46 s/step; the job: 238.5 min spent, about 8.7 min of steps left; card peak 14.4 GB
158
+ [14483.5s] step 3900/4000: task 0.0743 (mse 0.074 cos 0.961); quiet A 0.00085 off 0.00035; 1,456 sites; 3.46 s/step; the job: 241.4 min spent, about 5.8 min of steps left; card peak 14.4 GB
159
+ [14656.9s] step 3950/4000: task 0.0718 (mse 0.072 cos 0.963); quiet A 0.00084 off 0.00045; 1,457 sites; 3.46 s/step; the job: 244.3 min spent, about 2.9 min of steps left; card peak 14.4 GB
160
+ [14830.0s] step 4000/4000: task 0.0732 (mse 0.073 cos 0.962); quiet A 0.00083 off 0.00039; 1,551 sites; 3.46 s/step; the job: 247.2 min spent, about 0.0 min of steps left; card peak 14.4 GB
161
+ [14830.2s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXB_o1_close4_s10_qwen.safetensors
162
+ [15038.4s] 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
+ [15038.4s] block 8: 0.731 -> 0.953 / 0.952; A 0.9971; mse 0.075 cos 0.962
164
+ [15038.4s] block 12: 0.640 -> 0.953 / 0.953; A 0.9969; mse 0.064 cos 0.969
165
+ [15038.4s] block 16: 0.526 -> 0.960 / 0.957; A 0.9958; mse 0.054 cos 0.974
166
+ [15038.4s] block 18: 0.493 -> 0.964 / 0.963; A 0.9969; mse 0.051 cos 0.976
167
+ [15038.4s] block 20: 0.477 -> 0.963 / 0.963; A 0.9936; mse 0.050 cos 0.977
168
+ [15038.4s] block 24: 0.450 -> 0.963 / 0.963; A 0.9949; mse 0.049 cos 0.978
169
+ [15038.4s] block 28: 0.401 -> 0.957 / 0.956; A 0.9945; mse 0.059 cos 0.973
170
+ [15038.4s] block 31: 0.343 -> 0.915 / 0.914; A 0.9847; mse 0.127 cos 0.961
171
+ [15038.4s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.951 / 0.955 / 0.959 / 0.960 / 0.952
172
+ [15038.4s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.915 against the bar .85; B2: the lowest A alignment 0.9847 against .995; fineweb share +0.00032 +- 0.00015 against +.012; partners' shares s1 +0.0001, s2 +0.0002, s3 +0.0002, s4 +0.0002, s5 +0.0001, s6 +0.0001, s7 +0.0002, s8 +0.0001
173
+ [15038.4s] gate mean 0.169 / widest 0.485; compiled chain: 15 cache entries (limit 256), 0 limit warnings; docs drawn 128,945 (skipped {'test': 195, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 14.4 GB; host RSS 13.6 GB
174
+ [15038.4s] DONE qarm_mse_gXB_o1: 4000 steps in 231.3 min of steps (3.47 s/step); 4 closes; 15038 s in all; card peak 14.4 GB; host RSS 13.6 GB; compiled chain: 15 cache entries (limit 256), 0 limit warnings