Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
cf185d4
·
verified ·
1 Parent(s): ed98a10

surface arm qwen3/mse_gXA_o0: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/qwen3/mse_gXA_o0/closes/qarm_mse_gXA_o0_close1_s10_qwen.safetensors +3 -0
  2. surface/qwen3/mse_gXA_o0/closes/qarm_mse_gXA_o0_close2_s10_qwen.safetensors +3 -0
  3. surface/qwen3/mse_gXA_o0/closes/qarm_mse_gXA_o0_close3_s10_qwen.safetensors +3 -0
  4. surface/qwen3/mse_gXA_o0/closes/qarm_mse_gXA_o0_close4_s10_qwen.safetensors +3 -0
  5. surface/qwen3/mse_gXA_o0/manifest.json +20 -0
  6. surface/qwen3/mse_gXA_o0/readout.md +127 -0
  7. surface/qwen3/mse_gXA_o0/results.json +0 -0
  8. surface/qwen3/mse_gXA_o0/s10_qwen.safetensors +3 -0
  9. surface/qwen3/mse_gXA_o0/stats.pt +3 -0
  10. surface/qwen3/mse_gXA_o0/train.log +174 -0
surface/qwen3/mse_gXA_o0/closes/qarm_mse_gXA_o0_close1_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8b57d7c8477348e2e38fbe696431c7dc559f97e38ff2ac5d43d6a78a6010ea73
3
+ size 54819192
surface/qwen3/mse_gXA_o0/closes/qarm_mse_gXA_o0_close2_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:676ed8a7e717c418b3ec91b7ec9cb1ab480c260c90c5e192c5ba34b20c0b6bb5
3
+ size 54819192
surface/qwen3/mse_gXA_o0/closes/qarm_mse_gXA_o0_close3_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4500aa46a1cd0c46f02715b46fc541af7855ae8b79066950d803f24ab6ab6ef2
3
+ size 54819192
surface/qwen3/mse_gXA_o0/closes/qarm_mse_gXA_o0_close4_s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d240b902c8f8ad90fa637f99b0fa933df49a1f59fc4ba05ba2f9baa5a224e790
3
+ size 54819192
surface/qwen3/mse_gXA_o0/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "qarm_mse_gXA_o0",
3
+ "run": "mse_gXA_o0",
4
+ "convention": "qwen3",
5
+ "member": "s10_qwen",
6
+ "final_anchor": "qarm_mse_gXA_o0_close4_s10_qwen.safetensors",
7
+ "closes": [
8
+ "qarm_mse_gXA_o0_close1_s10_qwen.safetensors",
9
+ "qarm_mse_gXA_o0_close2_s10_qwen.safetensors",
10
+ "qarm_mse_gXA_o0_close3_s10_qwen.safetensors",
11
+ "qarm_mse_gXA_o0_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 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_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_gXA_o0: 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
12
+
13
+ - B1 the gap: MET over every served block >= 12 (the lowest 0.895 at block 31 on draw 1, 0.892 at block 31 on draw 2; bar 0.85)
14
+ - B2 silence, whitened: NOT MET (the lowest A alignment 0.9835 at block 31 on draw 1, 0.9822 on draw 2; bar 0.995)
15
+ - B2 silence, in bits: MET (web +0.00030 bpb; the worst partner s3_rules +0.00054; bar +0.012)
16
+
17
+ ## The closes
18
+
19
+ ### qarm_mse_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 | 0.705 / 0.696 | 0.863 / 0.875 / 0.853 | 0.9525 / 0.9567 | +0.00018 | 0.116 / 0.168 |
25
+ | 2 | 2000 | 0.824 / 0.817 | 0.925 / 0.936 / 0.924 | 0.9580 / 0.9554 | +0.00042 | 0.136 / 0.279 |
26
+ | 3 | 3000 | 0.873 / 0.866 | 0.944 / 0.958 / 0.949 | 0.9733 / 0.9762 | +0.00049 | 0.151 / 0.396 |
27
+ | 4 | 4000 | 0.895 / 0.892 | 0.957 / 0.969 / 0.960 | 0.9835 / 0.9822 | +0.00030 | 0.163 / 0.477 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### qarm_mse_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 | 0.898 | 0.9996 | 0.363 | 0.775 |
36
+ | 4 | 0.790 | 0.948 | 0.948 | 0.9984 | 0.090 | 0.951 |
37
+ | 8 | 0.733 | 0.955 | 0.954 | 0.9977 | 0.069 | 0.965 |
38
+ | 12 | 0.640 | 0.957 | 0.958 | 0.9966 | 0.058 | 0.972 |
39
+ | 16 | 0.532 | 0.962 | 0.963 | 0.9949 | 0.049 | 0.977 |
40
+ | 18 | 0.502 | 0.967 | 0.966 | 0.9976 | 0.046 | 0.978 |
41
+ | 20 | 0.486 | 0.969 | 0.968 | 0.9971 | 0.045 | 0.979 |
42
+ | 22 | 0.472 | 0.971 | 0.970 | 0.9975 | 0.044 | 0.981 |
43
+ | 24 | 0.457 | 0.968 | 0.967 | 0.9973 | 0.044 | 0.980 |
44
+ | 26 | 0.430 | 0.967 | 0.967 | 0.9963 | 0.048 | 0.979 |
45
+ | 28 | 0.407 | 0.960 | 0.959 | 0.9963 | 0.058 | 0.975 |
46
+ | 29 | 0.398 | 0.954 | 0.954 | 0.9961 | 0.067 | 0.973 |
47
+ | 30 | 0.376 | 0.944 | 0.944 | 0.9941 | 0.111 | 0.970 |
48
+ | 31 | 0.322 | 0.895 | 0.895 | 0.9835 | 0.314 | 0.960 |
49
+
50
+ ### qarm_mse_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 | 0.886 | 0.886 | 0.9985 | 0.374 | 0.770 |
55
+ | 4 | 0.795 | 0.947 | 0.947 | 0.9982 | 0.091 | 0.950 |
56
+ | 8 | 0.738 | 0.953 | 0.953 | 0.9976 | 0.070 | 0.964 |
57
+ | 12 | 0.644 | 0.955 | 0.954 | 0.9974 | 0.059 | 0.971 |
58
+ | 16 | 0.540 | 0.960 | 0.960 | 0.9969 | 0.050 | 0.976 |
59
+ | 18 | 0.505 | 0.963 | 0.962 | 0.9965 | 0.047 | 0.977 |
60
+ | 20 | 0.492 | 0.964 | 0.964 | 0.9973 | 0.046 | 0.978 |
61
+ | 22 | 0.481 | 0.963 | 0.963 | 0.9964 | 0.046 | 0.980 |
62
+ | 24 | 0.465 | 0.967 | 0.967 | 0.9971 | 0.046 | 0.979 |
63
+ | 26 | 0.440 | 0.962 | 0.961 | 0.9957 | 0.050 | 0.978 |
64
+ | 28 | 0.419 | 0.958 | 0.958 | 0.9955 | 0.060 | 0.974 |
65
+ | 29 | 0.409 | 0.953 | 0.953 | 0.9926 | 0.069 | 0.972 |
66
+ | 30 | 0.388 | 0.940 | 0.940 | 0.9934 | 0.114 | 0.969 |
67
+ | 31 | 0.319 | 0.892 | 0.893 | 0.9822 | 0.307 | 0.959 |
68
+
69
+ The member's share in bits per byte (all on minus the member masked): s1_perspective +0.00001, s2_concept -0.00002, s3_rules +0.00054, s4_arith +0.00023, s5_causal +0.00012, s6_tryfail +0.00001, s7_mixed +0.00008, s8_register +0.00002, web +0.00030
70
+
71
+ ## The training curves
72
+
73
+ ### qarm_mse_gXA_o0
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 | 1.9286 | 1.929 | 0.254 | None | 0.00033 | 0.00016 |
79
+ | 250 | 1.7568 | 1.757 | 0.293 | None | 0.00177 | 0.00065 |
80
+ | 500 | 0.8949 | 0.895 | 0.595 | None | 0.00122 | 0.00044 |
81
+ | 1000 | 0.4951 | 0.495 | 0.764 | None | 0.00088 | 0.00032 |
82
+ | 1500 | 0.3057 | 0.306 | 0.851 | None | 0.00099 | 0.00039 |
83
+ | 2000 | 0.1946 | 0.195 | 0.904 | None | 0.00106 | 0.00047 |
84
+ | 2500 | 0.1307 | 0.131 | 0.934 | None | 0.00107 | 0.00112 |
85
+ | 3000 | 0.0991 | 0.099 | 0.950 | None | 0.00102 | 0.00045 |
86
+ | 3500 | 0.0758 | 0.076 | 0.961 | None | 0.00084 | 0.00039 |
87
+ | 4000 | 0.0692 | 0.069 | 0.964 | None | 0.00077 | 0.00034 |
88
+ | 1 | 1.9230 | 1.923 | 0.260 | None | 0.00025 | 0.00012 |
89
+ | 50 | 1.9286 | 1.929 | 0.254 | None | 0.00033 | 0.00016 |
90
+ | 250 | 1.7568 | 1.757 | 0.293 | None | 0.00177 | 0.00065 |
91
+ | 500 | 0.8949 | 0.895 | 0.595 | None | 0.00122 | 0.00044 |
92
+ | 1000 | 0.4951 | 0.495 | 0.764 | None | 0.00088 | 0.00032 |
93
+ | 1500 | 0.3057 | 0.306 | 0.851 | None | 0.00099 | 0.00039 |
94
+ | 2000 | 0.1946 | 0.195 | 0.904 | None | 0.00106 | 0.00047 |
95
+ | 2500 | 0.1307 | 0.131 | 0.934 | None | 0.00107 | 0.00112 |
96
+ | 3000 | 0.0991 | 0.099 | 0.950 | None | 0.00102 | 0.00045 |
97
+ | 3500 | 0.0758 | 0.076 | 0.961 | None | 0.00084 | 0.00039 |
98
+ | 4000 | 0.0692 | 0.069 | 0.964 | None | 0.00077 | 0.00034 |
99
+ | 1 | 1.9230 | 1.923 | 0.260 | None | 0.00025 | 0.00012 |
100
+ | 50 | 1.9286 | 1.929 | 0.254 | None | 0.00033 | 0.00016 |
101
+ | 250 | 1.7568 | 1.757 | 0.293 | None | 0.00177 | 0.00065 |
102
+ | 500 | 0.8949 | 0.895 | 0.595 | None | 0.00122 | 0.00044 |
103
+ | 1000 | 0.4951 | 0.495 | 0.764 | None | 0.00088 | 0.00032 |
104
+ | 1500 | 0.3057 | 0.306 | 0.851 | None | 0.00099 | 0.00039 |
105
+ | 2000 | 0.1946 | 0.195 | 0.904 | None | 0.00106 | 0.00047 |
106
+ | 2500 | 0.1307 | 0.131 | 0.934 | None | 0.00107 | 0.00112 |
107
+ | 3000 | 0.0991 | 0.099 | 0.950 | None | 0.00102 | 0.00045 |
108
+ | 3500 | 0.0758 | 0.076 | 0.961 | None | 0.00084 | 0.00039 |
109
+ | 4000 | 0.0692 | 0.069 | 0.964 | None | 0.00077 | 0.00034 |
110
+ | 1 | 1.9230 | 1.923 | 0.260 | None | 0.00025 | 0.00012 |
111
+ | 50 | 1.9286 | 1.929 | 0.254 | None | 0.00033 | 0.00016 |
112
+ | 250 | 1.7568 | 1.757 | 0.293 | None | 0.00177 | 0.00065 |
113
+ | 500 | 0.8949 | 0.895 | 0.595 | None | 0.00122 | 0.00044 |
114
+ | 1000 | 0.4951 | 0.495 | 0.764 | None | 0.00088 | 0.00032 |
115
+ | 1500 | 0.3057 | 0.306 | 0.851 | None | 0.00099 | 0.00039 |
116
+ | 2000 | 0.1946 | 0.195 | 0.904 | None | 0.00106 | 0.00047 |
117
+ | 2500 | 0.1307 | 0.131 | 0.934 | None | 0.00107 | 0.00112 |
118
+ | 3000 | 0.0991 | 0.099 | 0.950 | None | 0.00102 | 0.00045 |
119
+ | 3500 | 0.0758 | 0.076 | 0.961 | None | 0.00084 | 0.00039 |
120
+ | 4000 | 0.0692 | 0.069 | 0.964 | None | 0.00077 | 0.00034 |
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/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/qwen3/mse_gXA_o0/s10_qwen.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d240b902c8f8ad90fa637f99b0fa933df49a1f59fc4ba05ba2f9baa5a224e790
3
+ size 54819192
surface/qwen3/mse_gXA_o0/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:39ec14a65e9b56c168af03d460caf463dc0553547453ceba89a594b00bf8c076
3
+ size 264213
surface/qwen3/mse_gXA_o0/train.log ADDED
@@ -0,0 +1,174 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 0.4s] THE JOB: qarm_mse_gXA_o0: the Qwen tokenizer arm (s10_qwen) over the frozen eight of gXA; 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
+ [ 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
+ [ 29.3s] 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
+ [ 55.0s] 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
+ [ 68.4s] 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
+ [ 68.5s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 116), 4,920 sites
24
+ [ 68.5s] 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
+ [ 81.7s] 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
+ [ 87.5s] 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_stats.pt
28
+ [ 350.1s] CLOSE 0 (step 0): the member's alignment read on draw 1 (512 captions, 4,920 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
29
+ [ 350.1s] block 8: 0.733 -> 0.733 / 0.733; A 1.0000; mse 1.718 cos 0.366
30
+ [ 350.1s] block 12: 0.640 -> 0.640 / 0.640; A 1.0000; mse 1.718 cos 0.307
31
+ [ 350.1s] block 16: 0.532 -> 0.532 / 0.532; A 1.0000; mse 1.699 cos 0.250
32
+ [ 350.1s] block 18: 0.502 -> 0.502 / 0.502; A 1.0000; mse 1.682 cos 0.236
33
+ [ 350.1s] block 20: 0.486 -> 0.486 / 0.486; A 1.0000; mse 1.699 cos 0.233
34
+ [ 350.1s] block 24: 0.457 -> 0.457 / 0.457; A 1.0000; mse 1.858 cos 0.231
35
+ [ 350.1s] block 28: 0.407 -> 0.407 / 0.407; A 1.0000; mse 2.189 cos 0.194
36
+ [ 350.1s] block 31: 0.322 -> 0.322 / 0.322; A 1.0000; mse 5.319 cos 0.105
37
+ [ 350.1s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.644 / 0.540 / 0.492 / 0.465 / 0.419
38
+ [ 350.1s] 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
+ [ 350.1s] 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 13.0 GB
40
+ [stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
41
+ [ 359.2s] step 1/4000: task 1.9230 (mse 1.923 cos 0.260); quiet A 0.00025 off 0.00012; 1,503 sites; 9.02 s/step; the job: 6.0 min spent, about 601.2 min of steps left; card peak 14.2 GB
42
+ [ 528.4s] step 50/4000: task 1.9286 (mse 1.929 cos 0.254); quiet A 0.00033 off 0.00016; 1,508 sites; 3.56 s/step; the job: 8.8 min spent, about 234.7 min of steps left; card peak 14.3 GB
43
+ [ 701.2s] step 100/4000: task 1.9073 (mse 1.907 cos 0.258); quiet A 0.00054 off 0.00023; 1,506 sites; 3.51 s/step; the job: 11.7 min spent, about 228.2 min of steps left; card peak 14.3 GB
44
+ [ 874.4s] step 150/4000: task 1.8695 (mse 1.870 cos 0.267); quiet A 0.00083 off 0.00034; 1,471 sites; 3.49 s/step; the job: 14.6 min spent, about 224.2 min of steps left; card peak 14.3 GB
45
+ [ 1048.4s] step 200/4000: task 1.8167 (mse 1.817 cos 0.279); quiet A 0.00133 off 0.00055; 1,483 sites; 3.49 s/step; the job: 17.5 min spent, about 221.1 min of steps left; card peak 14.3 GB
46
+ [ 1221.6s] step 250/4000: task 1.7568 (mse 1.757 cos 0.293); quiet A 0.00177 off 0.00065; 1,485 sites; 3.49 s/step; the job: 20.4 min spent, about 217.9 min of steps left; card peak 14.3 GB
47
+ [ 1394.3s] step 300/4000: task 1.6762 (mse 1.676 cos 0.313); quiet A 0.00227 off 0.00070; 1,497 sites; 3.48 s/step; the job: 23.2 min spent, about 214.6 min of steps left; card peak 14.4 GB
48
+ [ 1567.7s] step 350/4000: task 1.5460 (mse 1.546 cos 0.350); quiet A 0.00246 off 0.00051; 1,487 sites; 3.48 s/step; the job: 26.1 min spent, about 211.6 min of steps left; card peak 14.4 GB
49
+ [ 1741.4s] step 400/4000: task 1.3094 (mse 1.309 cos 0.419); quiet A 0.00225 off 0.00067; 1,505 sites; 3.48 s/step; the job: 29.0 min spent, about 208.7 min of steps left; card peak 14.4 GB
50
+ [ 1914.4s] step 450/4000: task 1.0378 (mse 1.038 cos 0.530); quiet A 0.00153 off 0.00057; 1,593 sites; 3.48 s/step; the job: 31.9 min spent, about 205.7 min of steps left; card peak 14.4 GB
51
+ [ 2088.1s] step 500/4000: task 0.8949 (mse 0.895 cos 0.595); quiet A 0.00122 off 0.00044; 1,535 sites; 3.48 s/step; the job: 34.8 min spent, about 202.8 min of steps left; card peak 14.4 GB
52
+ [ 2261.1s] step 550/4000: task 0.8098 (mse 0.810 cos 0.632); quiet A 0.00111 off 0.00037; 1,433 sites; 3.47 s/step; the job: 37.7 min spent, about 199.8 min of steps left; card peak 14.4 GB
53
+ [ 2434.9s] step 600/4000: task 0.7493 (mse 0.749 cos 0.658); quiet A 0.00108 off 0.00038; 1,487 sites; 3.47 s/step; the job: 40.6 min spent, about 196.9 min of steps left; card peak 14.4 GB
54
+ [ 2608.3s] step 650/4000: task 0.6948 (mse 0.695 cos 0.682); quiet A 0.00100 off 0.00035; 1,558 sites; 3.47 s/step; the job: 43.5 min spent, about 194.0 min of steps left; card peak 14.4 GB
55
+ [ 2781.2s] step 700/4000: task 0.6510 (mse 0.651 cos 0.699); quiet A 0.00094 off 0.00036; 1,429 sites; 3.47 s/step; the job: 46.4 min spent, about 191.0 min of steps left; card peak 14.4 GB
56
+ [ 2954.4s] step 750/4000: task 0.6163 (mse 0.616 cos 0.714); quiet A 0.00090 off 0.00031; 1,468 sites; 3.47 s/step; the job: 49.2 min spent, about 188.1 min of steps left; card peak 14.4 GB
57
+ [ 3127.1s] step 800/4000: task 0.5852 (mse 0.585 cos 0.725); quiet A 0.00093 off 0.00031; 1,541 sites; 3.47 s/step; the job: 52.1 min spent, about 185.1 min of steps left; card peak 14.4 GB
58
+ [ 3299.9s] step 850/4000: task 0.5619 (mse 0.562 cos 0.736); quiet A 0.00092 off 0.00034; 1,440 sites; 3.47 s/step; the job: 55.0 min spent, about 182.2 min of steps left; card peak 14.4 GB
59
+ [ 3472.8s] step 900/4000: task 0.5322 (mse 0.532 cos 0.748); quiet A 0.00095 off 0.00033; 1,652 sites; 3.47 s/step; the job: 57.9 min spent, about 179.3 min of steps left; card peak 14.4 GB
60
+ [ 3645.4s] step 950/4000: task 0.5113 (mse 0.511 cos 0.757); quiet A 0.00091 off 0.00033; 1,550 sites; 3.47 s/step; the job: 60.8 min spent, about 176.3 min of steps left; card peak 14.4 GB
61
+ [ 3818.5s] step 1000/4000: task 0.4951 (mse 0.495 cos 0.764); quiet A 0.00088 off 0.00032; 1,493 sites; 3.47 s/step; the job: 63.6 min spent, about 173.4 min of steps left; card peak 14.4 GB
62
+ [ 3818.8s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXA_o0_close1_s10_qwen.safetensors
63
+ [ 4024.5s] 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
+ [ 4024.5s] block 8: 0.733 -> 0.869 / 0.873; A 0.9908; mse 0.369 cos 0.817
65
+ [ 4024.5s] block 12: 0.640 -> 0.863 / 0.869; A 0.9901; mse 0.296 cos 0.854
66
+ [ 4024.5s] block 16: 0.532 -> 0.867 / 0.872; A 0.9900; mse 0.239 cos 0.884
67
+ [ 4024.5s] block 18: 0.502 -> 0.869 / 0.875; A 0.9903; mse 0.225 cos 0.891
68
+ [ 4024.5s] block 20: 0.486 -> 0.875 / 0.880; A 0.9912; mse 0.217 cos 0.897
69
+ [ 4024.5s] block 24: 0.457 -> 0.879 / 0.884; A 0.9869; mse 0.212 cos 0.905
70
+ [ 4024.5s] block 28: 0.407 -> 0.853 / 0.859; A 0.9867; mse 0.261 cos 0.884
71
+ [ 4024.5s] block 31: 0.322 -> 0.705 / 0.720; A 0.9525; mse 1.143 cos 0.836
72
+ [ 4024.5s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.859 / 0.861 / 0.871 / 0.875 / 0.848
73
+ [ 4024.5s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.705 against the bar .85; B2: the lowest A alignment 0.9525 against .995; fineweb share +0.00018 +- 0.00012 against +.012; partners' shares s1 +0.0001, s2 +0.0002, s3 +0.0003, s4 +0.0002, s5 +0.0002, s6 -0.0000, s7 -0.0001, s8 +0.0000
74
+ [ 4024.5s] 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 13.5 GB
75
+ [ 4197.3s] step 1050/4000: task 0.4695 (mse 0.469 cos 0.775); quiet A 0.00089 off 0.00032; 1,510 sites; 3.46 s/step; the job: 70.0 min spent, about 170.5 min of steps left; card peak 14.4 GB
76
+ [ 4370.3s] step 1100/4000: task 0.4459 (mse 0.446 cos 0.784); quiet A 0.00084 off 0.00033; 1,491 sites; 3.46 s/step; the job: 72.8 min spent, about 167.6 min of steps left; card peak 14.4 GB
77
+ [ 4543.2s] step 1150/4000: task 0.4326 (mse 0.433 cos 0.790); quiet A 0.00089 off 0.00040; 1,475 sites; 3.46 s/step; the job: 75.7 min spent, about 164.7 min of steps left; card peak 14.4 GB
78
+ [ 4715.9s] step 1200/4000: task 0.4138 (mse 0.414 cos 0.799); quiet A 0.00103 off 0.00095; 1,492 sites; 3.46 s/step; the job: 78.6 min spent, about 161.8 min of steps left; card peak 14.4 GB
79
+ [ 4889.0s] step 1250/4000: task 0.3914 (mse 0.391 cos 0.810); quiet A 0.00093 off 0.00033; 1,473 sites; 3.46 s/step; the job: 81.5 min spent, about 158.9 min of steps left; card peak 14.4 GB
80
+ [ 5061.9s] step 1300/4000: task 0.3742 (mse 0.374 cos 0.818); quiet A 0.00092 off 0.00033; 1,475 sites; 3.46 s/step; the job: 84.4 min spent, about 156.0 min of steps left; card peak 14.4 GB
81
+ [ 5235.2s] step 1350/4000: task 0.3559 (mse 0.356 cos 0.826); quiet A 0.00093 off 0.00036; 1,476 sites; 3.46 s/step; the job: 87.3 min spent, about 153.1 min of steps left; card peak 14.4 GB
82
+ [ 5408.5s] step 1400/4000: task 0.3368 (mse 0.337 cos 0.836); quiet A 0.00097 off 0.00035; 1,519 sites; 3.46 s/step; the job: 90.1 min spent, about 150.2 min of steps left; card peak 14.4 GB
83
+ [ 5581.3s] step 1450/4000: task 0.3221 (mse 0.322 cos 0.843); quiet A 0.00097 off 0.00049; 1,500 sites; 3.46 s/step; the job: 93.0 min spent, about 147.3 min of steps left; card peak 14.4 GB
84
+ [ 5754.5s] step 1500/4000: task 0.3057 (mse 0.306 cos 0.851); quiet A 0.00099 off 0.00039; 1,442 sites; 3.46 s/step; the job: 95.9 min spent, about 144.4 min of steps left; card peak 14.4 GB
85
+ [ 5928.2s] step 1550/4000: task 0.2902 (mse 0.290 cos 0.859); quiet A 0.00101 off 0.00042; 1,466 sites; 3.46 s/step; the job: 98.8 min spent, about 141.5 min of steps left; card peak 14.4 GB
86
+ [ 6101.9s] step 1600/4000: task 0.2757 (mse 0.276 cos 0.865); quiet A 0.00102 off 0.00040; 1,480 sites; 3.46 s/step; the job: 101.7 min spent, about 138.6 min of steps left; card peak 14.4 GB
87
+ [ 6275.1s] step 1650/4000: task 0.2619 (mse 0.262 cos 0.872); quiet A 0.00102 off 0.00039; 1,535 sites; 3.46 s/step; the job: 104.6 min spent, about 135.8 min of steps left; card peak 14.4 GB
88
+ [ 6448.3s] step 1700/4000: task 0.2546 (mse 0.255 cos 0.876); quiet A 0.00110 off 0.00043; 1,499 sites; 3.46 s/step; the job: 107.5 min spent, about 132.9 min of steps left; card peak 14.4 GB
89
+ [ 6622.0s] step 1750/4000: task 0.2376 (mse 0.238 cos 0.883); quiet A 0.00105 off 0.00040; 1,463 sites; 3.46 s/step; the job: 110.4 min spent, about 130.0 min of steps left; card peak 14.4 GB
90
+ [ 6795.6s] step 1800/4000: task 0.2285 (mse 0.228 cos 0.888); quiet A 0.00103 off 0.00052; 1,531 sites; 3.46 s/step; the job: 113.3 min spent, about 127.1 min of steps left; card peak 14.4 GB
91
+ [ 6969.6s] step 1850/4000: task 0.2182 (mse 0.218 cos 0.893); quiet A 0.00105 off 0.00044; 1,498 sites; 3.46 s/step; the job: 116.2 min spent, about 124.2 min of steps left; card peak 14.4 GB
92
+ [ 7143.4s] step 1900/4000: task 0.2130 (mse 0.213 cos 0.895); quiet A 0.00108 off 0.00046; 1,601 sites; 3.47 s/step; the job: 119.1 min spent, about 121.3 min of steps left; card peak 14.4 GB
93
+ [ 7317.1s] step 1950/4000: task 0.2054 (mse 0.205 cos 0.899); quiet A 0.00107 off 0.00047; 1,500 sites; 3.47 s/step; the job: 122.0 min spent, about 118.5 min of steps left; card peak 14.4 GB
94
+ [ 7491.1s] step 2000/4000: task 0.1946 (mse 0.195 cos 0.904); quiet A 0.00106 off 0.00047; 1,501 sites; 3.47 s/step; the job: 124.9 min spent, about 115.6 min of steps left; card peak 14.4 GB
95
+ [ 7491.3s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXA_o0_close2_s10_qwen.safetensors
96
+ [ 7697.2s] 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
+ [ 7697.2s] block 8: 0.733 -> 0.927 / 0.929; A 0.9914; mse 0.149 cos 0.924
98
+ [ 7697.2s] block 12: 0.640 -> 0.925 / 0.929; A 0.9913; mse 0.125 cos 0.939
99
+ [ 7697.2s] block 16: 0.532 -> 0.931 / 0.935; A 0.9900; mse 0.104 cos 0.950
100
+ [ 7697.2s] block 18: 0.502 -> 0.935 / 0.938; A 0.9932; mse 0.098 cos 0.953
101
+ [ 7697.2s] block 20: 0.486 -> 0.936 / 0.939; A 0.9925; mse 0.095 cos 0.955
102
+ [ 7697.2s] block 24: 0.457 -> 0.939 / 0.943; A 0.9910; mse 0.093 cos 0.959
103
+ [ 7697.2s] block 28: 0.407 -> 0.924 / 0.929; A 0.9905; mse 0.118 cos 0.949
104
+ [ 7697.2s] block 31: 0.322 -> 0.824 / 0.837; A 0.9580; mse 0.608 cos 0.923
105
+ [ 7697.2s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.923 / 0.928 / 0.934 / 0.937 / 0.920
106
+ [ 7697.2s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.824 against the bar .85; B2: the lowest A alignment 0.9580 against .995; fineweb share +0.00042 +- 0.00016 against +.012; partners' shares s1 +0.0001, s2 +0.0003, s3 +0.0005, s4 +0.0008, s5 +0.0001, s6 +0.0001, s7 +0.0002, s8 +0.0001
107
+ [ 7697.2s] gate mean 0.136 / widest 0.279; 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 13.6 GB
108
+ [ 7884.5s] step 2050/4000: task 0.1858 (mse 0.186 cos 0.908); quiet A 0.00109 off 0.00045; 1,509 sites; 3.75 s/step; the job: 131.4 min spent, about 112.9 min of steps left; card peak 14.4 GB
109
+ [ 8058.1s] step 2100/4000: task 0.1840 (mse 0.184 cos 0.909); quiet A 0.00123 off 0.00062; 1,504 sites; 3.61 s/step; the job: 134.3 min spent, about 110.0 min of steps left; card peak 14.4 GB
110
+ [ 8232.2s] step 2150/4000: task 0.1729 (mse 0.173 cos 0.914); quiet A 0.00107 off 0.00043; 1,495 sites; 3.57 s/step; the job: 137.2 min spent, about 107.1 min of steps left; card peak 14.4 GB
111
+ [ 8405.2s] step 2200/4000: task 0.1671 (mse 0.167 cos 0.917); quiet A 0.00111 off 0.00047; 1,510 sites; 3.54 s/step; the job: 140.1 min spent, about 104.2 min of steps left; card peak 14.4 GB
112
+ [ 8578.9s] step 2250/4000: task 0.1634 (mse 0.163 cos 0.919); quiet A 0.00107 off 0.00043; 1,547 sites; 3.53 s/step; the job: 143.0 min spent, about 101.3 min of steps left; card peak 14.4 GB
113
+ [ 8752.1s] step 2300/4000: task 0.1567 (mse 0.157 cos 0.922); quiet A 0.00101 off 0.00043; 1,479 sites; 3.52 s/step; the job: 145.9 min spent, about 98.4 min of steps left; card peak 14.4 GB
114
+ [ 8924.8s] step 2350/4000: task 0.1549 (mse 0.155 cos 0.923); quiet A 0.00107 off 0.00047; 1,529 sites; 3.51 s/step; the job: 148.7 min spent, about 95.5 min of steps left; card peak 14.4 GB
115
+ [ 9097.7s] step 2400/4000: task 0.1469 (mse 0.147 cos 0.927); quiet A 0.00104 off 0.00047; 1,459 sites; 3.50 s/step; the job: 151.6 min spent, about 92.6 min of steps left; card peak 14.4 GB
116
+ [ 9270.8s] step 2450/4000: task 0.1376 (mse 0.138 cos 0.931); quiet A 0.00107 off 0.00047; 1,471 sites; 3.50 s/step; the job: 154.5 min spent, about 89.7 min of steps left; card peak 14.4 GB
117
+ [ 9444.0s] step 2500/4000: task 0.1307 (mse 0.131 cos 0.934); quiet A 0.00107 off 0.00112; 1,553 sites; 3.49 s/step; the job: 157.4 min spent, about 86.8 min of steps left; card peak 14.4 GB
118
+ [ 9617.4s] step 2550/4000: task 0.1225 (mse 0.122 cos 0.938); quiet A 0.00107 off 0.00047; 1,506 sites; 3.49 s/step; the job: 160.3 min spent, about 83.9 min of steps left; card peak 14.4 GB
119
+ [ 9790.6s] step 2600/4000: task 0.1268 (mse 0.127 cos 0.937); quiet A 0.00101 off 0.00042; 1,443 sites; 3.49 s/step; the job: 163.2 min spent, about 81.0 min of steps left; card peak 14.4 GB
120
+ [ 9963.9s] step 2650/4000: task 0.1185 (mse 0.119 cos 0.940); quiet A 0.00097 off 0.00042; 1,447 sites; 3.49 s/step; the job: 166.1 min spent, about 78.1 min of steps left; card peak 14.4 GB
121
+ [10136.9s] step 2700/4000: task 0.1137 (mse 0.114 cos 0.943); quiet A 0.00097 off 0.00042; 1,511 sites; 3.49 s/step; the job: 168.9 min spent, about 75.2 min of steps left; card peak 14.4 GB
122
+ [10310.1s] step 2750/4000: task 0.1102 (mse 0.110 cos 0.944); quiet A 0.00096 off 0.00049; 1,482 sites; 3.48 s/step; the job: 171.8 min spent, about 72.3 min of steps left; card peak 14.4 GB
123
+ [10483.0s] step 2800/4000: task 0.1109 (mse 0.111 cos 0.944); quiet A 0.00099 off 0.00040; 1,489 sites; 3.48 s/step; the job: 174.7 min spent, about 69.4 min of steps left; card peak 14.4 GB
124
+ [10656.3s] step 2850/4000: task 0.1043 (mse 0.104 cos 0.947); quiet A 0.00096 off 0.00047; 1,523 sites; 3.48 s/step; the job: 177.6 min spent, about 66.5 min of steps left; card peak 14.4 GB
125
+ [10829.3s] step 2900/4000: task 0.1133 (mse 0.113 cos 0.945); quiet A 0.00160 off 0.00045; 1,518 sites; 3.48 s/step; the job: 180.5 min spent, about 63.6 min of steps left; card peak 14.5 GB
126
+ [11002.3s] step 2950/4000: task 0.1033 (mse 0.103 cos 0.947); quiet A 0.00117 off 0.00063; 1,497 sites; 3.48 s/step; the job: 183.4 min spent, about 60.7 min of steps left; card peak 14.5 GB
127
+ [11175.4s] step 3000/4000: task 0.0991 (mse 0.099 cos 0.950); quiet A 0.00102 off 0.00045; 1,480 sites; 3.48 s/step; the job: 186.3 min spent, about 57.9 min of steps left; card peak 14.5 GB
128
+ [11175.6s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXA_o0_close3_s10_qwen.safetensors
129
+ [11381.0s] 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
+ [11381.0s] block 8: 0.733 -> 0.946 / 0.946; A 0.9939; mse 0.092 cos 0.953
131
+ [11381.0s] block 12: 0.640 -> 0.944 / 0.946; A 0.9952; mse 0.078 cos 0.962
132
+ [11381.0s] block 16: 0.532 -> 0.953 / 0.953; A 0.9937; mse 0.065 cos 0.969
133
+ [11381.0s] block 18: 0.502 -> 0.956 / 0.957; A 0.9966; mse 0.062 cos 0.971
134
+ [11381.0s] block 20: 0.486 -> 0.958 / 0.958; A 0.9958; mse 0.060 cos 0.972
135
+ [11381.0s] block 24: 0.457 -> 0.959 / 0.958; A 0.9958; mse 0.059 cos 0.974
136
+ [11381.0s] block 28: 0.407 -> 0.949 / 0.949; A 0.9941; mse 0.076 cos 0.967
137
+ [11381.0s] block 31: 0.322 -> 0.873 / 0.876; A 0.9733; mse 0.399 cos 0.950
138
+ [11381.0s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.946 / 0.947 / 0.953 / 0.958 / 0.946
139
+ [11381.0s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.873 against the bar .85; B2: the lowest A alignment 0.9733 against .995; fineweb share +0.00049 +- 0.00019 against +.012; partners' shares s1 +0.0001, s2 -0.0000, s3 +0.0007, s4 +0.0004, s5 +0.0001, s6 +0.0000, s7 +0.0000, s8 +0.0000
140
+ [11381.0s] gate mean 0.151 / widest 0.396; 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 13.6 GB
141
+ [11554.4s] step 3050/4000: task 0.0948 (mse 0.095 cos 0.952); quiet A 0.00098 off 0.00054; 1,458 sites; 3.47 s/step; the job: 192.6 min spent, about 55.0 min of steps left; card peak 14.5 GB
142
+ [11727.5s] step 3100/4000: task 0.0954 (mse 0.095 cos 0.952); quiet A 0.00097 off 0.00069; 1,561 sites; 3.46 s/step; the job: 195.5 min spent, about 52.1 min of steps left; card peak 14.5 GB
143
+ [11900.7s] step 3150/4000: task 0.0920 (mse 0.092 cos 0.953); quiet A 0.00095 off 0.00042; 1,440 sites; 3.46 s/step; the job: 198.3 min spent, about 49.2 min of steps left; card peak 14.5 GB
144
+ [12073.7s] step 3200/4000: task 0.0884 (mse 0.088 cos 0.955); quiet A 0.00092 off 0.00039; 1,519 sites; 3.46 s/step; the job: 201.2 min spent, about 46.3 min of steps left; card peak 14.5 GB
145
+ [12246.9s] step 3250/4000: task 0.0852 (mse 0.085 cos 0.956); quiet A 0.00088 off 0.00046; 1,475 sites; 3.46 s/step; the job: 204.1 min spent, about 43.4 min of steps left; card peak 14.5 GB
146
+ [12420.1s] step 3300/4000: task 0.0857 (mse 0.086 cos 0.956); quiet A 0.00092 off 0.00351; 1,474 sites; 3.46 s/step; the job: 207.0 min spent, about 40.5 min of steps left; card peak 14.5 GB
147
+ [12593.5s] step 3350/4000: task 0.0901 (mse 0.090 cos 0.955); quiet A 0.00091 off 0.00045; 1,541 sites; 3.46 s/step; the job: 209.9 min spent, about 37.6 min of steps left; card peak 14.5 GB
148
+ [12766.8s] step 3400/4000: task 0.0809 (mse 0.081 cos 0.958); quiet A 0.00088 off 0.00039; 1,520 sites; 3.46 s/step; the job: 212.8 min spent, about 34.7 min of steps left; card peak 14.5 GB
149
+ [12940.5s] step 3450/4000: task 0.0798 (mse 0.080 cos 0.959); quiet A 0.00086 off 0.00040; 1,454 sites; 3.47 s/step; the job: 215.7 min spent, about 31.8 min of steps left; card peak 14.5 GB
150
+ [13113.7s] step 3500/4000: task 0.0758 (mse 0.076 cos 0.961); quiet A 0.00084 off 0.00039; 1,505 sites; 3.47 s/step; the job: 218.6 min spent, about 28.9 min of steps left; card peak 14.5 GB
151
+ [13287.2s] step 3550/4000: task 0.0775 (mse 0.077 cos 0.960); quiet A 0.00085 off 0.00104; 1,525 sites; 3.47 s/step; the job: 221.5 min spent, about 26.0 min of steps left; card peak 14.5 GB
152
+ [13460.6s] step 3600/4000: task 0.0761 (mse 0.076 cos 0.961); quiet A 0.00085 off 0.00040; 1,519 sites; 3.47 s/step; the job: 224.3 min spent, about 23.1 min of steps left; card peak 14.5 GB
153
+ [13634.2s] step 3650/4000: task 0.0770 (mse 0.077 cos 0.961); quiet A 0.00084 off 0.00038; 1,472 sites; 3.47 s/step; the job: 227.2 min spent, about 20.2 min of steps left; card peak 14.5 GB
154
+ [13807.8s] step 3700/4000: task 0.0722 (mse 0.072 cos 0.963); quiet A 0.00082 off 0.00038; 1,463 sites; 3.47 s/step; the job: 230.1 min spent, about 17.4 min of steps left; card peak 14.5 GB
155
+ [13981.6s] step 3750/4000: task 0.0717 (mse 0.072 cos 0.963); quiet A 0.00081 off 0.00047; 1,570 sites; 3.47 s/step; the job: 233.0 min spent, about 14.5 min of steps left; card peak 14.5 GB
156
+ [14155.5s] step 3800/4000: task 0.0717 (mse 0.072 cos 0.963); quiet A 0.00079 off 0.00036; 1,532 sites; 3.47 s/step; the job: 235.9 min spent, about 11.6 min of steps left; card peak 14.5 GB
157
+ [14330.0s] step 3850/4000: task 0.0696 (mse 0.070 cos 0.964); quiet A 0.00080 off 0.00039; 1,507 sites; 3.47 s/step; the job: 238.8 min spent, about 8.7 min of steps left; card peak 14.5 GB
158
+ [14503.3s] step 3900/4000: task 0.0686 (mse 0.069 cos 0.965); quiet A 0.00079 off 0.00034; 1,435 sites; 3.47 s/step; the job: 241.7 min spent, about 5.8 min of steps left; card peak 14.5 GB
159
+ [14677.1s] step 3950/4000: task 0.0682 (mse 0.068 cos 0.965); quiet A 0.00076 off 0.00040; 1,500 sites; 3.47 s/step; the job: 244.6 min spent, about 2.9 min of steps left; card peak 14.5 GB
160
+ [14850.8s] step 4000/4000: task 0.0692 (mse 0.069 cos 0.964); quiet A 0.00077 off 0.00034; 1,494 sites; 3.47 s/step; the job: 247.5 min spent, about 0.0 min of steps left; card peak 14.5 GB
161
+ [14851.0s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_gXA_o0_close4_s10_qwen.safetensors
162
+ [15056.7s] 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
+ [15056.7s] block 8: 0.733 -> 0.955 / 0.954; A 0.9977; mse 0.069 cos 0.965
164
+ [15056.7s] block 12: 0.640 -> 0.957 / 0.958; A 0.9966; mse 0.058 cos 0.972
165
+ [15056.7s] block 16: 0.532 -> 0.962 / 0.963; A 0.9949; mse 0.049 cos 0.977
166
+ [15056.7s] block 18: 0.502 -> 0.967 / 0.966; A 0.9976; mse 0.046 cos 0.978
167
+ [15056.7s] block 20: 0.486 -> 0.969 / 0.968; A 0.9971; mse 0.045 cos 0.979
168
+ [15056.7s] block 24: 0.457 -> 0.968 / 0.967; A 0.9973; mse 0.044 cos 0.980
169
+ [15056.7s] block 28: 0.407 -> 0.960 / 0.959; A 0.9963; mse 0.058 cos 0.975
170
+ [15056.7s] block 31: 0.322 -> 0.895 / 0.895; A 0.9835; mse 0.314 cos 0.960
171
+ [15056.7s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.955 / 0.960 / 0.964 / 0.967 / 0.958
172
+ [15056.7s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.895 against the bar .85; B2: the lowest A alignment 0.9835 against .995; fineweb share +0.00030 +- 0.00019 against +.012; partners' shares s1 +0.0000, s2 -0.0000, s3 +0.0005, s4 +0.0002, s5 +0.0001, s6 +0.0000, s7 +0.0001, s8 +0.0000
173
+ [15056.7s] gate mean 0.163 / widest 0.477; 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 13.6 GB
174
+ [15056.7s] DONE qarm_mse_gXA_o0: 4000 steps in 231.4 min of steps (3.47 s/step); 4 closes; 15057 s in all; card peak 14.5 GB; host RSS 13.6 GB; compiled chain: 15 cache entries (limit 256), 0 limit warnings