Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
6a3e504
·
verified ·
1 Parent(s): 2467484

surface arm qwen3/mse_solo_o0_bytes: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/qwen3/mse_solo_o0_bytes/closes/qarm_mse_solo_o0_bytes_close1_s14_qwen_bytes.safetensors +3 -0
  2. surface/qwen3/mse_solo_o0_bytes/closes/qarm_mse_solo_o0_bytes_close2_s14_qwen_bytes.safetensors +3 -0
  3. surface/qwen3/mse_solo_o0_bytes/closes/qarm_mse_solo_o0_bytes_close3_s14_qwen_bytes.safetensors +3 -0
  4. surface/qwen3/mse_solo_o0_bytes/closes/qarm_mse_solo_o0_bytes_close4_s14_qwen_bytes.safetensors +3 -0
  5. surface/qwen3/mse_solo_o0_bytes/manifest.json +20 -0
  6. surface/qwen3/mse_solo_o0_bytes/readout.md +127 -0
  7. surface/qwen3/mse_solo_o0_bytes/results.json +0 -0
  8. surface/qwen3/mse_solo_o0_bytes/s14_qwen_bytes.safetensors +3 -0
  9. surface/qwen3/mse_solo_o0_bytes/stats.pt +3 -0
  10. surface/qwen3/mse_solo_o0_bytes/train.log +159 -0
surface/qwen3/mse_solo_o0_bytes/closes/qarm_mse_solo_o0_bytes_close1_s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ede23d492ed3117e69583777361aec2dd639543b3a2e46bae46e8cf1e9b05da4
3
+ size 54819344
surface/qwen3/mse_solo_o0_bytes/closes/qarm_mse_solo_o0_bytes_close2_s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ba3f0681ebb4037208aa00582deff6af24719d7c4c0b598a3c69e73f9e61e97c
3
+ size 54819344
surface/qwen3/mse_solo_o0_bytes/closes/qarm_mse_solo_o0_bytes_close3_s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aa9886cd97f70e00987ef2eae4f5e19dbeed2dfb6dde812f33d05dbdb5de0fc3
3
+ size 54819344
surface/qwen3/mse_solo_o0_bytes/closes/qarm_mse_solo_o0_bytes_close4_s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ec52760e07a1cfaad4dda5d491ec6e739b4297a6f821e0c56132238fad6f7037
3
+ size 54819344
surface/qwen3/mse_solo_o0_bytes/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "qarm_mse_solo_o0_bytes",
3
+ "run": "mse_solo_o0_bytes",
4
+ "convention": "qwen3",
5
+ "member": "s14_qwen_bytes",
6
+ "final_anchor": "qarm_mse_solo_o0_bytes_close4_s14_qwen_bytes.safetensors",
7
+ "closes": [
8
+ "qarm_mse_solo_o0_bytes_close1_s14_qwen_bytes.safetensors",
9
+ "qarm_mse_solo_o0_bytes_close2_s14_qwen_bytes.safetensors",
10
+ "qarm_mse_solo_o0_bytes_close3_s14_qwen_bytes.safetensors",
11
+ "qarm_mse_solo_o0_bytes_close4_s14_qwen_bytes.safetensors"
12
+ ],
13
+ "finished": true,
14
+ "form": "mse",
15
+ "base": "solo",
16
+ "steps": 4000,
17
+ "trunk_step": 245674,
18
+ "shipped": "2026-10-08 23:57 local",
19
+ "note": "every close's anchor ships (mids included); the final close's anchor is the run's s14_qwen_bytes.safetensors"
20
+ }
surface/qwen3/mse_solo_o0_bytes/readout.md ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # The surface arm's readout: Qwen3's spelling (the GPT-2 byte map)
2
+
3
+ A surface arm makes the trunk read a tokenizer's spelling of a text as it reads the text's own bytes. Every number below is read on the dual-extraction read's held-out captions (512 of each COCO draw, never trained on; draw 2 is the repeat), one document a row, at every token's closing byte, as whitened Procrustes alignment (1 = the same geometry up to a rotation; two folds by caption). The pre-registered bars are in the plan; B3 and B4 (against Qwen's own states, the mood axis) are the instrument's read.
4
+
5
+ ## The runs
6
+
7
+ - qarm_mse_solo_o0_bytes: per-site MSE; alone on the bare trunk; blocks 0-31; 4000 of 4000 steps; seed offset 0
8
+
9
+ ## The verdicts (the final close of each run)
10
+
11
+ ### qarm_mse_solo_o0_bytes
12
+
13
+ - B1 the gap: MET over every served block >= 12 (the lowest 0.963 at block 12 on draw 1, 0.963 at block 12 on draw 2; bar 0.85)
14
+ - B2 silence, whitened: NOT MET (the lowest A alignment 0.9976 at block 6 on draw 1, 0.9947 on draw 2; bar 0.995)
15
+ - B2 silence, in bits: MET (web +0.00008 bpb; bar +0.012)
16
+
17
+ ## The closes
18
+
19
+ ### qarm_mse_solo_o0_bytes
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 | 0.499 / 0.513 | 0.688 / 0.580 / 0.598 | 1.0000 / 1.0000 | +0.00000 | 0.047 / 0.047 |
24
+ | 1 | 1000 | 0.896 / 0.900 | 0.897 / 0.915 / 0.940 | 0.9922 / 0.9918 | +0.00005 | 0.135 / 0.169 |
25
+ | 2 | 2000 | 0.943 / 0.946 | 0.943 / 0.962 / 0.973 | 0.9912 / 0.9909 | +0.00031 | 0.160 / 0.301 |
26
+ | 3 | 3000 | 0.958 / 0.958 | 0.958 / 0.972 / 0.980 | 0.9950 / 0.9933 | +0.00028 | 0.180 / 0.414 |
27
+ | 4 | 4000 | 0.963 / 0.963 | 0.963 / 0.977 / 0.984 | 0.9976 / 0.9947 | +0.00008 | 0.199 / 0.485 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### qarm_mse_solo_o0_bytes, draw1 (512 captions, 25,597 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 | 0.878 | 0.878 | 0.879 | 0.9998 | 0.324 | 0.828 |
36
+ | 4 | 0.826 | 0.941 | 0.942 | 0.9989 | 0.082 | 0.954 |
37
+ | 8 | 0.766 | 0.948 | 0.949 | 0.9989 | 0.066 | 0.963 |
38
+ | 12 | 0.688 | 0.963 | 0.963 | 0.9987 | 0.054 | 0.971 |
39
+ | 16 | 0.628 | 0.973 | 0.972 | 0.9990 | 0.043 | 0.978 |
40
+ | 18 | 0.599 | 0.974 | 0.974 | 0.9988 | 0.039 | 0.980 |
41
+ | 20 | 0.580 | 0.977 | 0.977 | 0.9982 | 0.036 | 0.982 |
42
+ | 22 | 0.589 | 0.981 | 0.981 | 0.9989 | 0.031 | 0.985 |
43
+ | 24 | 0.605 | 0.984 | 0.983 | 0.9992 | 0.028 | 0.987 |
44
+ | 26 | 0.610 | 0.988 | 0.988 | 0.9992 | 0.027 | 0.987 |
45
+ | 28 | 0.598 | 0.984 | 0.985 | 0.9991 | 0.028 | 0.987 |
46
+ | 29 | 0.586 | 0.986 | 0.986 | 0.9990 | 0.028 | 0.986 |
47
+ | 30 | 0.550 | 0.982 | 0.983 | 0.9981 | 0.028 | 0.985 |
48
+ | 31 | 0.499 | 0.981 | 0.981 | 0.9977 | 0.028 | 0.984 |
49
+
50
+ ### qarm_mse_solo_o0_bytes, draw2 (512 captions, 26,687 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 | 0.873 | 0.874 | 0.874 | 0.9998 | 0.324 | 0.827 |
55
+ | 4 | 0.833 | 0.940 | 0.941 | 0.9988 | 0.081 | 0.954 |
56
+ | 8 | 0.773 | 0.952 | 0.952 | 0.9989 | 0.066 | 0.963 |
57
+ | 12 | 0.698 | 0.963 | 0.964 | 0.9988 | 0.053 | 0.971 |
58
+ | 16 | 0.641 | 0.974 | 0.974 | 0.9988 | 0.042 | 0.978 |
59
+ | 18 | 0.615 | 0.973 | 0.974 | 0.9985 | 0.039 | 0.980 |
60
+ | 20 | 0.595 | 0.980 | 0.980 | 0.9980 | 0.036 | 0.982 |
61
+ | 22 | 0.604 | 0.983 | 0.983 | 0.9990 | 0.032 | 0.985 |
62
+ | 24 | 0.620 | 0.982 | 0.982 | 0.9990 | 0.029 | 0.986 |
63
+ | 26 | 0.621 | 0.985 | 0.985 | 0.9991 | 0.027 | 0.987 |
64
+ | 28 | 0.607 | 0.986 | 0.986 | 0.9990 | 0.028 | 0.986 |
65
+ | 29 | 0.596 | 0.982 | 0.982 | 0.9990 | 0.029 | 0.986 |
66
+ | 30 | 0.563 | 0.983 | 0.984 | 0.9986 | 0.029 | 0.985 |
67
+ | 31 | 0.513 | 0.978 | 0.979 | 0.9975 | 0.029 | 0.983 |
68
+
69
+ The member's share in bits per byte (all on minus the member masked): web +0.00008
70
+
71
+ ## The training curves
72
+
73
+ ### qarm_mse_solo_o0_bytes
74
+
75
+ | step | task loss | mse | cos | InfoNCE | quiet on A | quiet off-domain |
76
+ |---|---|---|---|---|---|---|
77
+ | 1 | 1.0383 | 1.038 | 0.504 | None | 0.00000 | 0.00000 |
78
+ | 50 | 1.0354 | 1.035 | 0.504 | None | 0.00031 | 0.00024 |
79
+ | 250 | 0.8816 | 0.882 | 0.571 | None | 0.00093 | 0.00046 |
80
+ | 500 | 0.6587 | 0.659 | 0.671 | None | 0.00071 | 0.00037 |
81
+ | 1000 | 0.3259 | 0.326 | 0.833 | None | 0.00066 | 0.00039 |
82
+ | 1500 | 0.2047 | 0.205 | 0.894 | None | 0.00075 | 0.00052 |
83
+ | 2000 | 0.1290 | 0.129 | 0.932 | None | 0.00083 | 0.00057 |
84
+ | 2500 | 0.0878 | 0.088 | 0.953 | None | 0.00078 | 0.00050 |
85
+ | 3000 | 0.0733 | 0.073 | 0.961 | None | 0.00076 | 0.00057 |
86
+ | 3500 | 0.0615 | 0.062 | 0.967 | None | 0.00062 | 0.00077 |
87
+ | 4000 | 0.0575 | 0.058 | 0.969 | None | 0.00061 | 0.00040 |
88
+ | 1 | 1.0383 | 1.038 | 0.504 | None | 0.00000 | 0.00000 |
89
+ | 50 | 1.0354 | 1.035 | 0.504 | None | 0.00031 | 0.00024 |
90
+ | 250 | 0.8816 | 0.882 | 0.571 | None | 0.00093 | 0.00046 |
91
+ | 500 | 0.6587 | 0.659 | 0.671 | None | 0.00071 | 0.00037 |
92
+ | 1000 | 0.3259 | 0.326 | 0.833 | None | 0.00066 | 0.00039 |
93
+ | 1500 | 0.2047 | 0.205 | 0.894 | None | 0.00075 | 0.00052 |
94
+ | 2000 | 0.1290 | 0.129 | 0.932 | None | 0.00083 | 0.00057 |
95
+ | 2500 | 0.0878 | 0.088 | 0.953 | None | 0.00078 | 0.00050 |
96
+ | 3000 | 0.0733 | 0.073 | 0.961 | None | 0.00076 | 0.00057 |
97
+ | 3500 | 0.0615 | 0.062 | 0.967 | None | 0.00062 | 0.00077 |
98
+ | 4000 | 0.0575 | 0.058 | 0.969 | None | 0.00061 | 0.00040 |
99
+ | 1 | 1.0383 | 1.038 | 0.504 | None | 0.00000 | 0.00000 |
100
+ | 50 | 1.0354 | 1.035 | 0.504 | None | 0.00031 | 0.00024 |
101
+ | 250 | 0.8816 | 0.882 | 0.571 | None | 0.00093 | 0.00046 |
102
+ | 500 | 0.6587 | 0.659 | 0.671 | None | 0.00071 | 0.00037 |
103
+ | 1000 | 0.3259 | 0.326 | 0.833 | None | 0.00066 | 0.00039 |
104
+ | 1500 | 0.2047 | 0.205 | 0.894 | None | 0.00075 | 0.00052 |
105
+ | 2000 | 0.1290 | 0.129 | 0.932 | None | 0.00083 | 0.00057 |
106
+ | 2500 | 0.0878 | 0.088 | 0.953 | None | 0.00078 | 0.00050 |
107
+ | 3000 | 0.0733 | 0.073 | 0.961 | None | 0.00076 | 0.00057 |
108
+ | 3500 | 0.0615 | 0.062 | 0.967 | None | 0.00062 | 0.00077 |
109
+ | 4000 | 0.0575 | 0.058 | 0.969 | None | 0.00061 | 0.00040 |
110
+ | 1 | 1.0383 | 1.038 | 0.504 | None | 0.00000 | 0.00000 |
111
+ | 50 | 1.0354 | 1.035 | 0.504 | None | 0.00031 | 0.00024 |
112
+ | 250 | 0.8816 | 0.882 | 0.571 | None | 0.00093 | 0.00046 |
113
+ | 500 | 0.6587 | 0.659 | 0.671 | None | 0.00071 | 0.00037 |
114
+ | 1000 | 0.3259 | 0.326 | 0.833 | None | 0.00066 | 0.00039 |
115
+ | 1500 | 0.2047 | 0.205 | 0.894 | None | 0.00075 | 0.00052 |
116
+ | 2000 | 0.1290 | 0.129 | 0.932 | None | 0.00083 | 0.00057 |
117
+ | 2500 | 0.0878 | 0.088 | 0.953 | None | 0.00078 | 0.00050 |
118
+ | 3000 | 0.0733 | 0.073 | 0.961 | None | 0.00076 | 0.00057 |
119
+ | 3500 | 0.0615 | 0.062 | 0.967 | None | 0.00062 | 0.00077 |
120
+ | 4000 | 0.0575 | 0.058 | 0.969 | None | 0.00061 | 0.00040 |
121
+
122
+ ## The recipe
123
+
124
+ - per step: the teacher (member masked) reads the A rows: block outputs at the sites + logits; the armed model reads the B rows: per-site MSE of standardized LayerNorm'd block outputs against the teacher's at each token's closing byte, mean over served blocks; quiet on A: lambda x KL(teacher || armed) on the A rows; quiet off-domain: lambda x KL(member masked || armed) on one web row + one partner row
125
+ - lambda 2.0; pure Adam lr 0.001 wd 0; 2 x 4096-byte paired rows a step; the arm LAWFUL_16x8 at every block ({'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}); the trunk mini-beatrix-3 at step 245,674; alephllm 0.10.7, amoe-lora 0.2.11, torch 2.8.0+cu128
126
+ - the surface: Qwen3's byte-level BPE spelling (tokenizer.json sha256 be75606093db2094..., vocab 151,669); documents = one caption-like field of a caption-pack row (caption, short, long, tags, medium), sources drawn by {'dff': 0.161, 'scf': 0.161, 'danbooru': 0.164, 'cc12m': 0.406, 'coco': 0.108}
127
+ - the test set: the first 512 captions of each draw of coco_caps_2x2048.json (every caption of the file kept out of training)
surface/qwen3/mse_solo_o0_bytes/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/qwen3/mse_solo_o0_bytes/s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ec52760e07a1cfaad4dda5d491ec6e739b4297a6f821e0c56132238fad6f7037
3
+ size 54819344
surface/qwen3/mse_solo_o0_bytes/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7afaf713671b6189a487e362c2cf6ea819fcbd73285982a81896e5712f61ffe8
3
+ size 264333
surface/qwen3/mse_solo_o0_bytes/train.log ADDED
@@ -0,0 +1,159 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 0.3s] THE JOB: qarm_mse_solo_o0_bytes: the surface arm s14_qwen_bytes (gpt2 spelling by /workspace/qarm/dl/qwen3_tokenizer; A = the text written; sites = bytes) alone on the bare trunk; form mse; blocks all; 4000 steps, a close every 1000; lambda 2.0; pure Adam lr 0.001; the compiled chain with recompute; card 30.5 GB with 30.4 GB free; host RSS 0.6 GB
2
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
3
+ [ 17.4s] 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
+ [ 18.0s] 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, 2, 3), (1, 3, 4), (1, 4, 5), (1, 5, 6), (1, 6, 7), (2, 7, 9), (2, 8, 10), (2, 9, 11), (2, 10, 12), (2, 11, 13)]
5
+ [ 18.0s] 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)
6
+ [ 31.4s] caption pack (/workspace/qarm/dl/mini-beatrix-3/caption_pack/pack_v2, train): dff 121,153 rows, scf 42,080 rows, danbooru 603,026 rows, cc12m 1,242,176 rows, coco 19,227 rows; host RSS 3.7 GB
7
+ [ 31.6s] streams: web text; partners none (two web rows a quiet chunk); held-out sets ['web'] x 8 batches
8
+ [ 31.7s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 116), 25,597 sites
9
+ [ 31.8s] read rows draw2: 512 captions (0 unspellable), A rows (512, 151), B rows (512, 176), 26,687 sites
10
+ [arms] attached 's14_qwen_bytes' for phase 'qwen_spelling': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
11
+ [ 52.7s] BIRTH: the member attached over the trunk: web rows read exactly as before; the hand-run trunk against the model's forward: armed 0.0e+00, masked 0.0e+00 (gates: 0.0)
12
+ [ 68.5s] standardization statistics from 55,408 teacher sites on 8 paired batches (a mean and a spread per block and dimension); sigma range at block 20 0.779..1.090; saved /workspace/qarm/out/anchors/qarm_mse_solo_o0_bytes_stats.pt
13
+ [ 515.9s] CLOSE 0 (step 0): the member's alignment read on draw 1 (512 captions, 25,597 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
14
+ [ 515.9s] block 8: 0.766 -> 0.766 / 0.766; A 1.0000; mse 0.834 cos 0.610
15
+ [ 515.9s] block 12: 0.688 -> 0.688 / 0.688; A 1.0000; mse 0.981 cos 0.530
16
+ [ 515.9s] block 16: 0.628 -> 0.628 / 0.628; A 1.0000; mse 1.093 cos 0.466
17
+ [ 515.9s] block 18: 0.599 -> 0.599 / 0.599; A 1.0000; mse 1.140 cos 0.439
18
+ [ 515.9s] block 20: 0.580 -> 0.580 / 0.580; A 1.0000; mse 1.189 cos 0.418
19
+ [ 515.9s] block 24: 0.605 -> 0.605 / 0.605; A 1.0000; mse 1.215 cos 0.422
20
+ [ 515.9s] block 28: 0.598 -> 0.598 / 0.598; A 1.0000; mse 1.262 cos 0.375
21
+ [ 515.9s] block 31: 0.499 -> 0.499 / 0.499; A 1.0000; mse 1.008 cos 0.321
22
+ [ 515.9s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.698 / 0.641 / 0.595 / 0.620 / 0.607
23
+ [ 515.9s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.499 against the bar .85; B2: the lowest A alignment 1.0000 against .995; fineweb share +0.00000 +- 0.00000 against +.012
24
+ [ 515.9s] gate mean 0.047 / widest 0.047; compiled chain: 2 cache entries (limit 256), 0 limit warnings; docs drawn 256 (skipped {'test': 2, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 3.6 GB; host RSS 5.3 GB
25
+ [stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
26
+ [ 531.1s] step 1/4000: task 1.0383 (mse 1.038 cos 0.504); quiet A 0.00000 off 0.00000; 6,930 sites; 15.22 s/step; the job: 8.9 min spent, about 1014.7 min of steps left; card peak 16.5 GB
27
+ [ 655.9s] step 50/4000: task 1.0354 (mse 1.035 cos 0.504); quiet A 0.00031 off 0.00024; 6,909 sites; 2.80 s/step; the job: 10.9 min spent, about 184.3 min of steps left; card peak 16.6 GB
28
+ [ 783.2s] step 100/4000: task 1.0078 (mse 1.008 cos 0.516); quiet A 0.00051 off 0.00034; 6,940 sites; 2.67 s/step; the job: 13.1 min spent, about 173.7 min of steps left; card peak 16.6 GB
29
+ [ 910.6s] step 150/4000: task 0.9703 (mse 0.970 cos 0.532); quiet A 0.00060 off 0.00038; 6,900 sites; 2.63 s/step; the job: 15.2 min spent, about 168.9 min of steps left; card peak 16.6 GB
30
+ [ 1037.8s] step 200/4000: task 0.9248 (mse 0.925 cos 0.553); quiet A 0.00076 off 0.00048; 6,952 sites; 2.61 s/step; the job: 17.3 min spent, about 165.3 min of steps left; card peak 16.6 GB
31
+ [ 1165.1s] step 250/4000: task 0.8816 (mse 0.882 cos 0.571); quiet A 0.00093 off 0.00046; 6,908 sites; 2.60 s/step; the job: 19.4 min spent, about 162.3 min of steps left; card peak 16.6 GB
32
+ [ 1292.2s] step 300/4000: task 0.8417 (mse 0.842 cos 0.589); quiet A 0.00100 off 0.00041; 6,907 sites; 2.59 s/step; the job: 21.5 min spent, about 159.6 min of steps left; card peak 16.6 GB
33
+ [ 1419.3s] step 350/4000: task 0.7956 (mse 0.796 cos 0.610); quiet A 0.00093 off 0.00037; 6,920 sites; 2.58 s/step; the job: 23.7 min spent, about 157.0 min of steps left; card peak 16.6 GB
34
+ [ 1546.7s] step 400/4000: task 0.7570 (mse 0.757 cos 0.626); quiet A 0.00078 off 0.00040; 6,921 sites; 2.58 s/step; the job: 25.8 min spent, about 154.6 min of steps left; card peak 16.6 GB
35
+ [ 1673.8s] step 450/4000: task 0.7114 (mse 0.711 cos 0.647); quiet A 0.00075 off 0.00037; 6,941 sites; 2.57 s/step; the job: 27.9 min spent, about 152.2 min of steps left; card peak 16.6 GB
36
+ [ 1800.8s] step 500/4000: task 0.6587 (mse 0.659 cos 0.671); quiet A 0.00071 off 0.00037; 6,985 sites; 2.57 s/step; the job: 30.0 min spent, about 149.9 min of steps left; card peak 16.6 GB
37
+ [ 1928.1s] step 550/4000: task 0.5977 (mse 0.598 cos 0.699); quiet A 0.00071 off 0.00040; 6,940 sites; 2.57 s/step; the job: 32.1 min spent, about 147.6 min of steps left; card peak 16.6 GB
38
+ [ 2055.2s] step 600/4000: task 0.5246 (mse 0.525 cos 0.735); quiet A 0.00071 off 0.00049; 6,908 sites; 2.57 s/step; the job: 34.3 min spent, about 145.4 min of steps left; card peak 16.6 GB
39
+ [ 2182.9s] step 650/4000: task 0.4702 (mse 0.470 cos 0.763); quiet A 0.00069 off 0.00038; 6,956 sites; 2.56 s/step; the job: 36.4 min spent, about 143.2 min of steps left; card peak 16.6 GB
40
+ [ 2310.2s] step 700/4000: task 0.4355 (mse 0.436 cos 0.779); quiet A 0.00064 off 0.00037; 6,894 sites; 2.56 s/step; the job: 38.5 min spent, about 141.0 min of steps left; card peak 16.6 GB
41
+ [ 2437.5s] step 750/4000: task 0.4082 (mse 0.408 cos 0.793); quiet A 0.00063 off 0.00042; 6,919 sites; 2.56 s/step; the job: 40.6 min spent, about 138.8 min of steps left; card peak 16.6 GB
42
+ [ 2564.7s] step 800/4000: task 0.3869 (mse 0.387 cos 0.803); quiet A 0.00064 off 0.00037; 6,932 sites; 2.56 s/step; the job: 42.7 min spent, about 136.6 min of steps left; card peak 16.6 GB
43
+ [ 2691.9s] step 850/4000: task 0.3697 (mse 0.370 cos 0.811); quiet A 0.00065 off 0.00036; 6,943 sites; 2.56 s/step; the job: 44.9 min spent, about 134.4 min of steps left; card peak 16.6 GB
44
+ [ 2819.2s] step 900/4000: task 0.3496 (mse 0.350 cos 0.821); quiet A 0.00076 off 0.00040; 7,215 sites; 2.56 s/step; the job: 47.0 min spent, about 132.2 min of steps left; card peak 16.6 GB
45
+ [ 2946.6s] step 950/4000: task 0.3360 (mse 0.336 cos 0.828); quiet A 0.00069 off 0.00038; 6,932 sites; 2.56 s/step; the job: 49.1 min spent, about 130.1 min of steps left; card peak 16.6 GB
46
+ [ 3073.8s] step 1000/4000: task 0.3259 (mse 0.326 cos 0.833); quiet A 0.00066 off 0.00039; 6,866 sites; 2.56 s/step; the job: 51.2 min spent, about 127.9 min of steps left; card peak 16.6 GB
47
+ [ 3074.0s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o0_bytes_close1_s14_qwen_bytes.safetensors
48
+ [ 3408.7s] CLOSE 1 (step 1000): the member's alignment read on draw 1 (512 captions, 25,597 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
49
+ [ 3408.7s] block 8: 0.766 -> 0.883 / 0.886; A 0.9961; mse 0.256 cos 0.861
50
+ [ 3408.7s] block 12: 0.688 -> 0.897 / 0.899; A 0.9922; mse 0.205 cos 0.889
51
+ [ 3408.7s] block 16: 0.628 -> 0.911 / 0.913; A 0.9960; mse 0.168 cos 0.913
52
+ [ 3408.7s] block 18: 0.599 -> 0.911 / 0.914; A 0.9948; mse 0.161 cos 0.918
53
+ [ 3408.7s] block 20: 0.580 -> 0.915 / 0.918; A 0.9957; mse 0.153 cos 0.923
54
+ [ 3408.7s] block 24: 0.605 -> 0.938 / 0.940; A 0.9974; mse 0.127 cos 0.939
55
+ [ 3408.7s] block 28: 0.598 -> 0.940 / 0.943; A 0.9947; mse 0.123 cos 0.939
56
+ [ 3408.7s] block 31: 0.499 -> 0.924 / 0.926; A 0.9941; mse 0.111 cos 0.933
57
+ [ 3408.7s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.902 / 0.911 / 0.920 / 0.935 / 0.941
58
+ [ 3408.7s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.896 against the bar .85; B2: the lowest A alignment 0.9922 against .995; fineweb share +0.00005 +- 0.00013 against +.012
59
+ [ 3408.7s] gate mean 0.135 / widest 0.169; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 32,523 (skipped {'test': 44, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 16.6 GB; host RSS 8.2 GB
60
+ [ 3535.9s] step 1050/4000: task 0.3104 (mse 0.310 cos 0.840); quiet A 0.00068 off 0.00041; 6,919 sites; 2.54 s/step; the job: 58.9 min spent, about 125.7 min of steps left; card peak 16.6 GB
61
+ [ 3663.5s] step 1100/4000: task 0.2962 (mse 0.296 cos 0.847); quiet A 0.00068 off 0.00037; 6,967 sites; 2.55 s/step; the job: 61.1 min spent, about 123.6 min of steps left; card peak 16.6 GB
62
+ [ 3791.3s] step 1150/4000: task 0.2884 (mse 0.288 cos 0.851); quiet A 0.00070 off 0.00039; 6,942 sites; 2.55 s/step; the job: 63.2 min spent, about 121.5 min of steps left; card peak 16.6 GB
63
+ [ 3918.8s] step 1200/4000: task 0.2738 (mse 0.274 cos 0.858); quiet A 0.00069 off 0.00040; 6,958 sites; 2.55 s/step; the job: 65.3 min spent, about 119.3 min of steps left; card peak 16.6 GB
64
+ [ 4046.4s] step 1250/4000: task 0.2625 (mse 0.263 cos 0.864); quiet A 0.00074 off 0.00078; 6,968 sites; 2.55 s/step; the job: 67.4 min spent, about 117.2 min of steps left; card peak 16.6 GB
65
+ [ 4173.5s] step 1300/4000: task 0.2518 (mse 0.252 cos 0.869); quiet A 0.00075 off 0.00044; 6,944 sites; 2.55 s/step; the job: 69.6 min spent, about 115.0 min of steps left; card peak 16.6 GB
66
+ [ 4301.2s] step 1350/4000: task 0.2393 (mse 0.239 cos 0.876); quiet A 0.00072 off 0.00043; 6,923 sites; 2.55 s/step; the job: 71.7 min spent, about 112.9 min of steps left; card peak 16.6 GB
67
+ [ 4428.4s] step 1400/4000: task 0.2266 (mse 0.227 cos 0.882); quiet A 0.00072 off 0.00045; 6,919 sites; 2.55 s/step; the job: 73.8 min spent, about 110.7 min of steps left; card peak 16.6 GB
68
+ [ 4555.7s] step 1450/4000: task 0.2158 (mse 0.216 cos 0.888); quiet A 0.00074 off 0.00049; 6,968 sites; 2.55 s/step; the job: 75.9 min spent, about 108.6 min of steps left; card peak 16.6 GB
69
+ [ 4683.0s] step 1500/4000: task 0.2047 (mse 0.205 cos 0.894); quiet A 0.00075 off 0.00052; 6,948 sites; 2.55 s/step; the job: 78.1 min spent, about 106.4 min of steps left; card peak 16.6 GB
70
+ [ 4810.3s] step 1550/4000: task 0.1951 (mse 0.195 cos 0.899); quiet A 0.00078 off 0.00052; 6,932 sites; 2.55 s/step; the job: 80.2 min spent, about 104.3 min of steps left; card peak 16.6 GB
71
+ [ 4937.9s] step 1600/4000: task 0.1841 (mse 0.184 cos 0.904); quiet A 0.00079 off 0.00047; 6,941 sites; 2.55 s/step; the job: 82.3 min spent, about 102.2 min of steps left; card peak 16.6 GB
72
+ [ 5065.1s] step 1650/4000: task 0.1760 (mse 0.176 cos 0.908); quiet A 0.00084 off 0.00079; 6,962 sites; 2.55 s/step; the job: 84.4 min spent, about 100.0 min of steps left; card peak 16.6 GB
73
+ [ 5192.5s] step 1700/4000: task 0.1699 (mse 0.170 cos 0.912); quiet A 0.00087 off 0.00056; 6,949 sites; 2.55 s/step; the job: 86.5 min spent, about 97.9 min of steps left; card peak 16.6 GB
74
+ [ 5319.8s] step 1750/4000: task 0.1583 (mse 0.158 cos 0.917); quiet A 0.00086 off 0.00055; 6,916 sites; 2.55 s/step; the job: 88.7 min spent, about 95.8 min of steps left; card peak 16.6 GB
75
+ [ 5447.3s] step 1800/4000: task 0.1521 (mse 0.152 cos 0.921); quiet A 0.00084 off 0.00060; 6,956 sites; 2.55 s/step; the job: 90.8 min spent, about 93.6 min of steps left; card peak 16.6 GB
76
+ [ 5574.3s] step 1850/4000: task 0.1452 (mse 0.145 cos 0.924); quiet A 0.00086 off 0.00054; 6,949 sites; 2.55 s/step; the job: 92.9 min spent, about 91.5 min of steps left; card peak 16.6 GB
77
+ [ 5701.9s] step 1900/4000: task 0.1416 (mse 0.142 cos 0.926); quiet A 0.00088 off 0.00062; 6,967 sites; 2.55 s/step; the job: 95.0 min spent, about 89.4 min of steps left; card peak 16.6 GB
78
+ [ 5829.2s] step 1950/4000: task 0.1368 (mse 0.137 cos 0.928); quiet A 0.00087 off 0.00054; 6,956 sites; 2.55 s/step; the job: 97.2 min spent, about 87.2 min of steps left; card peak 16.6 GB
79
+ [ 5956.5s] step 2000/4000: task 0.1290 (mse 0.129 cos 0.932); quiet A 0.00083 off 0.00057; 6,915 sites; 2.55 s/step; the job: 99.3 min spent, about 85.1 min of steps left; card peak 16.6 GB
80
+ [ 5956.7s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o0_bytes_close2_s14_qwen_bytes.safetensors
81
+ [ 6290.6s] CLOSE 2 (step 2000): the member's alignment read on draw 1 (512 captions, 25,597 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
82
+ [ 6290.6s] block 8: 0.766 -> 0.929 / 0.931; A 0.9977; mse 0.113 cos 0.938
83
+ [ 6290.6s] block 12: 0.688 -> 0.943 / 0.944; A 0.9973; mse 0.091 cos 0.950
84
+ [ 6290.6s] block 16: 0.628 -> 0.958 / 0.958; A 0.9972; mse 0.073 cos 0.962
85
+ [ 6290.6s] block 18: 0.599 -> 0.958 / 0.958; A 0.9967; mse 0.067 cos 0.966
86
+ [ 6290.6s] block 20: 0.580 -> 0.962 / 0.962; A 0.9976; mse 0.062 cos 0.969
87
+ [ 6290.6s] block 24: 0.605 -> 0.975 / 0.976; A 0.9981; mse 0.050 cos 0.976
88
+ [ 6290.6s] block 28: 0.598 -> 0.973 / 0.973; A 0.9978; mse 0.049 cos 0.976
89
+ [ 6290.6s] block 31: 0.499 -> 0.967 / 0.969; A 0.9957; mse 0.048 cos 0.972
90
+ [ 6290.6s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.947 / 0.957 / 0.963 / 0.972 / 0.974
91
+ [ 6290.6s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.943 against the bar .85; B2: the lowest A alignment 0.9912 against .995; fineweb share +0.00031 +- 0.00018 against +.012
92
+ [ 6290.6s] gate mean 0.160 / widest 0.301; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 64,564 (skipped {'test': 97, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 16.6 GB; host RSS 8.3 GB
93
+ [ 6418.2s] step 2050/4000: task 0.1225 (mse 0.122 cos 0.936); quiet A 0.00086 off 0.00056; 6,927 sites; 2.55 s/step; the job: 107.0 min spent, about 83.0 min of steps left; card peak 16.6 GB
94
+ [ 6545.4s] step 2100/4000: task 0.1213 (mse 0.121 cos 0.936); quiet A 0.00087 off 0.00062; 6,959 sites; 2.55 s/step; the job: 109.1 min spent, about 80.8 min of steps left; card peak 16.6 GB
95
+ [ 6672.3s] step 2150/4000: task 0.1139 (mse 0.114 cos 0.940); quiet A 0.00081 off 0.00052; 6,930 sites; 2.54 s/step; the job: 111.2 min spent, about 78.7 min of steps left; card peak 16.6 GB
96
+ [ 6799.5s] step 2200/4000: task 0.1100 (mse 0.110 cos 0.942); quiet A 0.00084 off 0.00051; 6,969 sites; 2.54 s/step; the job: 113.3 min spent, about 76.6 min of steps left; card peak 16.6 GB
97
+ [ 6926.9s] step 2250/4000: task 0.1080 (mse 0.108 cos 0.943); quiet A 0.00085 off 0.00056; 6,986 sites; 2.55 s/step; the job: 115.4 min spent, about 74.4 min of steps left; card peak 16.6 GB
98
+ [ 7054.6s] step 2300/4000: task 0.1023 (mse 0.102 cos 0.946); quiet A 0.00081 off 0.00052; 6,936 sites; 2.55 s/step; the job: 117.6 min spent, about 72.3 min of steps left; card peak 16.6 GB
99
+ [ 7181.9s] step 2350/4000: task 0.1005 (mse 0.101 cos 0.947); quiet A 0.00081 off 0.00051; 6,938 sites; 2.55 s/step; the job: 119.7 min spent, about 70.2 min of steps left; card peak 16.6 GB
100
+ [ 7309.1s] step 2400/4000: task 0.0954 (mse 0.095 cos 0.950); quiet A 0.00078 off 0.00050; 6,829 sites; 2.55 s/step; the job: 121.8 min spent, about 68.0 min of steps left; card peak 16.6 GB
101
+ [ 7436.4s] step 2450/4000: task 0.0910 (mse 0.091 cos 0.952); quiet A 0.00078 off 0.00055; 6,960 sites; 2.55 s/step; the job: 123.9 min spent, about 65.9 min of steps left; card peak 16.6 GB
102
+ [ 7563.7s] step 2500/4000: task 0.0878 (mse 0.088 cos 0.953); quiet A 0.00078 off 0.00050; 6,959 sites; 2.55 s/step; the job: 126.1 min spent, about 63.8 min of steps left; card peak 16.6 GB
103
+ [ 7691.0s] step 2550/4000: task 0.0831 (mse 0.083 cos 0.956); quiet A 0.00076 off 0.00051; 6,910 sites; 2.55 s/step; the job: 128.2 min spent, about 61.7 min of steps left; card peak 16.6 GB
104
+ [ 7818.1s] step 2600/4000: task 0.0872 (mse 0.087 cos 0.954); quiet A 0.00076 off 0.00049; 6,918 sites; 2.55 s/step; the job: 130.3 min spent, about 59.5 min of steps left; card peak 16.6 GB
105
+ [ 7945.6s] step 2650/4000: task 0.0823 (mse 0.082 cos 0.956); quiet A 0.00072 off 0.00056; 6,926 sites; 2.55 s/step; the job: 132.4 min spent, about 57.4 min of steps left; card peak 16.6 GB
106
+ [ 8072.9s] step 2700/4000: task 0.0796 (mse 0.080 cos 0.958); quiet A 0.00072 off 0.00048; 6,926 sites; 2.55 s/step; the job: 134.5 min spent, about 55.3 min of steps left; card peak 16.6 GB
107
+ [ 8199.9s] step 2750/4000: task 0.0779 (mse 0.078 cos 0.959); quiet A 0.00072 off 0.00055; 6,880 sites; 2.55 s/step; the job: 136.7 min spent, about 53.1 min of steps left; card peak 16.6 GB
108
+ [ 8327.2s] step 2800/4000: task 0.0789 (mse 0.079 cos 0.958); quiet A 0.00072 off 0.00057; 6,952 sites; 2.55 s/step; the job: 138.8 min spent, about 51.0 min of steps left; card peak 16.6 GB
109
+ [ 8454.5s] step 2850/4000: task 0.0747 (mse 0.075 cos 0.960); quiet A 0.00070 off 0.00048; 6,969 sites; 2.55 s/step; the job: 140.9 min spent, about 48.9 min of steps left; card peak 16.6 GB
110
+ [ 8581.7s] step 2900/4000: task 0.0777 (mse 0.078 cos 0.959); quiet A 0.00160 off 0.00045; 6,920 sites; 2.55 s/step; the job: 143.0 min spent, about 46.8 min of steps left; card peak 16.6 GB
111
+ [ 8709.0s] step 2950/4000: task 0.0761 (mse 0.076 cos 0.960); quiet A 0.00088 off 0.00055; 6,967 sites; 2.55 s/step; the job: 145.1 min spent, about 44.6 min of steps left; card peak 16.6 GB
112
+ [ 8836.5s] step 3000/4000: task 0.0733 (mse 0.073 cos 0.961); quiet A 0.00076 off 0.00057; 6,944 sites; 2.55 s/step; the job: 147.3 min spent, about 42.5 min of steps left; card peak 16.6 GB
113
+ [ 8836.7s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o0_bytes_close3_s14_qwen_bytes.safetensors
114
+ [ 9171.0s] CLOSE 3 (step 3000): the member's alignment read on draw 1 (512 captions, 25,597 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
115
+ [ 9171.0s] block 8: 0.766 -> 0.941 / 0.942; A 0.9985; mse 0.079 cos 0.956
116
+ [ 9171.0s] block 12: 0.688 -> 0.958 / 0.958; A 0.9972; mse 0.064 cos 0.965
117
+ [ 9171.0s] block 16: 0.628 -> 0.968 / 0.968; A 0.9985; mse 0.052 cos 0.973
118
+ [ 9171.0s] block 18: 0.599 -> 0.971 / 0.970; A 0.9980; mse 0.048 cos 0.976
119
+ [ 9171.0s] block 20: 0.580 -> 0.972 / 0.973; A 0.9968; mse 0.044 cos 0.978
120
+ [ 9171.0s] block 24: 0.605 -> 0.981 / 0.981; A 0.9987; mse 0.035 cos 0.983
121
+ [ 9171.0s] block 28: 0.598 -> 0.980 / 0.981; A 0.9987; mse 0.035 cos 0.983
122
+ [ 9171.0s] block 31: 0.499 -> 0.977 / 0.978; A 0.9972; mse 0.034 cos 0.980
123
+ [ 9171.0s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.959 / 0.968 / 0.973 / 0.979 / 0.983
124
+ [ 9171.0s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.958 against the bar .85; B2: the lowest A alignment 0.9950 against .995; fineweb share +0.00028 +- 0.00007 against +.012
125
+ [ 9171.0s] gate mean 0.180 / widest 0.414; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 96,501 (skipped {'test': 156, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 16.6 GB; host RSS 8.3 GB
126
+ [ 9298.3s] step 3050/4000: task 0.0707 (mse 0.071 cos 0.962); quiet A 0.00071 off 0.00046; 6,968 sites; 2.55 s/step; the job: 155.0 min spent, about 40.4 min of steps left; card peak 16.6 GB
127
+ [ 9425.7s] step 3100/4000: task 0.0704 (mse 0.070 cos 0.962); quiet A 0.00068 off 0.00059; 6,963 sites; 2.55 s/step; the job: 157.1 min spent, about 38.3 min of steps left; card peak 16.6 GB
128
+ [ 9553.0s] step 3150/4000: task 0.0691 (mse 0.069 cos 0.963); quiet A 0.00070 off 0.00044; 6,828 sites; 2.55 s/step; the job: 159.2 min spent, about 36.1 min of steps left; card peak 16.6 GB
129
+ [ 9680.3s] step 3200/4000: task 0.0673 (mse 0.067 cos 0.964); quiet A 0.00067 off 0.00050; 6,943 sites; 2.55 s/step; the job: 161.3 min spent, about 34.0 min of steps left; card peak 16.6 GB
130
+ [ 9807.5s] step 3250/4000: task 0.0659 (mse 0.066 cos 0.965); quiet A 0.00066 off 0.00044; 6,931 sites; 2.55 s/step; the job: 163.5 min spent, about 31.9 min of steps left; card peak 16.6 GB
131
+ [ 9935.3s] step 3300/4000: task 0.0657 (mse 0.066 cos 0.965); quiet A 0.00065 off 0.00045; 6,915 sites; 2.55 s/step; the job: 165.6 min spent, about 29.8 min of steps left; card peak 16.6 GB
132
+ [10062.5s] step 3350/4000: task 0.0692 (mse 0.069 cos 0.963); quiet A 0.00066 off 0.00042; 6,905 sites; 2.55 s/step; the job: 167.7 min spent, about 27.6 min of steps left; card peak 16.6 GB
133
+ [10189.8s] step 3400/4000: task 0.0638 (mse 0.064 cos 0.966); quiet A 0.00065 off 0.00042; 6,937 sites; 2.55 s/step; the job: 169.8 min spent, about 25.5 min of steps left; card peak 16.6 GB
134
+ [10317.5s] step 3450/4000: task 0.0635 (mse 0.063 cos 0.966); quiet A 0.00065 off 0.00054; 6,928 sites; 2.55 s/step; the job: 172.0 min spent, about 23.4 min of steps left; card peak 16.6 GB
135
+ [10444.9s] step 3500/4000: task 0.0615 (mse 0.062 cos 0.967); quiet A 0.00062 off 0.00077; 6,940 sites; 2.55 s/step; the job: 174.1 min spent, about 21.3 min of steps left; card peak 16.6 GB
136
+ [10572.1s] step 3550/4000: task 0.0636 (mse 0.064 cos 0.966); quiet A 0.00072 off 0.00045; 6,906 sites; 2.55 s/step; the job: 176.2 min spent, about 19.1 min of steps left; card peak 16.6 GB
137
+ [10699.3s] step 3600/4000: task 0.0619 (mse 0.062 cos 0.967); quiet A 0.00066 off 0.00043; 6,957 sites; 2.55 s/step; the job: 178.3 min spent, about 17.0 min of steps left; card peak 16.6 GB
138
+ [10826.6s] step 3650/4000: task 0.0621 (mse 0.062 cos 0.967); quiet A 0.00064 off 0.00045; 6,942 sites; 2.55 s/step; the job: 180.4 min spent, about 14.9 min of steps left; card peak 16.6 GB
139
+ [10953.9s] step 3700/4000: task 0.0597 (mse 0.060 cos 0.968); quiet A 0.00063 off 0.00043; 6,949 sites; 2.55 s/step; the job: 182.6 min spent, about 12.7 min of steps left; card peak 16.6 GB
140
+ [11081.5s] step 3750/4000: task 0.0589 (mse 0.059 cos 0.968); quiet A 0.00062 off 0.00048; 6,967 sites; 2.55 s/step; the job: 184.7 min spent, about 10.6 min of steps left; card peak 16.6 GB
141
+ [11208.7s] step 3800/4000: task 0.0592 (mse 0.059 cos 0.968); quiet A 0.00062 off 0.00043; 6,974 sites; 2.55 s/step; the job: 186.8 min spent, about 8.5 min of steps left; card peak 16.6 GB
142
+ [11336.0s] step 3850/4000: task 0.0579 (mse 0.058 cos 0.969); quiet A 0.00063 off 0.00041; 6,947 sites; 2.55 s/step; the job: 188.9 min spent, about 6.4 min of steps left; card peak 16.6 GB
143
+ [11463.3s] step 3900/4000: task 0.0570 (mse 0.057 cos 0.969); quiet A 0.00060 off 0.00041; 6,967 sites; 2.55 s/step; the job: 191.1 min spent, about 4.2 min of steps left; card peak 16.6 GB
144
+ [11590.5s] step 3950/4000: task 0.0574 (mse 0.057 cos 0.969); quiet A 0.00059 off 0.00041; 6,947 sites; 2.55 s/step; the job: 193.2 min spent, about 2.1 min of steps left; card peak 16.6 GB
145
+ [11717.9s] step 4000/4000: task 0.0575 (mse 0.058 cos 0.969); quiet A 0.00061 off 0.00040; 6,944 sites; 2.55 s/step; the job: 195.3 min spent, about 0.0 min of steps left; card peak 16.6 GB
146
+ [11718.1s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o0_bytes_close4_s14_qwen_bytes.safetensors
147
+ [12052.5s] CLOSE 4 (step 4000): the member's alignment read on draw 1 (512 captions, 25,597 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
148
+ [12052.5s] block 8: 0.766 -> 0.948 / 0.949; A 0.9989; mse 0.066 cos 0.963
149
+ [12052.5s] block 12: 0.688 -> 0.963 / 0.963; A 0.9987; mse 0.054 cos 0.971
150
+ [12052.5s] block 16: 0.628 -> 0.973 / 0.972; A 0.9990; mse 0.043 cos 0.978
151
+ [12052.5s] block 18: 0.599 -> 0.974 / 0.974; A 0.9988; mse 0.039 cos 0.980
152
+ [12052.5s] block 20: 0.580 -> 0.977 / 0.977; A 0.9982; mse 0.036 cos 0.982
153
+ [12052.5s] block 24: 0.605 -> 0.984 / 0.983; A 0.9992; mse 0.028 cos 0.987
154
+ [12052.5s] block 28: 0.598 -> 0.984 / 0.985; A 0.9991; mse 0.028 cos 0.987
155
+ [12052.5s] block 31: 0.499 -> 0.981 / 0.981; A 0.9977; mse 0.028 cos 0.984
156
+ [12052.5s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.963 / 0.974 / 0.980 / 0.982 / 0.986
157
+ [12052.5s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.963 against the bar .85; B2: the lowest A alignment 0.9976 against .995; fineweb share +0.00008 +- 0.00009 against +.012
158
+ [12052.5s] gate mean 0.199 / widest 0.485; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 128,688 (skipped {'test': 203, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 16.6 GB; host RSS 8.3 GB
159
+ [12052.6s] DONE qarm_mse_solo_o0_bytes: 4000 steps in 170.0 min of steps (2.55 s/step); 4 closes; 12053 s in all; card peak 16.6 GB; host RSS 8.3 GB; compiled chain: 5 cache entries (limit 256), 0 limit warnings