Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
b73c650
·
verified ·
1 Parent(s): 3e85e91

surface arm qwen3/mse_solo_o1_bytes: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/qwen3/mse_solo_o1_bytes/closes/qarm_mse_solo_o1_bytes_close1_s14_qwen_bytes.safetensors +3 -0
  2. surface/qwen3/mse_solo_o1_bytes/closes/qarm_mse_solo_o1_bytes_close2_s14_qwen_bytes.safetensors +3 -0
  3. surface/qwen3/mse_solo_o1_bytes/closes/qarm_mse_solo_o1_bytes_close3_s14_qwen_bytes.safetensors +3 -0
  4. surface/qwen3/mse_solo_o1_bytes/closes/qarm_mse_solo_o1_bytes_close4_s14_qwen_bytes.safetensors +3 -0
  5. surface/qwen3/mse_solo_o1_bytes/manifest.json +20 -0
  6. surface/qwen3/mse_solo_o1_bytes/readout.md +127 -0
  7. surface/qwen3/mse_solo_o1_bytes/results.json +0 -0
  8. surface/qwen3/mse_solo_o1_bytes/s14_qwen_bytes.safetensors +3 -0
  9. surface/qwen3/mse_solo_o1_bytes/stats.pt +3 -0
  10. surface/qwen3/mse_solo_o1_bytes/train.log +159 -0
surface/qwen3/mse_solo_o1_bytes/closes/qarm_mse_solo_o1_bytes_close1_s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e77705cb0ae810b4345cdf73773856685e543acce9706a0a6e893f3d3f9083e4
3
+ size 54819344
surface/qwen3/mse_solo_o1_bytes/closes/qarm_mse_solo_o1_bytes_close2_s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e9e736c72207660e2fdc2b9657518b3ebde75909fa282f959f5579ce1f001a1f
3
+ size 54819344
surface/qwen3/mse_solo_o1_bytes/closes/qarm_mse_solo_o1_bytes_close3_s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:386d16fdaed2509296c309abba642889cd7c448668a1fee6a5d927faea386702
3
+ size 54819344
surface/qwen3/mse_solo_o1_bytes/closes/qarm_mse_solo_o1_bytes_close4_s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:90925cb2d45b908a30af874d3155fc9e999c46e3fe946940867c1a2c1e10fb8d
3
+ size 54819344
surface/qwen3/mse_solo_o1_bytes/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "qarm_mse_solo_o1_bytes",
3
+ "run": "mse_solo_o1_bytes",
4
+ "convention": "qwen3",
5
+ "member": "s14_qwen_bytes",
6
+ "final_anchor": "qarm_mse_solo_o1_bytes_close4_s14_qwen_bytes.safetensors",
7
+ "closes": [
8
+ "qarm_mse_solo_o1_bytes_close1_s14_qwen_bytes.safetensors",
9
+ "qarm_mse_solo_o1_bytes_close2_s14_qwen_bytes.safetensors",
10
+ "qarm_mse_solo_o1_bytes_close3_s14_qwen_bytes.safetensors",
11
+ "qarm_mse_solo_o1_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": "2026-10-09 11:16 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_o1_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_o1_bytes: per-site MSE; alone on the bare trunk; blocks 0-31; 4000 of 4000 steps; seed offset 1
8
+
9
+ ## The verdicts (the final close of each run)
10
+
11
+ ### qarm_mse_solo_o1_bytes
12
+
13
+ - B1 the gap: MET over every served block >= 12 (the lowest 0.964 at block 12 on draw 1, 0.963 at block 12 on draw 2; bar 0.85)
14
+ - B2 silence, whitened: MET (the lowest A alignment 0.9974 at block 7 on draw 1, 0.9955 on draw 2; bar 0.995)
15
+ - B2 silence, in bits: MET (web +0.00016 bpb; bar +0.012)
16
+
17
+ ## The closes
18
+
19
+ ### qarm_mse_solo_o1_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 / 0.916 / 0.940 | 0.9915 / 0.9907 | -0.00000 | 0.135 / 0.166 |
25
+ | 2 | 2000 | 0.944 / 0.946 | 0.944 / 0.963 / 0.974 | 0.9944 / 0.9931 | +0.00030 | 0.161 / 0.294 |
26
+ | 3 | 3000 | 0.956 / 0.958 | 0.956 / 0.974 / 0.981 | 0.9968 / 0.9925 | +0.00020 | 0.180 / 0.412 |
27
+ | 4 | 4000 | 0.964 / 0.963 | 0.964 / 0.977 / 0.985 | 0.9974 / 0.9955 | +0.00016 | 0.198 / 0.482 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### qarm_mse_solo_o1_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 | 0.878 | 0.9997 | 0.322 | 0.829 |
36
+ | 4 | 0.826 | 0.941 | 0.941 | 0.9979 | 0.081 | 0.954 |
37
+ | 8 | 0.766 | 0.947 | 0.947 | 0.9984 | 0.065 | 0.964 |
38
+ | 12 | 0.688 | 0.964 | 0.964 | 0.9987 | 0.052 | 0.971 |
39
+ | 16 | 0.628 | 0.973 | 0.972 | 0.9990 | 0.042 | 0.978 |
40
+ | 18 | 0.599 | 0.976 | 0.976 | 0.9988 | 0.038 | 0.980 |
41
+ | 20 | 0.580 | 0.977 | 0.978 | 0.9977 | 0.035 | 0.983 |
42
+ | 22 | 0.589 | 0.981 | 0.981 | 0.9990 | 0.031 | 0.985 |
43
+ | 24 | 0.605 | 0.984 | 0.984 | 0.9991 | 0.028 | 0.987 |
44
+ | 26 | 0.610 | 0.988 | 0.988 | 0.9991 | 0.026 | 0.988 |
45
+ | 28 | 0.598 | 0.985 | 0.985 | 0.9988 | 0.027 | 0.987 |
46
+ | 29 | 0.586 | 0.987 | 0.986 | 0.9988 | 0.028 | 0.986 |
47
+ | 30 | 0.550 | 0.983 | 0.983 | 0.9986 | 0.027 | 0.986 |
48
+ | 31 | 0.499 | 0.982 | 0.982 | 0.9977 | 0.028 | 0.984 |
49
+
50
+ ### qarm_mse_solo_o1_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 | 0.873 | 0.873 | 0.9998 | 0.322 | 0.829 |
55
+ | 4 | 0.833 | 0.940 | 0.940 | 0.9961 | 0.080 | 0.954 |
56
+ | 8 | 0.773 | 0.952 | 0.951 | 0.9990 | 0.064 | 0.964 |
57
+ | 12 | 0.698 | 0.963 | 0.963 | 0.9989 | 0.052 | 0.972 |
58
+ | 16 | 0.641 | 0.974 | 0.975 | 0.9978 | 0.041 | 0.978 |
59
+ | 18 | 0.615 | 0.973 | 0.975 | 0.9984 | 0.038 | 0.980 |
60
+ | 20 | 0.595 | 0.980 | 0.977 | 0.9973 | 0.035 | 0.982 |
61
+ | 22 | 0.604 | 0.983 | 0.983 | 0.9988 | 0.031 | 0.985 |
62
+ | 24 | 0.620 | 0.983 | 0.983 | 0.9989 | 0.028 | 0.987 |
63
+ | 26 | 0.621 | 0.984 | 0.984 | 0.9990 | 0.027 | 0.987 |
64
+ | 28 | 0.607 | 0.986 | 0.987 | 0.9986 | 0.028 | 0.986 |
65
+ | 29 | 0.596 | 0.983 | 0.984 | 0.9988 | 0.028 | 0.986 |
66
+ | 30 | 0.563 | 0.984 | 0.983 | 0.9985 | 0.028 | 0.985 |
67
+ | 31 | 0.513 | 0.980 | 0.980 | 0.9978 | 0.028 | 0.983 |
68
+
69
+ The member's share in bits per byte (all on minus the member masked): web +0.00016
70
+
71
+ ## The training curves
72
+
73
+ ### qarm_mse_solo_o1_bytes
74
+
75
+ | step | task loss | mse | cos | InfoNCE | quiet on A | quiet off-domain |
76
+ |---|---|---|---|---|---|---|
77
+ | 1 | 1.0450 | 1.045 | 0.493 | None | 0.00000 | 0.00000 |
78
+ | 50 | 1.0314 | 1.031 | 0.503 | None | 0.00031 | 0.00023 |
79
+ | 250 | 0.8788 | 0.879 | 0.570 | None | 0.00092 | 0.00049 |
80
+ | 500 | 0.6571 | 0.657 | 0.670 | None | 0.00073 | 0.00043 |
81
+ | 1000 | 0.3212 | 0.321 | 0.834 | None | 0.00071 | 0.00040 |
82
+ | 1500 | 0.2005 | 0.201 | 0.895 | None | 0.00074 | 0.00048 |
83
+ | 2000 | 0.1303 | 0.130 | 0.931 | None | 0.00077 | 0.00052 |
84
+ | 2500 | 0.0885 | 0.089 | 0.953 | None | 0.00074 | 0.00061 |
85
+ | 3000 | 0.0710 | 0.071 | 0.962 | None | 0.00067 | 0.00051 |
86
+ | 3500 | 0.0605 | 0.061 | 0.967 | None | 0.00062 | 0.00040 |
87
+ | 4000 | 0.0570 | 0.057 | 0.969 | None | 0.00058 | 0.00039 |
88
+ | 1 | 1.0450 | 1.045 | 0.493 | None | 0.00000 | 0.00000 |
89
+ | 50 | 1.0314 | 1.031 | 0.503 | None | 0.00031 | 0.00023 |
90
+ | 250 | 0.8788 | 0.879 | 0.570 | None | 0.00092 | 0.00049 |
91
+ | 500 | 0.6571 | 0.657 | 0.670 | None | 0.00073 | 0.00043 |
92
+ | 1000 | 0.3212 | 0.321 | 0.834 | None | 0.00071 | 0.00040 |
93
+ | 1500 | 0.2005 | 0.201 | 0.895 | None | 0.00074 | 0.00048 |
94
+ | 2000 | 0.1303 | 0.130 | 0.931 | None | 0.00077 | 0.00052 |
95
+ | 2500 | 0.0885 | 0.089 | 0.953 | None | 0.00074 | 0.00061 |
96
+ | 3000 | 0.0710 | 0.071 | 0.962 | None | 0.00067 | 0.00051 |
97
+ | 3500 | 0.0605 | 0.061 | 0.967 | None | 0.00062 | 0.00040 |
98
+ | 4000 | 0.0570 | 0.057 | 0.969 | None | 0.00058 | 0.00039 |
99
+ | 1 | 1.0450 | 1.045 | 0.493 | None | 0.00000 | 0.00000 |
100
+ | 50 | 1.0314 | 1.031 | 0.503 | None | 0.00031 | 0.00023 |
101
+ | 250 | 0.8788 | 0.879 | 0.570 | None | 0.00092 | 0.00049 |
102
+ | 500 | 0.6571 | 0.657 | 0.670 | None | 0.00073 | 0.00043 |
103
+ | 1000 | 0.3212 | 0.321 | 0.834 | None | 0.00071 | 0.00040 |
104
+ | 1500 | 0.2005 | 0.201 | 0.895 | None | 0.00074 | 0.00048 |
105
+ | 2000 | 0.1303 | 0.130 | 0.931 | None | 0.00077 | 0.00052 |
106
+ | 2500 | 0.0885 | 0.089 | 0.953 | None | 0.00074 | 0.00061 |
107
+ | 3000 | 0.0710 | 0.071 | 0.962 | None | 0.00067 | 0.00051 |
108
+ | 3500 | 0.0605 | 0.061 | 0.967 | None | 0.00062 | 0.00040 |
109
+ | 4000 | 0.0570 | 0.057 | 0.969 | None | 0.00058 | 0.00039 |
110
+ | 1 | 1.0450 | 1.045 | 0.493 | None | 0.00000 | 0.00000 |
111
+ | 50 | 1.0314 | 1.031 | 0.503 | None | 0.00031 | 0.00023 |
112
+ | 250 | 0.8788 | 0.879 | 0.570 | None | 0.00092 | 0.00049 |
113
+ | 500 | 0.6571 | 0.657 | 0.670 | None | 0.00073 | 0.00043 |
114
+ | 1000 | 0.3212 | 0.321 | 0.834 | None | 0.00071 | 0.00040 |
115
+ | 1500 | 0.2005 | 0.201 | 0.895 | None | 0.00074 | 0.00048 |
116
+ | 2000 | 0.1303 | 0.130 | 0.931 | None | 0.00077 | 0.00052 |
117
+ | 2500 | 0.0885 | 0.089 | 0.953 | None | 0.00074 | 0.00061 |
118
+ | 3000 | 0.0710 | 0.071 | 0.962 | None | 0.00067 | 0.00051 |
119
+ | 3500 | 0.0605 | 0.061 | 0.967 | None | 0.00062 | 0.00040 |
120
+ | 4000 | 0.0570 | 0.057 | 0.969 | None | 0.00058 | 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_solo_o1_bytes/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/qwen3/mse_solo_o1_bytes/s14_qwen_bytes.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:90925cb2d45b908a30af874d3155fc9e999c46e3fe946940867c1a2c1e10fb8d
3
+ size 54819344
surface/qwen3/mse_solo_o1_bytes/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:77420d656ac0c505f71f171b57d8f4aaa4948f600fcf2494822c9efc7a2d6b8b
3
+ size 264333
surface/qwen3/mse_solo_o1_bytes/train.log ADDED
@@ -0,0 +1,159 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 0.3s] THE JOB: qarm_mse_solo_o1_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
+ [ 16.8s] 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
4
+ [ 17.5s] 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
+ [ 17.5s] 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
+ [ 30.6s] 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
+ [ 30.8s] streams: web text; partners none (two web rows a quiet chunk); held-out sets ['web'] x 8 batches
8
+ [ 30.8s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 116), 25,597 sites
9
+ [ 30.9s] 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
+ [ 51.9s] 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
+ [ 68.3s] standardization statistics from 55,421 teacher sites on 8 paired batches (a mean and a spread per block and dimension); sigma range at block 20 0.773..1.089; saved /workspace/qarm/out/anchors/qarm_mse_solo_o1_bytes_stats.pt
13
+ [ 517.1s] 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
+ [ 517.1s] block 8: 0.766 -> 0.766 / 0.766; A 1.0000; mse 0.827 cos 0.610
15
+ [ 517.1s] block 12: 0.688 -> 0.688 / 0.688; A 1.0000; mse 0.973 cos 0.530
16
+ [ 517.1s] block 16: 0.628 -> 0.628 / 0.628; A 1.0000; mse 1.088 cos 0.466
17
+ [ 517.1s] block 18: 0.599 -> 0.599 / 0.599; A 1.0000; mse 1.136 cos 0.439
18
+ [ 517.1s] block 20: 0.580 -> 0.580 / 0.580; A 1.0000; mse 1.185 cos 0.416
19
+ [ 517.1s] block 24: 0.605 -> 0.605 / 0.605; A 1.0000; mse 1.210 cos 0.420
20
+ [ 517.1s] block 28: 0.598 -> 0.598 / 0.598; A 1.0000; mse 1.258 cos 0.373
21
+ [ 517.1s] block 31: 0.499 -> 0.499 / 0.499; A 1.0000; mse 1.011 cos 0.317
22
+ [ 517.1s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.698 / 0.641 / 0.595 / 0.620 / 0.607
23
+ [ 517.1s] 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
+ [ 517.1s] gate mean 0.047 / widest 0.047; compiled chain: 2 cache entries (limit 256), 0 limit warnings; docs drawn 274 (skipped {'test': 3, '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
+ [ 532.9s] step 1/4000: task 1.0450 (mse 1.045 cos 0.493); quiet A 0.00000 off 0.00000; 6,945 sites; 15.80 s/step; the job: 8.9 min spent, about 1052.8 min of steps left; card peak 16.5 GB
27
+ [ 657.6s] step 50/4000: task 1.0314 (mse 1.031 cos 0.503); quiet A 0.00031 off 0.00023; 6,930 sites; 2.81 s/step; the job: 11.0 min spent, about 185.0 min of steps left; card peak 16.6 GB
28
+ [ 784.8s] step 100/4000: task 0.9925 (mse 0.992 cos 0.520); quiet A 0.00050 off 0.00035; 6,969 sites; 2.68 s/step; the job: 13.1 min spent, about 174.0 min of steps left; card peak 16.6 GB
29
+ [ 912.1s] step 150/4000: task 0.9595 (mse 0.959 cos 0.534); quiet A 0.00055 off 0.00037; 6,939 sites; 2.63 s/step; the job: 15.2 min spent, about 169.0 min of steps left; card peak 16.6 GB
30
+ [ 1039.4s] step 200/4000: task 0.9136 (mse 0.914 cos 0.555); quiet A 0.00078 off 0.00052; 6,865 sites; 2.61 s/step; the job: 17.3 min spent, about 165.4 min of steps left; card peak 16.6 GB
31
+ [ 1166.7s] step 250/4000: task 0.8788 (mse 0.879 cos 0.570); quiet A 0.00092 off 0.00049; 6,905 sites; 2.60 s/step; the job: 19.4 min spent, about 162.4 min of steps left; card peak 16.6 GB
32
+ [ 1294.3s] step 300/4000: task 0.8390 (mse 0.839 cos 0.587); quiet A 0.00108 off 0.00039; 6,877 sites; 2.59 s/step; the job: 21.6 min spent, about 159.8 min of steps left; card peak 16.6 GB
33
+ [ 1421.5s] step 350/4000: task 0.7897 (mse 0.790 cos 0.610); quiet A 0.00095 off 0.00039; 6,954 sites; 2.58 s/step; the job: 23.7 min spent, about 157.2 min of steps left; card peak 16.6 GB
34
+ [ 1548.8s] step 400/4000: task 0.7533 (mse 0.753 cos 0.626); quiet A 0.00082 off 0.00040; 6,966 sites; 2.58 s/step; the job: 25.8 min spent, about 154.8 min of steps left; card peak 16.6 GB
35
+ [ 1676.2s] step 450/4000: task 0.7092 (mse 0.709 cos 0.646); quiet A 0.00074 off 0.00038; 6,927 sites; 2.58 s/step; the job: 27.9 min spent, about 152.4 min of steps left; card peak 16.6 GB
36
+ [ 1803.6s] step 500/4000: task 0.6571 (mse 0.657 cos 0.670); quiet A 0.00073 off 0.00043; 6,910 sites; 2.57 s/step; the job: 30.1 min spent, about 150.1 min of steps left; card peak 16.6 GB
37
+ [ 1931.3s] step 550/4000: task 0.5948 (mse 0.595 cos 0.699); quiet A 0.00068 off 0.00037; 6,911 sites; 2.57 s/step; the job: 32.2 min spent, about 147.9 min of steps left; card peak 16.6 GB
38
+ [ 2058.7s] step 600/4000: task 0.5206 (mse 0.521 cos 0.736); quiet A 0.00067 off 0.00038; 6,933 sites; 2.57 s/step; the job: 34.3 min spent, about 145.6 min of steps left; card peak 16.6 GB
39
+ [ 2185.9s] step 650/4000: task 0.4691 (mse 0.469 cos 0.762); quiet A 0.00064 off 0.00036; 6,951 sites; 2.57 s/step; the job: 36.4 min spent, about 143.3 min of steps left; card peak 16.6 GB
40
+ [ 2313.1s] step 700/4000: task 0.4368 (mse 0.437 cos 0.778); quiet A 0.00066 off 0.00042; 6,937 sites; 2.57 s/step; the job: 38.6 min spent, about 141.1 min of steps left; card peak 16.6 GB
41
+ [ 2440.3s] step 750/4000: task 0.4060 (mse 0.406 cos 0.793); quiet A 0.00065 off 0.00036; 6,935 sites; 2.56 s/step; the job: 40.7 min spent, about 138.9 min of steps left; card peak 16.6 GB
42
+ [ 2567.8s] step 800/4000: task 0.3859 (mse 0.386 cos 0.802); quiet A 0.00065 off 0.00042; 6,996 sites; 2.56 s/step; the job: 42.8 min spent, about 136.7 min of steps left; card peak 16.6 GB
43
+ [ 2695.0s] step 850/4000: task 0.3665 (mse 0.367 cos 0.812); quiet A 0.00069 off 0.00036; 6,932 sites; 2.56 s/step; the job: 44.9 min spent, about 134.5 min of steps left; card peak 16.6 GB
44
+ [ 2822.3s] step 900/4000: task 0.3493 (mse 0.349 cos 0.820); quiet A 0.00069 off 0.00038; 6,887 sites; 2.56 s/step; the job: 47.0 min spent, about 132.3 min of steps left; card peak 16.6 GB
45
+ [ 2949.7s] step 950/4000: task 0.3346 (mse 0.335 cos 0.828); quiet A 0.00069 off 0.00039; 6,910 sites; 2.56 s/step; the job: 49.2 min spent, about 130.2 min of steps left; card peak 16.6 GB
46
+ [ 3077.4s] step 1000/4000: task 0.3212 (mse 0.321 cos 0.834); quiet A 0.00071 off 0.00040; 6,978 sites; 2.56 s/step; the job: 51.3 min spent, about 128.0 min of steps left; card peak 16.6 GB
47
+ [ 3077.7s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o1_bytes_close1_s14_qwen_bytes.safetensors
48
+ [ 3412.8s] 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
+ [ 3412.8s] block 8: 0.766 -> 0.883 / 0.886; A 0.9956; mse 0.254 cos 0.862
50
+ [ 3412.8s] block 12: 0.688 -> 0.897 / 0.899; A 0.9930; mse 0.203 cos 0.890
51
+ [ 3412.8s] block 16: 0.628 -> 0.911 / 0.913; A 0.9952; mse 0.167 cos 0.913
52
+ [ 3412.8s] block 18: 0.599 -> 0.911 / 0.914; A 0.9918; mse 0.159 cos 0.919
53
+ [ 3412.8s] block 20: 0.580 -> 0.916 / 0.919; A 0.9954; mse 0.152 cos 0.924
54
+ [ 3412.8s] block 24: 0.605 -> 0.938 / 0.941; A 0.9967; mse 0.125 cos 0.939
55
+ [ 3412.8s] block 28: 0.598 -> 0.940 / 0.943; A 0.9956; mse 0.122 cos 0.940
56
+ [ 3412.8s] block 31: 0.499 -> 0.925 / 0.928; A 0.9921; mse 0.111 cos 0.933
57
+ [ 3412.8s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.903 / 0.911 / 0.920 / 0.935 / 0.941
58
+ [ 3412.8s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.896 against the bar .85; B2: the lowest A alignment 0.9915 against .995; fineweb share -0.00000 +- 0.00010 against +.012
59
+ [ 3412.8s] gate mean 0.135 / widest 0.166; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 32,659 (skipped {'test': 51, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 16.6 GB; host RSS 8.2 GB
60
+ [ 3540.1s] step 1050/4000: task 0.3105 (mse 0.311 cos 0.839); quiet A 0.00073 off 0.00039; 6,898 sites; 2.55 s/step; the job: 59.0 min spent, about 125.8 min of steps left; card peak 16.6 GB
61
+ [ 3667.3s] step 1100/4000: task 0.2978 (mse 0.298 cos 0.846); quiet A 0.00074 off 0.00042; 6,947 sites; 2.54 s/step; the job: 61.1 min spent, about 123.7 min of steps left; card peak 16.6 GB
62
+ [ 3794.5s] step 1150/4000: task 0.2839 (mse 0.284 cos 0.852); quiet A 0.00075 off 0.00041; 6,946 sites; 2.54 s/step; the job: 63.2 min spent, about 121.5 min of steps left; card peak 16.6 GB
63
+ [ 3921.7s] step 1200/4000: task 0.2716 (mse 0.272 cos 0.859); quiet A 0.00072 off 0.00041; 6,958 sites; 2.54 s/step; the job: 65.4 min spent, about 119.4 min of steps left; card peak 16.6 GB
64
+ [ 4049.1s] step 1250/4000: task 0.2612 (mse 0.261 cos 0.864); quiet A 0.00072 off 0.00042; 6,935 sites; 2.55 s/step; the job: 67.5 min spent, about 117.2 min of steps left; card peak 16.6 GB
65
+ [ 4176.5s] step 1300/4000: task 0.2473 (mse 0.247 cos 0.871); quiet A 0.00074 off 0.00044; 6,902 sites; 2.55 s/step; the job: 69.6 min spent, about 115.1 min of steps left; card peak 16.6 GB
66
+ [ 4304.1s] step 1350/4000: task 0.2327 (mse 0.233 cos 0.879); quiet A 0.00073 off 0.00045; 6,939 sites; 2.55 s/step; the job: 71.7 min spent, about 112.9 min of steps left; card peak 16.6 GB
67
+ [ 4431.3s] step 1400/4000: task 0.2234 (mse 0.223 cos 0.883); quiet A 0.00073 off 0.00044; 6,946 sites; 2.55 s/step; the job: 73.9 min spent, about 110.8 min of steps left; card peak 16.6 GB
68
+ [ 4559.1s] step 1450/4000: task 0.2102 (mse 0.210 cos 0.890); quiet A 0.00075 off 0.00045; 6,934 sites; 2.55 s/step; the job: 76.0 min spent, about 108.6 min of steps left; card peak 16.6 GB
69
+ [ 4686.5s] step 1500/4000: task 0.2005 (mse 0.201 cos 0.895); quiet A 0.00074 off 0.00048; 6,976 sites; 2.55 s/step; the job: 78.1 min spent, about 106.5 min of steps left; card peak 16.6 GB
70
+ [ 4813.9s] step 1550/4000: task 0.1915 (mse 0.192 cos 0.900); quiet A 0.00075 off 0.00049; 6,825 sites; 2.55 s/step; the job: 80.2 min spent, about 104.4 min of steps left; card peak 16.6 GB
71
+ [ 4941.7s] step 1600/4000: task 0.1832 (mse 0.183 cos 0.904); quiet A 0.00075 off 0.00052; 6,882 sites; 2.55 s/step; the job: 82.4 min spent, about 102.2 min of steps left; card peak 16.6 GB
72
+ [ 5068.7s] step 1650/4000: task 0.1772 (mse 0.177 cos 0.907); quiet A 0.00076 off 0.00051; 6,935 sites; 2.55 s/step; the job: 84.5 min spent, about 100.1 min of steps left; card peak 16.6 GB
73
+ [ 5196.5s] step 1700/4000: task 0.1678 (mse 0.168 cos 0.912); quiet A 0.00078 off 0.00062; 6,950 sites; 2.55 s/step; the job: 86.6 min spent, about 98.0 min of steps left; card peak 16.6 GB
74
+ [ 5323.8s] step 1750/4000: task 0.1583 (mse 0.158 cos 0.917); quiet A 0.00079 off 0.00048; 6,969 sites; 2.55 s/step; the job: 88.7 min spent, about 95.8 min of steps left; card peak 16.6 GB
75
+ [ 5451.0s] step 1800/4000: task 0.1535 (mse 0.154 cos 0.919); quiet A 0.00077 off 0.00048; 6,969 sites; 2.55 s/step; the job: 90.9 min spent, about 93.7 min of steps left; card peak 16.6 GB
76
+ [ 5578.5s] step 1850/4000: task 0.1472 (mse 0.147 cos 0.923); quiet A 0.00074 off 0.00048; 6,995 sites; 2.55 s/step; the job: 93.0 min spent, about 91.5 min of steps left; card peak 16.6 GB
77
+ [ 5706.0s] step 1900/4000: task 0.1425 (mse 0.143 cos 0.925); quiet A 0.00079 off 0.00048; 6,908 sites; 2.55 s/step; the job: 95.1 min spent, about 89.4 min of steps left; card peak 16.6 GB
78
+ [ 5833.3s] step 1950/4000: task 0.1375 (mse 0.138 cos 0.928); quiet A 0.00077 off 0.00048; 6,895 sites; 2.55 s/step; the job: 97.2 min spent, about 87.3 min of steps left; card peak 16.6 GB
79
+ [ 5960.4s] step 2000/4000: task 0.1303 (mse 0.130 cos 0.931); quiet A 0.00077 off 0.00052; 6,907 sites; 2.55 s/step; the job: 99.3 min spent, about 85.1 min of steps left; card peak 16.6 GB
80
+ [ 5960.7s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o1_bytes_close2_s14_qwen_bytes.safetensors
81
+ [ 6295.9s] 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
+ [ 6295.9s] block 8: 0.766 -> 0.929 / 0.930; A 0.9965; mse 0.111 cos 0.938
83
+ [ 6295.9s] block 12: 0.688 -> 0.944 / 0.944; A 0.9974; mse 0.089 cos 0.951
84
+ [ 6295.9s] block 16: 0.628 -> 0.958 / 0.958; A 0.9977; mse 0.071 cos 0.963
85
+ [ 6295.9s] block 18: 0.599 -> 0.959 / 0.959; A 0.9969; mse 0.066 cos 0.966
86
+ [ 6295.9s] block 20: 0.580 -> 0.963 / 0.964; A 0.9973; mse 0.061 cos 0.969
87
+ [ 6295.9s] block 24: 0.605 -> 0.974 / 0.975; A 0.9977; mse 0.049 cos 0.977
88
+ [ 6295.9s] block 28: 0.598 -> 0.974 / 0.975; A 0.9978; mse 0.047 cos 0.977
89
+ [ 6295.9s] block 31: 0.499 -> 0.968 / 0.969; A 0.9957; mse 0.047 cos 0.973
90
+ [ 6295.9s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.947 / 0.958 / 0.962 / 0.973 / 0.974
91
+ [ 6295.9s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.944 against the bar .85; B2: the lowest A alignment 0.9944 against .995; fineweb share +0.00030 +- 0.00020 against +.012
92
+ [ 6295.9s] gate mean 0.161 / widest 0.294; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 64,676 (skipped {'test': 88, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 16.6 GB; host RSS 8.2 GB
93
+ [ 6423.2s] step 2050/4000: task 0.1237 (mse 0.124 cos 0.935); quiet A 0.00079 off 0.00053; 6,932 sites; 2.55 s/step; the job: 107.1 min spent, about 83.0 min of steps left; card peak 16.6 GB
94
+ [ 6550.5s] step 2100/4000: task 0.1210 (mse 0.121 cos 0.936); quiet A 0.00076 off 0.00046; 6,920 sites; 2.55 s/step; the job: 109.2 min spent, about 80.9 min of steps left; card peak 16.6 GB
95
+ [ 6678.1s] step 2150/4000: task 0.1165 (mse 0.116 cos 0.939); quiet A 0.00077 off 0.00052; 6,925 sites; 2.55 s/step; the job: 111.3 min spent, about 78.7 min of steps left; card peak 16.6 GB
96
+ [ 6805.5s] step 2200/4000: task 0.1126 (mse 0.113 cos 0.941); quiet A 0.00074 off 0.00048; 6,972 sites; 2.55 s/step; the job: 113.4 min spent, about 76.6 min of steps left; card peak 16.6 GB
97
+ [ 6932.9s] step 2250/4000: task 0.1092 (mse 0.109 cos 0.942); quiet A 0.00077 off 0.00057; 6,897 sites; 2.55 s/step; the job: 115.5 min spent, about 74.5 min of steps left; card peak 16.6 GB
98
+ [ 7060.2s] step 2300/4000: task 0.1042 (mse 0.104 cos 0.945); quiet A 0.00076 off 0.00054; 6,916 sites; 2.55 s/step; the job: 117.7 min spent, about 72.3 min of steps left; card peak 16.6 GB
99
+ [ 7187.4s] step 2350/4000: task 0.0984 (mse 0.098 cos 0.948); quiet A 0.00074 off 0.00050; 6,944 sites; 2.55 s/step; the job: 119.8 min spent, about 70.2 min of steps left; card peak 16.6 GB
100
+ [ 7315.1s] step 2400/4000: task 0.0965 (mse 0.097 cos 0.949); quiet A 0.00076 off 0.00047; 6,934 sites; 2.55 s/step; the job: 121.9 min spent, about 68.1 min of steps left; card peak 16.6 GB
101
+ [ 7442.4s] step 2450/4000: task 0.0935 (mse 0.094 cos 0.950); quiet A 0.00073 off 0.00051; 6,948 sites; 2.55 s/step; the job: 124.0 min spent, about 65.9 min of steps left; card peak 16.6 GB
102
+ [ 7570.0s] step 2500/4000: task 0.0885 (mse 0.089 cos 0.953); quiet A 0.00074 off 0.00061; 6,933 sites; 2.55 s/step; the job: 126.2 min spent, about 63.8 min of steps left; card peak 16.6 GB
103
+ [ 7697.4s] step 2550/4000: task 0.0891 (mse 0.089 cos 0.953); quiet A 0.00079 off 0.00051; 6,827 sites; 2.55 s/step; the job: 128.3 min spent, about 61.7 min of steps left; card peak 16.6 GB
104
+ [ 7824.6s] step 2600/4000: task 0.0888 (mse 0.089 cos 0.953); quiet A 0.00073 off 0.00050; 6,918 sites; 2.55 s/step; the job: 130.4 min spent, about 59.6 min of steps left; card peak 16.6 GB
105
+ [ 7952.5s] step 2650/4000: task 0.0795 (mse 0.079 cos 0.957); quiet A 0.00070 off 0.00048; 6,943 sites; 2.55 s/step; the job: 132.5 min spent, about 57.4 min of steps left; card peak 16.6 GB
106
+ [ 8080.0s] step 2700/4000: task 0.0787 (mse 0.079 cos 0.958); quiet A 0.00070 off 0.00044; 6,889 sites; 2.55 s/step; the job: 134.7 min spent, about 55.3 min of steps left; card peak 16.6 GB
107
+ [ 8208.5s] step 2750/4000: task 0.0755 (mse 0.076 cos 0.960); quiet A 0.00068 off 0.00046; 6,923 sites; 2.55 s/step; the job: 136.8 min spent, about 53.2 min of steps left; card peak 16.6 GB
108
+ [ 8335.9s] step 2800/4000: task 0.0774 (mse 0.077 cos 0.959); quiet A 0.00068 off 0.00047; 6,921 sites; 2.55 s/step; the job: 138.9 min spent, about 51.1 min of steps left; card peak 16.6 GB
109
+ [ 8463.3s] step 2850/4000: task 0.0747 (mse 0.075 cos 0.960); quiet A 0.00068 off 0.00044; 6,881 sites; 2.55 s/step; the job: 141.1 min spent, about 48.9 min of steps left; card peak 16.6 GB
110
+ [ 8590.5s] step 2900/4000: task 0.0709 (mse 0.071 cos 0.962); quiet A 0.00066 off 0.00042; 6,941 sites; 2.55 s/step; the job: 143.2 min spent, about 46.8 min of steps left; card peak 16.6 GB
111
+ [ 8718.0s] step 2950/4000: task 0.0712 (mse 0.071 cos 0.962); quiet A 0.00067 off 0.00044; 6,946 sites; 2.55 s/step; the job: 145.3 min spent, about 44.7 min of steps left; card peak 16.6 GB
112
+ [ 8845.2s] step 3000/4000: task 0.0710 (mse 0.071 cos 0.962); quiet A 0.00067 off 0.00051; 6,852 sites; 2.55 s/step; the job: 147.4 min spent, about 42.5 min of steps left; card peak 16.6 GB
113
+ [ 8845.5s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o1_bytes_close3_s14_qwen_bytes.safetensors
114
+ [ 9179.9s] 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
+ [ 9179.9s] block 8: 0.766 -> 0.940 / 0.941; A 0.9988; mse 0.079 cos 0.956
116
+ [ 9179.9s] block 12: 0.688 -> 0.956 / 0.956; A 0.9982; mse 0.064 cos 0.965
117
+ [ 9179.9s] block 16: 0.628 -> 0.968 / 0.968; A 0.9981; mse 0.051 cos 0.973
118
+ [ 9179.9s] block 18: 0.599 -> 0.972 / 0.970; A 0.9979; mse 0.047 cos 0.976
119
+ [ 9179.9s] block 20: 0.580 -> 0.974 / 0.974; A 0.9981; mse 0.043 cos 0.979
120
+ [ 9179.9s] block 24: 0.605 -> 0.981 / 0.981; A 0.9988; mse 0.034 cos 0.984
121
+ [ 9179.9s] block 28: 0.598 -> 0.981 / 0.982; A 0.9987; mse 0.033 cos 0.984
122
+ [ 9179.9s] block 31: 0.499 -> 0.978 / 0.978; A 0.9972; mse 0.033 cos 0.980
123
+ [ 9179.9s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.958 / 0.968 / 0.974 / 0.980 / 0.983
124
+ [ 9179.9s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.956 against the bar .85; B2: the lowest A alignment 0.9968 against .995; fineweb share +0.00020 +- 0.00017 against +.012
125
+ [ 9179.9s] gate mean 0.180 / widest 0.412; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 96,834 (skipped {'test': 137, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 16.6 GB; host RSS 8.3 GB
126
+ [ 9307.2s] step 3050/4000: task 0.0748 (mse 0.075 cos 0.960); quiet A 0.00067 off 0.00045; 6,974 sites; 2.55 s/step; the job: 155.1 min spent, about 40.4 min of steps left; card peak 16.6 GB
127
+ [ 9434.6s] step 3100/4000: task 0.0670 (mse 0.067 cos 0.964); quiet A 0.00069 off 0.00103; 6,931 sites; 2.55 s/step; the job: 157.2 min spent, about 38.3 min of steps left; card peak 16.6 GB
128
+ [ 9562.0s] step 3150/4000: task 0.0696 (mse 0.070 cos 0.963); quiet A 0.00084 off 0.00051; 6,930 sites; 2.55 s/step; the job: 159.4 min spent, about 36.2 min of steps left; card peak 16.6 GB
129
+ [ 9689.7s] step 3200/4000: task 0.0687 (mse 0.069 cos 0.963); quiet A 0.00068 off 0.00043; 6,904 sites; 2.55 s/step; the job: 161.5 min spent, about 34.0 min of steps left; card peak 16.6 GB
130
+ [ 9817.0s] step 3250/4000: task 0.0647 (mse 0.065 cos 0.965); quiet A 0.00067 off 0.00043; 6,922 sites; 2.55 s/step; the job: 163.6 min spent, about 31.9 min of steps left; card peak 16.6 GB
131
+ [ 9944.2s] step 3300/4000: task 0.0672 (mse 0.067 cos 0.964); quiet A 0.00065 off 0.00046; 6,936 sites; 2.55 s/step; the job: 165.7 min spent, about 29.8 min of steps left; card peak 16.6 GB
132
+ [10071.8s] step 3350/4000: task 0.0645 (mse 0.064 cos 0.965); quiet A 0.00067 off 0.00045; 6,915 sites; 2.55 s/step; the job: 167.9 min spent, about 27.6 min of steps left; card peak 16.6 GB
133
+ [10199.1s] step 3400/4000: task 0.0644 (mse 0.064 cos 0.966); quiet A 0.00062 off 0.00043; 6,952 sites; 2.55 s/step; the job: 170.0 min spent, about 25.5 min of steps left; card peak 16.6 GB
134
+ [10326.4s] step 3450/4000: task 0.0614 (mse 0.061 cos 0.967); quiet A 0.00061 off 0.00040; 6,962 sites; 2.55 s/step; the job: 172.1 min spent, about 23.4 min of steps left; card peak 16.6 GB
135
+ [10453.8s] step 3500/4000: task 0.0605 (mse 0.061 cos 0.967); quiet A 0.00062 off 0.00040; 6,910 sites; 2.55 s/step; the job: 174.2 min spent, about 21.3 min of steps left; card peak 16.6 GB
136
+ [10581.3s] step 3550/4000: task 0.0610 (mse 0.061 cos 0.967); quiet A 0.00059 off 0.00039; 6,906 sites; 2.55 s/step; the job: 176.4 min spent, about 19.1 min of steps left; card peak 16.6 GB
137
+ [10708.7s] step 3600/4000: task 0.0626 (mse 0.063 cos 0.967); quiet A 0.00061 off 0.00044; 6,934 sites; 2.55 s/step; the job: 178.5 min spent, about 17.0 min of steps left; card peak 16.6 GB
138
+ [10836.1s] step 3650/4000: task 0.0588 (mse 0.059 cos 0.968); quiet A 0.00061 off 0.00040; 6,929 sites; 2.55 s/step; the job: 180.6 min spent, about 14.9 min of steps left; card peak 16.6 GB
139
+ [10963.6s] step 3700/4000: task 0.0588 (mse 0.059 cos 0.968); quiet A 0.00063 off 0.00056; 6,895 sites; 2.55 s/step; the job: 182.7 min spent, about 12.8 min of steps left; card peak 16.6 GB
140
+ [11091.0s] step 3750/4000: task 0.0579 (mse 0.058 cos 0.969); quiet A 0.00060 off 0.00041; 6,918 sites; 2.55 s/step; the job: 184.9 min spent, about 10.6 min of steps left; card peak 16.6 GB
141
+ [11218.6s] step 3800/4000: task 0.0593 (mse 0.059 cos 0.968); quiet A 0.00059 off 0.00041; 6,937 sites; 2.55 s/step; the job: 187.0 min spent, about 8.5 min of steps left; card peak 16.6 GB
142
+ [11345.9s] step 3850/4000: task 0.0590 (mse 0.059 cos 0.968); quiet A 0.00059 off 0.00039; 6,946 sites; 2.55 s/step; the job: 189.1 min spent, about 6.4 min of steps left; card peak 16.6 GB
143
+ [11473.0s] step 3900/4000: task 0.0565 (mse 0.056 cos 0.970); quiet A 0.00059 off 0.00037; 6,931 sites; 2.55 s/step; the job: 191.2 min spent, about 4.3 min of steps left; card peak 16.6 GB
144
+ [11600.1s] step 3950/4000: task 0.0561 (mse 0.056 cos 0.970); quiet A 0.00058 off 0.00043; 6,912 sites; 2.55 s/step; the job: 193.3 min spent, about 2.1 min of steps left; card peak 16.6 GB
145
+ [11727.6s] step 4000/4000: task 0.0570 (mse 0.057 cos 0.969); quiet A 0.00058 off 0.00039; 6,940 sites; 2.55 s/step; the job: 195.5 min spent, about 0.0 min of steps left; card peak 16.6 GB
146
+ [11727.9s] anchor saved: /workspace/qarm/out/anchors/qarm_mse_solo_o1_bytes_close4_s14_qwen_bytes.safetensors
147
+ [12061.7s] 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
+ [12061.7s] block 8: 0.766 -> 0.947 / 0.947; A 0.9984; mse 0.065 cos 0.964
149
+ [12061.7s] block 12: 0.688 -> 0.964 / 0.964; A 0.9987; mse 0.052 cos 0.971
150
+ [12061.7s] block 16: 0.628 -> 0.973 / 0.972; A 0.9990; mse 0.042 cos 0.978
151
+ [12061.7s] block 18: 0.599 -> 0.976 / 0.976; A 0.9988; mse 0.038 cos 0.980
152
+ [12061.7s] block 20: 0.580 -> 0.977 / 0.978; A 0.9977; mse 0.035 cos 0.983
153
+ [12061.7s] block 24: 0.605 -> 0.984 / 0.984; A 0.9991; mse 0.028 cos 0.987
154
+ [12061.7s] block 28: 0.598 -> 0.985 / 0.985; A 0.9988; mse 0.027 cos 0.987
155
+ [12061.7s] block 31: 0.499 -> 0.982 / 0.982; A 0.9977; mse 0.028 cos 0.984
156
+ [12061.7s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.963 / 0.974 / 0.980 / 0.983 / 0.986
157
+ [12061.7s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.964 against the bar .85; B2: the lowest A alignment 0.9974 against .995; fineweb share +0.00016 +- 0.00015 against +.012
158
+ [12061.7s] gate mean 0.198 / widest 0.482; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 128,945 (skipped {'test': 195, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 16.6 GB; host RSS 8.3 GB
159
+ [12061.7s] DONE qarm_mse_solo_o1_bytes: 4000 steps in 170.1 min of steps (2.55 s/step); 4 closes; 12062 s in all; card peak 16.6 GB; host RSS 8.3 GB; compiled chain: 5 cache entries (limit 256), 0 limit warnings