Hugging Face's logo

AbstractPhil
/
beatrix-tokenizers

geolip-alephllm
byte-level
adapter
tokenizer
beatrix
Model card Files
xet
 Community
AbstractPhil commited on
Commit
08651d3
·
verified ·
1 Parent(s): 91cd831

surface arm t5/mse_solo_o1: anchors (every close), stats, results, log, readout

Browse files
Files changed (10) hide show
  1. surface/t5/mse_solo_o1/closes/sarm_t5_mse_solo_o1_close1_s11_t5.safetensors +3 -0
  2. surface/t5/mse_solo_o1/closes/sarm_t5_mse_solo_o1_close2_s11_t5.safetensors +3 -0
  3. surface/t5/mse_solo_o1/closes/sarm_t5_mse_solo_o1_close3_s11_t5.safetensors +3 -0
  4. surface/t5/mse_solo_o1/closes/sarm_t5_mse_solo_o1_close4_s11_t5.safetensors +3 -0
  5. surface/t5/mse_solo_o1/manifest.json +20 -0
  6. surface/t5/mse_solo_o1/readout.md +127 -0
  7. surface/t5/mse_solo_o1/results.json +0 -0
  8. surface/t5/mse_solo_o1/s11_t5.safetensors +3 -0
  9. surface/t5/mse_solo_o1/stats.pt +3 -0
  10. surface/t5/mse_solo_o1/train.log +160 -0
surface/t5/mse_solo_o1/closes/sarm_t5_mse_solo_o1_close1_s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b213dbbfaddedecb5ec1ed9b33297baaf70ca93a57488ee512741709b8b25aa3
3
+ size 54819296
surface/t5/mse_solo_o1/closes/sarm_t5_mse_solo_o1_close2_s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b06b08272e9db8fa17e115f81398ef10915d1dcb5e053f98bf6cf8069551ece
3
+ size 54819296
surface/t5/mse_solo_o1/closes/sarm_t5_mse_solo_o1_close3_s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:348506d398f5434998073e0ba45c93542db189a7c943b72bc70c48154cafe3bb
3
+ size 54819296
surface/t5/mse_solo_o1/closes/sarm_t5_mse_solo_o1_close4_s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d84548fcef5603b78c74fc0ce41386cee5f9ba6d266974ed228064e3b7d79d23
3
+ size 54819296
surface/t5/mse_solo_o1/manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tag": "sarm_t5_mse_solo_o1",
3
+ "run": "mse_solo_o1",
4
+ "convention": "t5",
5
+ "member": "s11_t5",
6
+ "final_anchor": "sarm_t5_mse_solo_o1_close4_s11_t5.safetensors",
7
+ "closes": [
8
+ "sarm_t5_mse_solo_o1_close1_s11_t5.safetensors",
9
+ "sarm_t5_mse_solo_o1_close2_s11_t5.safetensors",
10
+ "sarm_t5_mse_solo_o1_close3_s11_t5.safetensors",
11
+ "sarm_t5_mse_solo_o1_close4_s11_t5.safetensors"
12
+ ],
13
+ "finished": true,
14
+ "form": "mse",
15
+ "base": "solo",
16
+ "steps": 4000,
17
+ "trunk_step": 245674,
18
+ "shipped": "2026-10-09 04:02 local",
19
+ "note": "every close's anchor ships (mids included); the final close's anchor is the run's s11_t5.safetensors"
20
+ }
surface/t5/mse_solo_o1/readout.md ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # The surface arm's readout: t5-v1_1-xxl's spelling (the sentence-piece convention)
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
+ - sarm_t5_mse_solo_o1: 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
+ ### sarm_t5_mse_solo_o1
12
+
13
+ - B1 the gap: MET over every served block >= 12 (the lowest 0.948 at block 12 on draw 1, 0.941 at block 31 on draw 2; bar 0.85)
14
+ - B2 silence, whitened: NOT MET (the lowest A alignment 0.9925 at block 1 on draw 1, 0.9917 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
+ ### sarm_t5_mse_solo_o1
20
+
21
+ | close | step | B1 lowest (draw 1 / 2) | B1 at 12 / 20 / 28 (draw 1) | B2 lowest (draw 1 / 2) | web bits | gate mean / widest |
22
+ |---|---|---|---|---|---|---|
23
+ | 0 | 0 | 0.360 / 0.389 | 0.647 / 0.519 / 0.435 | 1.0000 / 1.0000 | +0.00000 | 0.047 / 0.047 |
24
+ | 1 | 1000 | 0.827 / 0.821 | 0.848 / 0.867 / 0.859 | 0.9890 / 0.9856 | +0.00015 | 0.108 / 0.147 |
25
+ | 2 | 2000 | 0.919 / 0.909 | 0.919 / 0.942 / 0.937 | 0.9901 / 0.9893 | +0.00022 | 0.142 / 0.290 |
26
+ | 3 | 3000 | 0.937 / 0.933 | 0.937 / 0.958 / 0.953 | 0.9895 / 0.9901 | +0.00037 | 0.168 / 0.402 |
27
+ | 4 | 4000 | 0.948 / 0.941 | 0.948 / 0.965 / 0.960 | 0.9925 / 0.9917 | +0.00008 | 0.188 / 0.487 |
28
+
29
+ ## Every block at the final close
30
+
31
+ ### sarm_t5_mse_solo_o1, draw1 (512 captions, 5,707 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 | 0.868 | 0.867 | 0.867 | 0.9998 | 0.361 | 0.805 |
36
+ | 4 | 0.797 | 0.921 | 0.920 | 0.9963 | 0.091 | 0.951 |
37
+ | 8 | 0.735 | 0.934 | 0.935 | 0.9962 | 0.078 | 0.960 |
38
+ | 12 | 0.647 | 0.948 | 0.948 | 0.9968 | 0.067 | 0.967 |
39
+ | 16 | 0.572 | 0.961 | 0.961 | 0.9956 | 0.055 | 0.973 |
40
+ | 18 | 0.545 | 0.961 | 0.960 | 0.9974 | 0.054 | 0.974 |
41
+ | 20 | 0.519 | 0.965 | 0.963 | 0.9975 | 0.052 | 0.975 |
42
+ | 22 | 0.495 | 0.965 | 0.964 | 0.9979 | 0.052 | 0.977 |
43
+ | 24 | 0.470 | 0.965 | 0.963 | 0.9966 | 0.049 | 0.978 |
44
+ | 26 | 0.448 | 0.964 | 0.963 | 0.9973 | 0.047 | 0.977 |
45
+ | 28 | 0.435 | 0.960 | 0.957 | 0.9944 | 0.049 | 0.975 |
46
+ | 29 | 0.421 | 0.958 | 0.956 | 0.9968 | 0.051 | 0.974 |
47
+ | 30 | 0.396 | 0.951 | 0.950 | 0.9967 | 0.051 | 0.971 |
48
+ | 31 | 0.360 | 0.949 | 0.949 | 0.9937 | 0.051 | 0.968 |
49
+
50
+ ### sarm_t5_mse_solo_o1, draw2 (512 captions, 5,988 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 | 0.865 | 0.864 | 0.864 | 0.9998 | 0.385 | 0.796 |
55
+ | 4 | 0.803 | 0.919 | 0.919 | 0.9978 | 0.099 | 0.948 |
56
+ | 8 | 0.750 | 0.937 | 0.938 | 0.9981 | 0.085 | 0.957 |
57
+ | 12 | 0.661 | 0.945 | 0.946 | 0.9967 | 0.072 | 0.964 |
58
+ | 16 | 0.586 | 0.954 | 0.954 | 0.9974 | 0.060 | 0.970 |
59
+ | 18 | 0.555 | 0.958 | 0.957 | 0.9973 | 0.059 | 0.971 |
60
+ | 20 | 0.526 | 0.960 | 0.959 | 0.9975 | 0.057 | 0.972 |
61
+ | 22 | 0.508 | 0.964 | 0.963 | 0.9976 | 0.056 | 0.974 |
62
+ | 24 | 0.489 | 0.961 | 0.960 | 0.9976 | 0.053 | 0.975 |
63
+ | 26 | 0.469 | 0.963 | 0.962 | 0.9964 | 0.052 | 0.975 |
64
+ | 28 | 0.457 | 0.959 | 0.958 | 0.9965 | 0.055 | 0.972 |
65
+ | 29 | 0.446 | 0.956 | 0.955 | 0.9955 | 0.056 | 0.971 |
66
+ | 30 | 0.426 | 0.950 | 0.948 | 0.9948 | 0.060 | 0.968 |
67
+ | 31 | 0.389 | 0.941 | 0.942 | 0.9917 | 0.057 | 0.965 |
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
+ ### sarm_t5_mse_solo_o1
74
+
75
+ | step | task loss | mse | cos | InfoNCE | quiet on A | quiet off-domain |
76
+ |---|---|---|---|---|---|---|
77
+ | 1 | 1.5949 | 1.595 | 0.323 | None | 0.00000 | 0.00000 |
78
+ | 50 | 1.5573 | 1.557 | 0.338 | None | 0.00030 | 0.00023 |
79
+ | 250 | 1.1463 | 1.146 | 0.464 | None | 0.00138 | 0.00085 |
80
+ | 500 | 0.7512 | 0.751 | 0.635 | None | 0.00054 | 0.00039 |
81
+ | 1000 | 0.4755 | 0.475 | 0.762 | None | 0.00072 | 0.00069 |
82
+ | 1500 | 0.3119 | 0.312 | 0.842 | None | 0.00070 | 0.00050 |
83
+ | 2000 | 0.2057 | 0.206 | 0.895 | None | 0.00086 | 0.00065 |
84
+ | 2500 | 0.1585 | 0.158 | 0.919 | None | 0.00087 | 0.00058 |
85
+ | 3000 | 0.1210 | 0.121 | 0.938 | None | 0.00088 | 0.00056 |
86
+ | 3500 | 0.0746 | 0.075 | 0.962 | None | 0.00077 | 0.00053 |
87
+ | 4000 | 0.0653 | 0.065 | 0.967 | None | 0.00070 | 0.00048 |
88
+ | 1 | 1.5949 | 1.595 | 0.323 | None | 0.00000 | 0.00000 |
89
+ | 50 | 1.5573 | 1.557 | 0.338 | None | 0.00030 | 0.00023 |
90
+ | 250 | 1.1463 | 1.146 | 0.464 | None | 0.00138 | 0.00085 |
91
+ | 500 | 0.7512 | 0.751 | 0.635 | None | 0.00054 | 0.00039 |
92
+ | 1000 | 0.4755 | 0.475 | 0.762 | None | 0.00072 | 0.00069 |
93
+ | 1500 | 0.3119 | 0.312 | 0.842 | None | 0.00070 | 0.00050 |
94
+ | 2000 | 0.2057 | 0.206 | 0.895 | None | 0.00086 | 0.00065 |
95
+ | 2500 | 0.1585 | 0.158 | 0.919 | None | 0.00087 | 0.00058 |
96
+ | 3000 | 0.1210 | 0.121 | 0.938 | None | 0.00088 | 0.00056 |
97
+ | 3500 | 0.0746 | 0.075 | 0.962 | None | 0.00077 | 0.00053 |
98
+ | 4000 | 0.0653 | 0.065 | 0.967 | None | 0.00070 | 0.00048 |
99
+ | 1 | 1.5949 | 1.595 | 0.323 | None | 0.00000 | 0.00000 |
100
+ | 50 | 1.5573 | 1.557 | 0.338 | None | 0.00030 | 0.00023 |
101
+ | 250 | 1.1463 | 1.146 | 0.464 | None | 0.00138 | 0.00085 |
102
+ | 500 | 0.7512 | 0.751 | 0.635 | None | 0.00054 | 0.00039 |
103
+ | 1000 | 0.4755 | 0.475 | 0.762 | None | 0.00072 | 0.00069 |
104
+ | 1500 | 0.3119 | 0.312 | 0.842 | None | 0.00070 | 0.00050 |
105
+ | 2000 | 0.2057 | 0.206 | 0.895 | None | 0.00086 | 0.00065 |
106
+ | 2500 | 0.1585 | 0.158 | 0.919 | None | 0.00087 | 0.00058 |
107
+ | 3000 | 0.1210 | 0.121 | 0.938 | None | 0.00088 | 0.00056 |
108
+ | 3500 | 0.0746 | 0.075 | 0.962 | None | 0.00077 | 0.00053 |
109
+ | 4000 | 0.0653 | 0.065 | 0.967 | None | 0.00070 | 0.00048 |
110
+ | 1 | 1.5949 | 1.595 | 0.323 | None | 0.00000 | 0.00000 |
111
+ | 50 | 1.5573 | 1.557 | 0.338 | None | 0.00030 | 0.00023 |
112
+ | 250 | 1.1463 | 1.146 | 0.464 | None | 0.00138 | 0.00085 |
113
+ | 500 | 0.7512 | 0.751 | 0.635 | None | 0.00054 | 0.00039 |
114
+ | 1000 | 0.4755 | 0.475 | 0.762 | None | 0.00072 | 0.00069 |
115
+ | 1500 | 0.3119 | 0.312 | 0.842 | None | 0.00070 | 0.00050 |
116
+ | 2000 | 0.2057 | 0.206 | 0.895 | None | 0.00086 | 0.00065 |
117
+ | 2500 | 0.1585 | 0.158 | 0.919 | None | 0.00087 | 0.00058 |
118
+ | 3000 | 0.1210 | 0.121 | 0.938 | None | 0.00088 | 0.00056 |
119
+ | 3500 | 0.0746 | 0.075 | 0.962 | None | 0.00077 | 0.00053 |
120
+ | 4000 | 0.0653 | 0.065 | 0.967 | None | 0.00070 | 0.00048 |
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: t5-v1_1-xxl's spelling (the sentence-piece convention; sha256 170798cc83dec7e6..., vocab 32,100; A = the text written; sites = tokens); 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/t5/mse_solo_o1/results.json ADDED
The diff for this file is too large to render. See raw diff
 
surface/t5/mse_solo_o1/s11_t5.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d84548fcef5603b78c74fc0ce41386cee5f9ba6d266974ed228064e3b7d79d23
3
+ size 54819296
surface/t5/mse_solo_o1/stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0d268fa2a2bfc825f0be2297a92742b956824a28fb5d1198b78cda079130686d
3
+ size 264309
surface/t5/mse_solo_o1/train.log ADDED
@@ -0,0 +1,160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [ 0.4s] THE JOB: sarm_t5_mse_solo_o1: the surface arm s11_t5 (sentencepiece spelling by google/t5-v1_1-xxl; A = the text written; sites = tokens) 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
+ [ 17.1s] tokenizer google/t5-v1_1-xxl (sentencepiece spelling; vocab 32100; sha256 170798cc83dec7e6): 'a taco truck' -> ['▁', 'a', '▁tac', 'o', '▁truck']; A 12 bytes, B 19 bytes; sites [(1, 1, 4), (2, 5, 10), (3, 6, 11)]
5
+ [ 17.1s] the read's captions: 512 of each draw of /workspace/qarm/dl/coco_caps_2x2048.json (4,090 captions of both draws kept out of training)
6
+ [ 30.0s] 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 3.7 GB
7
+ [ 30.1s] streams: web text; partners none (two web rows a quiet chunk); held-out sets ['web'] x 8 batches
8
+ [ 30.2s] read rows draw1: 512 captions (0 unspellable), A rows (512, 99), B rows (512, 136), 5,707 sites
9
+ [ 30.2s] read rows draw2: 512 captions (0 unspellable), A rows (512, 151), B rows (512, 204), 5,988 sites
10
+ [arms] attached 's11_t5' for phase 'sentencepiece_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
+ [ 41.7s] 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
+ [ 45.7s] standardization statistics from 12,131 teacher sites on 8 paired batches (a mean and a spread per block and dimension); sigma range at block 20 0.748..1.112; saved /workspace/qarm/out/anchors/sarm_t5_mse_solo_o1_stats.pt
13
+ [ 262.5s] CLOSE 0 (step 0): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
14
+ [ 262.5s] block 8: 0.735 -> 0.735 / 0.735; A 1.0000; mse 1.130 cos 0.491
15
+ [ 262.5s] block 12: 0.647 -> 0.647 / 0.647; A 1.0000; mse 1.195 cos 0.439
16
+ [ 262.5s] block 16: 0.572 -> 0.572 / 0.572; A 1.0000; mse 1.302 cos 0.388
17
+ [ 262.5s] block 18: 0.545 -> 0.545 / 0.545; A 1.0000; mse 1.325 cos 0.367
18
+ [ 262.5s] block 20: 0.519 -> 0.519 / 0.519; A 1.0000; mse 1.446 cos 0.315
19
+ [ 262.5s] block 24: 0.470 -> 0.470 / 0.470; A 1.0000; mse 1.702 cos 0.250
20
+ [ 262.5s] block 28: 0.435 -> 0.435 / 0.435; A 1.0000; mse 1.950 cos 0.176
21
+ [ 262.5s] block 31: 0.360 -> 0.360 / 0.360; A 1.0000; mse 3.310 cos 0.112
22
+ [ 262.5s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.661 / 0.586 / 0.526 / 0.489 / 0.457
23
+ [ 262.5s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.360 against the bar .85; B2: the lowest A alignment 1.0000 against .995; fineweb share +0.00000 +- 0.00000 against +.012
24
+ [ 262.5s] 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 2.3 GB; host RSS 11.7 GB
25
+ [stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
26
+ [ 269.3s] step 1/4000: task 1.5949 (mse 1.595 cos 0.323); quiet A 0.00000 off 0.00000; 1,521 sites; 6.83 s/step; the job: 4.5 min spent, about 455.3 min of steps left; card peak 13.7 GB
27
+ [ 390.0s] step 50/4000: task 1.5573 (mse 1.557 cos 0.338); quiet A 0.00030 off 0.00023; 1,501 sites; 2.55 s/step; the job: 6.5 min spent, about 167.9 min of steps left; card peak 13.9 GB
28
+ [ 512.9s] step 100/4000: task 1.5097 (mse 1.510 cos 0.353); quiet A 0.00044 off 0.00032; 1,565 sites; 2.50 s/step; the job: 8.5 min spent, about 162.7 min of steps left; card peak 13.9 GB
29
+ [ 635.8s] step 150/4000: task 1.4394 (mse 1.439 cos 0.375); quiet A 0.00048 off 0.00035; 1,532 sites; 2.49 s/step; the job: 10.6 min spent, about 159.7 min of steps left; card peak 13.9 GB
30
+ [ 758.7s] step 200/4000: task 1.3316 (mse 1.332 cos 0.406); quiet A 0.00077 off 0.00057; 1,546 sites; 2.48 s/step; the job: 12.6 min spent, about 157.1 min of steps left; card peak 13.9 GB
31
+ [ 881.4s] step 250/4000: task 1.1463 (mse 1.146 cos 0.464); quiet A 0.00138 off 0.00085; 1,469 sites; 2.48 s/step; the job: 14.7 min spent, about 154.7 min of steps left; card peak 13.9 GB
32
+ [transformers] Token indices sequence length is longer than the specified maximum sequence length for this model (539 > 512). Running this sequence through the model will result in indexing errors
33
+ [ 1004.3s] step 300/4000: task 0.9889 (mse 0.989 cos 0.522); quiet A 0.00138 off 0.00073; 1,468 sites; 2.47 s/step; the job: 16.7 min spent, about 152.5 min of steps left; card peak 13.9 GB
34
+ [ 1127.9s] step 350/4000: task 0.8932 (mse 0.893 cos 0.567); quiet A 0.00089 off 0.00049; 1,502 sites; 2.47 s/step; the job: 18.8 min spent, about 150.4 min of steps left; card peak 13.9 GB
35
+ [ 1251.2s] step 400/4000: task 0.8414 (mse 0.841 cos 0.592); quiet A 0.00052 off 0.00035; 1,543 sites; 2.47 s/step; the job: 20.9 min spent, about 148.3 min of steps left; card peak 13.9 GB
36
+ [ 1374.6s] step 450/4000: task 0.7974 (mse 0.797 cos 0.612); quiet A 0.00063 off 0.00037; 1,479 sites; 2.47 s/step; the job: 22.9 min spent, about 146.2 min of steps left; card peak 13.9 GB
37
+ [ 1497.9s] step 500/4000: task 0.7512 (mse 0.751 cos 0.635); quiet A 0.00054 off 0.00039; 1,512 sites; 2.47 s/step; the job: 25.0 min spent, about 144.1 min of steps left; card peak 13.9 GB
38
+ [ 1620.8s] step 550/4000: task 0.7118 (mse 0.712 cos 0.652); quiet A 0.00053 off 0.00034; 1,476 sites; 2.47 s/step; the job: 27.0 min spent, about 142.0 min of steps left; card peak 13.9 GB
39
+ [ 1743.9s] step 600/4000: task 0.6785 (mse 0.678 cos 0.669); quiet A 0.00055 off 0.00035; 1,514 sites; 2.47 s/step; the job: 29.1 min spent, about 139.9 min of steps left; card peak 13.9 GB
40
+ [ 1866.6s] step 650/4000: task 0.6472 (mse 0.647 cos 0.683); quiet A 0.00056 off 0.00036; 1,496 sites; 2.47 s/step; the job: 31.1 min spent, about 137.8 min of steps left; card peak 13.9 GB
41
+ [ 1989.3s] step 700/4000: task 0.6208 (mse 0.621 cos 0.695); quiet A 0.00056 off 0.00037; 1,512 sites; 2.47 s/step; the job: 33.2 min spent, about 135.7 min of steps left; card peak 13.9 GB
42
+ [ 2112.1s] step 750/4000: task 0.5911 (mse 0.591 cos 0.709); quiet A 0.00057 off 0.00038; 1,497 sites; 2.47 s/step; the job: 35.2 min spent, about 133.6 min of steps left; card peak 13.9 GB
43
+ [ 2235.1s] step 800/4000: task 0.5666 (mse 0.567 cos 0.720); quiet A 0.00061 off 0.00044; 1,669 sites; 2.47 s/step; the job: 37.3 min spent, about 131.5 min of steps left; card peak 13.9 GB
44
+ [ 2358.0s] step 850/4000: task 0.5410 (mse 0.541 cos 0.732); quiet A 0.00063 off 0.00038; 1,579 sites; 2.47 s/step; the job: 39.3 min spent, about 129.4 min of steps left; card peak 13.9 GB
45
+ [ 2481.4s] step 900/4000: task 0.5162 (mse 0.516 cos 0.743); quiet A 0.00062 off 0.00038; 1,518 sites; 2.47 s/step; the job: 41.4 min spent, about 127.4 min of steps left; card peak 13.9 GB
46
+ [ 2604.8s] step 950/4000: task 0.4934 (mse 0.493 cos 0.754); quiet A 0.00062 off 0.00045; 1,468 sites; 2.47 s/step; the job: 43.4 min spent, about 125.3 min of steps left; card peak 13.9 GB
47
+ [ 2728.1s] step 1000/4000: task 0.4755 (mse 0.475 cos 0.762); quiet A 0.00072 off 0.00069; 1,483 sites; 2.47 s/step; the job: 45.5 min spent, about 123.3 min of steps left; card peak 13.9 GB
48
+ [ 2728.4s] anchor saved: /workspace/qarm/out/anchors/sarm_t5_mse_solo_o1_close1_s11_t5.safetensors
49
+ [ 2890.9s] CLOSE 1 (step 1000): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
50
+ [ 2890.9s] block 8: 0.735 -> 0.842 / 0.843; A 0.9922; mse 0.374 cos 0.805
51
+ [ 2890.9s] block 12: 0.647 -> 0.848 / 0.850; A 0.9954; mse 0.316 cos 0.837
52
+ [ 2890.9s] block 16: 0.572 -> 0.855 / 0.857; A 0.9958; mse 0.270 cos 0.863
53
+ [ 2890.9s] block 18: 0.545 -> 0.862 / 0.863; A 0.9956; mse 0.255 cos 0.873
54
+ [ 2890.9s] block 20: 0.519 -> 0.867 / 0.868; A 0.9957; mse 0.241 cos 0.882
55
+ [ 2890.9s] block 24: 0.470 -> 0.875 / 0.876; A 0.9948; mse 0.215 cos 0.898
56
+ [ 2890.9s] block 28: 0.435 -> 0.859 / 0.860; A 0.9935; mse 0.220 cos 0.887
57
+ [ 2890.9s] block 31: 0.360 -> 0.827 / 0.825; A 0.9890; mse 0.208 cos 0.867
58
+ [ 2890.9s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.849 / 0.852 / 0.865 / 0.871 / 0.857
59
+ [ 2890.9s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.827 against the bar .85; B2: the lowest A alignment 0.9890 against .995; fineweb share +0.00015 +- 0.00008 against +.012
60
+ [ 2890.9s] gate mean 0.108 / widest 0.147; 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 13.9 GB; host RSS 14.9 GB
61
+ [ 3014.3s] step 1050/4000: task 0.4529 (mse 0.453 cos 0.773); quiet A 0.00065 off 0.00042; 1,517 sites; 2.47 s/step; the job: 50.2 min spent, about 121.2 min of steps left; card peak 13.9 GB
62
+ [ 3137.6s] step 1100/4000: task 0.4350 (mse 0.435 cos 0.782); quiet A 0.00064 off 0.00043; 1,570 sites; 2.47 s/step; the job: 52.3 min spent, about 119.2 min of steps left; card peak 13.9 GB
63
+ [ 3260.8s] step 1150/4000: task 0.4142 (mse 0.414 cos 0.792); quiet A 0.00064 off 0.00041; 1,520 sites; 2.47 s/step; the job: 54.3 min spent, about 117.1 min of steps left; card peak 13.9 GB
64
+ [ 3384.1s] step 1200/4000: task 0.3961 (mse 0.396 cos 0.800); quiet A 0.00066 off 0.00041; 1,559 sites; 2.47 s/step; the job: 56.4 min spent, about 115.1 min of steps left; card peak 13.9 GB
65
+ [ 3506.9s] step 1250/4000: task 0.3823 (mse 0.382 cos 0.807); quiet A 0.00067 off 0.00044; 1,526 sites; 2.46 s/step; the job: 58.4 min spent, about 113.0 min of steps left; card peak 13.9 GB
66
+ [ 3629.7s] step 1300/4000: task 0.3669 (mse 0.367 cos 0.815); quiet A 0.00067 off 0.00046; 1,522 sites; 2.46 s/step; the job: 60.5 min spent, about 110.9 min of steps left; card peak 13.9 GB
67
+ [ 3753.2s] step 1350/4000: task 0.3512 (mse 0.351 cos 0.822); quiet A 0.00067 off 0.00047; 1,513 sites; 2.46 s/step; the job: 62.6 min spent, about 108.9 min of steps left; card peak 13.9 GB
68
+ [ 3876.4s] step 1400/4000: task 0.3382 (mse 0.338 cos 0.829); quiet A 0.00068 off 0.00047; 1,440 sites; 2.46 s/step; the job: 64.6 min spent, about 106.8 min of steps left; card peak 13.9 GB
69
+ [ 3999.6s] step 1450/4000: task 0.3253 (mse 0.325 cos 0.836); quiet A 0.00068 off 0.00046; 1,521 sites; 2.46 s/step; the job: 66.7 min spent, about 104.8 min of steps left; card peak 13.9 GB
70
+ [ 4122.0s] step 1500/4000: task 0.3119 (mse 0.312 cos 0.842); quiet A 0.00070 off 0.00050; 1,462 sites; 2.46 s/step; the job: 68.7 min spent, about 102.7 min of steps left; card peak 13.9 GB
71
+ [ 4244.7s] step 1550/4000: task 0.2976 (mse 0.298 cos 0.849); quiet A 0.00072 off 0.00052; 1,538 sites; 2.46 s/step; the job: 70.7 min spent, about 100.6 min of steps left; card peak 13.9 GB
72
+ [ 4367.5s] step 1600/4000: task 0.2878 (mse 0.288 cos 0.854); quiet A 0.00073 off 0.00052; 1,513 sites; 2.46 s/step; the job: 72.8 min spent, about 98.6 min of steps left; card peak 13.9 GB
73
+ [ 4490.5s] step 1650/4000: task 0.2770 (mse 0.277 cos 0.859); quiet A 0.00076 off 0.00054; 1,508 sites; 2.46 s/step; the job: 74.8 min spent, about 96.5 min of steps left; card peak 13.9 GB
74
+ [ 4613.8s] step 1700/4000: task 0.2650 (mse 0.265 cos 0.866); quiet A 0.00080 off 0.00063; 1,591 sites; 2.46 s/step; the job: 76.9 min spent, about 94.4 min of steps left; card peak 13.9 GB
75
+ [ 4736.9s] step 1750/4000: task 0.2534 (mse 0.253 cos 0.871); quiet A 0.00081 off 0.00056; 1,578 sites; 2.46 s/step; the job: 78.9 min spent, about 92.4 min of steps left; card peak 13.9 GB
76
+ [ 4859.7s] step 1800/4000: task 0.2434 (mse 0.243 cos 0.876); quiet A 0.00081 off 0.00053; 1,575 sites; 2.46 s/step; the job: 81.0 min spent, about 90.3 min of steps left; card peak 13.9 GB
77
+ [ 4983.1s] step 1850/4000: task 0.2348 (mse 0.235 cos 0.880); quiet A 0.00079 off 0.00055; 1,461 sites; 2.46 s/step; the job: 83.1 min spent, about 88.3 min of steps left; card peak 13.9 GB
78
+ [ 5106.5s] step 1900/4000: task 0.2243 (mse 0.224 cos 0.886); quiet A 0.00085 off 0.00060; 1,518 sites; 2.46 s/step; the job: 85.1 min spent, about 86.2 min of steps left; card peak 13.9 GB
79
+ [ 5229.4s] step 1950/4000: task 0.2142 (mse 0.214 cos 0.891); quiet A 0.00086 off 0.00057; 1,462 sites; 2.46 s/step; the job: 87.2 min spent, about 84.2 min of steps left; card peak 13.9 GB
80
+ [ 5352.2s] step 2000/4000: task 0.2057 (mse 0.206 cos 0.895); quiet A 0.00086 off 0.00065; 1,474 sites; 2.46 s/step; the job: 89.2 min spent, about 82.1 min of steps left; card peak 13.9 GB
81
+ [ 5352.5s] anchor saved: /workspace/qarm/out/anchors/sarm_t5_mse_solo_o1_close2_s11_t5.safetensors
82
+ [ 5514.6s] CLOSE 2 (step 2000): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
83
+ [ 5514.6s] block 8: 0.735 -> 0.906 / 0.906; A 0.9965; mse 0.159 cos 0.917
84
+ [ 5514.6s] block 12: 0.647 -> 0.919 / 0.920; A 0.9960; mse 0.129 cos 0.934
85
+ [ 5514.6s] block 16: 0.572 -> 0.934 / 0.935; A 0.9965; mse 0.104 cos 0.948
86
+ [ 5514.6s] block 18: 0.545 -> 0.935 / 0.935; A 0.9971; mse 0.100 cos 0.951
87
+ [ 5514.6s] block 20: 0.519 -> 0.942 / 0.942; A 0.9966; mse 0.094 cos 0.955
88
+ [ 5514.6s] block 24: 0.470 -> 0.944 / 0.945; A 0.9962; mse 0.084 cos 0.961
89
+ [ 5514.6s] block 28: 0.435 -> 0.937 / 0.937; A 0.9952; mse 0.085 cos 0.958
90
+ [ 5514.6s] block 31: 0.360 -> 0.919 / 0.920; A 0.9916; mse 0.084 cos 0.948
91
+ [ 5514.6s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.919 / 0.930 / 0.936 / 0.938 / 0.934
92
+ [ 5514.6s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.919 against the bar .85; B2: the lowest A alignment 0.9901 against .995; fineweb share +0.00022 +- 0.00016 against +.012
93
+ [ 5514.6s] gate mean 0.142 / widest 0.290; 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 13.9 GB; host RSS 15.0 GB
94
+ [ 5637.7s] step 2050/4000: task 0.1968 (mse 0.197 cos 0.900); quiet A 0.00086 off 0.00062; 1,469 sites; 2.46 s/step; the job: 94.0 min spent, about 80.1 min of steps left; card peak 13.9 GB
95
+ [ 5760.7s] step 2100/4000: task 0.1905 (mse 0.191 cos 0.903); quiet A 0.00087 off 0.00059; 1,543 sites; 2.46 s/step; the job: 96.0 min spent, about 78.0 min of steps left; card peak 13.9 GB
96
+ [ 5884.0s] step 2150/4000: task 0.1880 (mse 0.188 cos 0.904); quiet A 0.00086 off 0.00060; 1,500 sites; 2.46 s/step; the job: 98.1 min spent, about 76.0 min of steps left; card peak 13.9 GB
97
+ [ 6006.9s] step 2200/4000: task 0.1825 (mse 0.182 cos 0.907); quiet A 0.00089 off 0.00070; 1,567 sites; 2.46 s/step; the job: 100.1 min spent, about 73.9 min of steps left; card peak 13.9 GB
98
+ [ 6130.1s] step 2250/4000: task 0.1771 (mse 0.177 cos 0.910); quiet A 0.00088 off 0.00072; 1,479 sites; 2.46 s/step; the job: 102.2 min spent, about 71.8 min of steps left; card peak 13.9 GB
99
+ [ 6253.4s] step 2300/4000: task 0.1737 (mse 0.174 cos 0.912); quiet A 0.00093 off 0.00068; 1,502 sites; 2.46 s/step; the job: 104.2 min spent, about 69.8 min of steps left; card peak 13.9 GB
100
+ [ 6377.1s] step 2350/4000: task 0.1693 (mse 0.169 cos 0.914); quiet A 0.00092 off 0.00068; 1,530 sites; 2.46 s/step; the job: 106.3 min spent, about 67.7 min of steps left; card peak 13.9 GB
101
+ [ 6500.3s] step 2400/4000: task 0.1661 (mse 0.166 cos 0.916); quiet A 0.00091 off 0.00056; 1,541 sites; 2.46 s/step; the job: 108.3 min spent, about 65.7 min of steps left; card peak 13.9 GB
102
+ [ 6623.7s] step 2450/4000: task 0.1621 (mse 0.162 cos 0.918); quiet A 0.00086 off 0.00064; 1,520 sites; 2.46 s/step; the job: 110.4 min spent, about 63.6 min of steps left; card peak 13.9 GB
103
+ [ 6747.1s] step 2500/4000: task 0.1585 (mse 0.158 cos 0.919); quiet A 0.00087 off 0.00058; 1,572 sites; 2.46 s/step; the job: 112.5 min spent, about 61.6 min of steps left; card peak 13.9 GB
104
+ [ 6870.6s] step 2550/4000: task 0.1552 (mse 0.155 cos 0.921); quiet A 0.00089 off 0.00057; 1,501 sites; 2.47 s/step; the job: 114.5 min spent, about 59.5 min of steps left; card peak 13.9 GB
105
+ [ 6993.9s] step 2600/4000: task 0.1546 (mse 0.155 cos 0.922); quiet A 0.00087 off 0.00059; 1,496 sites; 2.47 s/step; the job: 116.6 min spent, about 57.5 min of steps left; card peak 13.9 GB
106
+ [ 7117.1s] step 2650/4000: task 0.1467 (mse 0.147 cos 0.925); quiet A 0.00080 off 0.00058; 1,466 sites; 2.47 s/step; the job: 118.6 min spent, about 55.4 min of steps left; card peak 13.9 GB
107
+ [ 7240.5s] step 2700/4000: task 0.1444 (mse 0.144 cos 0.926); quiet A 0.00083 off 0.00054; 1,504 sites; 2.47 s/step; the job: 120.7 min spent, about 53.4 min of steps left; card peak 13.9 GB
108
+ [ 7363.9s] step 2750/4000: task 0.1432 (mse 0.143 cos 0.927); quiet A 0.00083 off 0.00060; 1,500 sites; 2.47 s/step; the job: 122.7 min spent, about 51.3 min of steps left; card peak 13.9 GB
109
+ [ 7487.3s] step 2800/4000: task 0.1387 (mse 0.139 cos 0.929); quiet A 0.00082 off 0.00059; 1,539 sites; 2.47 s/step; the job: 124.8 min spent, about 49.3 min of steps left; card peak 13.9 GB
110
+ [ 7610.3s] step 2850/4000: task 0.1367 (mse 0.137 cos 0.930); quiet A 0.00087 off 0.00061; 1,458 sites; 2.47 s/step; the job: 126.8 min spent, about 47.2 min of steps left; card peak 13.9 GB
111
+ [ 7734.0s] step 2900/4000: task 0.1328 (mse 0.133 cos 0.932); quiet A 0.00086 off 0.00062; 1,510 sites; 2.47 s/step; the job: 128.9 min spent, about 45.2 min of steps left; card peak 13.9 GB
112
+ [ 7857.2s] step 2950/4000: task 0.1280 (mse 0.128 cos 0.935); quiet A 0.00086 off 0.00056; 1,518 sites; 2.47 s/step; the job: 131.0 min spent, about 43.1 min of steps left; card peak 13.9 GB
113
+ [ 7980.6s] step 3000/4000: task 0.1210 (mse 0.121 cos 0.938); quiet A 0.00088 off 0.00056; 1,459 sites; 2.47 s/step; the job: 133.0 min spent, about 41.1 min of steps left; card peak 13.9 GB
114
+ [ 7980.9s] anchor saved: /workspace/qarm/out/anchors/sarm_t5_mse_solo_o1_close3_s11_t5.safetensors
115
+ [ 8143.3s] CLOSE 3 (step 3000): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
116
+ [ 8143.3s] block 8: 0.735 -> 0.926 / 0.926; A 0.9973; mse 0.101 cos 0.947
117
+ [ 8143.3s] block 12: 0.647 -> 0.937 / 0.937; A 0.9962; mse 0.085 cos 0.957
118
+ [ 8143.3s] block 16: 0.572 -> 0.952 / 0.953; A 0.9956; mse 0.069 cos 0.966
119
+ [ 8143.3s] block 18: 0.545 -> 0.952 / 0.952; A 0.9971; mse 0.068 cos 0.967
120
+ [ 8143.3s] block 20: 0.519 -> 0.958 / 0.956; A 0.9964; mse 0.065 cos 0.969
121
+ [ 8143.3s] block 24: 0.470 -> 0.957 / 0.956; A 0.9960; mse 0.059 cos 0.973
122
+ [ 8143.3s] block 28: 0.435 -> 0.953 / 0.951; A 0.9950; mse 0.060 cos 0.970
123
+ [ 8143.3s] block 31: 0.360 -> 0.940 / 0.939; A 0.9933; mse 0.061 cos 0.962
124
+ [ 8143.3s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.937 / 0.948 / 0.951 / 0.955 / 0.951
125
+ [ 8143.3s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.937 against the bar .85; B2: the lowest A alignment 0.9895 against .995; fineweb share +0.00037 +- 0.00020 against +.012
126
+ [ 8143.3s] gate mean 0.168 / widest 0.402; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 96,833 (skipped {'test': 137, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 15.1 GB
127
+ [ 8266.5s] step 3050/4000: task 0.1188 (mse 0.119 cos 0.940); quiet A 0.00090 off 0.00057; 1,519 sites; 2.46 s/step; the job: 137.8 min spent, about 39.0 min of steps left; card peak 13.9 GB
128
+ [ 8390.0s] step 3100/4000: task 0.1047 (mse 0.105 cos 0.947); quiet A 0.00088 off 0.00115; 1,485 sites; 2.47 s/step; the job: 139.8 min spent, about 37.0 min of steps left; card peak 13.9 GB
129
+ [ 8513.4s] step 3150/4000: task 0.1007 (mse 0.101 cos 0.949); quiet A 0.00096 off 0.00072; 1,490 sites; 2.47 s/step; the job: 141.9 min spent, about 34.9 min of steps left; card peak 13.9 GB
130
+ [ 8636.8s] step 3200/4000: task 0.0935 (mse 0.094 cos 0.952); quiet A 0.00091 off 0.00065; 1,532 sites; 2.47 s/step; the job: 143.9 min spent, about 32.9 min of steps left; card peak 13.9 GB
131
+ [ 8760.0s] step 3250/4000: task 0.0883 (mse 0.088 cos 0.955); quiet A 0.00087 off 0.00063; 1,459 sites; 2.47 s/step; the job: 146.0 min spent, about 30.8 min of steps left; card peak 13.9 GB
132
+ [ 8883.3s] step 3300/4000: task 0.0852 (mse 0.085 cos 0.957); quiet A 0.00083 off 0.00063; 1,505 sites; 2.47 s/step; the job: 148.1 min spent, about 28.8 min of steps left; card peak 13.9 GB
133
+ [ 9006.3s] step 3350/4000: task 0.0812 (mse 0.081 cos 0.959); quiet A 0.00079 off 0.00054; 1,491 sites; 2.47 s/step; the job: 150.1 min spent, about 26.7 min of steps left; card peak 13.9 GB
134
+ [ 9129.4s] step 3400/4000: task 0.0793 (mse 0.079 cos 0.960); quiet A 0.00075 off 0.00055; 1,527 sites; 2.47 s/step; the job: 152.2 min spent, about 24.6 min of steps left; card peak 13.9 GB
135
+ [ 9252.9s] step 3450/4000: task 0.0760 (mse 0.076 cos 0.961); quiet A 0.00076 off 0.00054; 1,516 sites; 2.47 s/step; the job: 154.2 min spent, about 22.6 min of steps left; card peak 13.9 GB
136
+ [ 9376.0s] step 3500/4000: task 0.0746 (mse 0.075 cos 0.962); quiet A 0.00077 off 0.00053; 1,541 sites; 2.47 s/step; the job: 156.3 min spent, about 20.5 min of steps left; card peak 13.9 GB
137
+ [ 9499.3s] step 3550/4000: task 0.0736 (mse 0.074 cos 0.963); quiet A 0.00075 off 0.00064; 1,469 sites; 2.47 s/step; the job: 158.3 min spent, about 18.5 min of steps left; card peak 13.9 GB
138
+ [ 9622.2s] step 3600/4000: task 0.0743 (mse 0.074 cos 0.962); quiet A 0.00084 off 0.00081; 1,502 sites; 2.46 s/step; the job: 160.4 min spent, about 16.4 min of steps left; card peak 13.9 GB
139
+ [ 9745.3s] step 3650/4000: task 0.0710 (mse 0.071 cos 0.964); quiet A 0.00078 off 0.00057; 1,505 sites; 2.46 s/step; the job: 162.4 min spent, about 14.4 min of steps left; card peak 13.9 GB
140
+ [ 9868.9s] step 3700/4000: task 0.0697 (mse 0.070 cos 0.964); quiet A 0.00073 off 0.00061; 1,473 sites; 2.47 s/step; the job: 164.5 min spent, about 12.3 min of steps left; card peak 13.9 GB
141
+ [ 9992.4s] step 3750/4000: task 0.0682 (mse 0.068 cos 0.965); quiet A 0.00073 off 0.00051; 1,473 sites; 2.47 s/step; the job: 166.5 min spent, about 10.3 min of steps left; card peak 13.9 GB
142
+ [10115.8s] step 3800/4000: task 0.0676 (mse 0.068 cos 0.965); quiet A 0.00071 off 0.00050; 1,454 sites; 2.47 s/step; the job: 168.6 min spent, about 8.2 min of steps left; card peak 13.9 GB
143
+ [10238.9s] step 3850/4000: task 0.0680 (mse 0.068 cos 0.965); quiet A 0.00071 off 0.00049; 1,489 sites; 2.47 s/step; the job: 170.6 min spent, about 6.2 min of steps left; card peak 13.9 GB
144
+ [10362.2s] step 3900/4000: task 0.0671 (mse 0.067 cos 0.966); quiet A 0.00070 off 0.00046; 1,492 sites; 2.47 s/step; the job: 172.7 min spent, about 4.1 min of steps left; card peak 13.9 GB
145
+ [10485.8s] step 3950/4000: task 0.0652 (mse 0.065 cos 0.967); quiet A 0.00069 off 0.00052; 1,477 sites; 2.47 s/step; the job: 174.8 min spent, about 2.1 min of steps left; card peak 13.9 GB
146
+ [10609.1s] step 4000/4000: task 0.0653 (mse 0.065 cos 0.967); quiet A 0.00070 off 0.00048; 1,615 sites; 2.47 s/step; the job: 176.8 min spent, about 0.0 min of steps left; card peak 13.9 GB
147
+ [10609.3s] anchor saved: /workspace/qarm/out/anchors/sarm_t5_mse_solo_o1_close4_s11_t5.safetensors
148
+ [10771.6s] CLOSE 4 (step 4000): the member's alignment read on draw 1 (512 captions, 5,707 sites); per block: before -> armed B vs masked A (target) / armed B vs armed A; A armed vs masked
149
+ [10771.6s] block 8: 0.735 -> 0.934 / 0.935; A 0.9962; mse 0.078 cos 0.960
150
+ [10771.6s] block 12: 0.647 -> 0.948 / 0.948; A 0.9968; mse 0.067 cos 0.967
151
+ [10771.6s] block 16: 0.572 -> 0.961 / 0.961; A 0.9956; mse 0.055 cos 0.973
152
+ [10771.6s] block 18: 0.545 -> 0.961 / 0.960; A 0.9974; mse 0.054 cos 0.974
153
+ [10771.6s] block 20: 0.519 -> 0.965 / 0.963; A 0.9975; mse 0.052 cos 0.975
154
+ [10771.6s] block 24: 0.470 -> 0.965 / 0.963; A 0.9966; mse 0.049 cos 0.978
155
+ [10771.6s] block 28: 0.435 -> 0.960 / 0.957; A 0.9944; mse 0.049 cos 0.975
156
+ [10771.6s] block 31: 0.360 -> 0.949 / 0.949; A 0.9937; mse 0.051 cos 0.968
157
+ [10771.6s] draw 2 (the repeat), target alignment at 12/16/20/24/28: 0.945 / 0.954 / 0.960 / 0.961 / 0.959
158
+ [10771.6s] B1 (served blocks >= 12, draw 1): the lowest target alignment 0.948 against the bar .85; B2: the lowest A alignment 0.9925 against .995; fineweb share +0.00008 +- 0.00012 against +.012
159
+ [10771.6s] gate mean 0.188 / widest 0.487; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 128,943 (skipped {'test': 195, 'spelling': 2, 'empty': 0, 'short': 0}); card peak 13.9 GB; host RSS 15.1 GB
160
+ [10771.7s] DONE sarm_t5_mse_solo_o1: 4000 steps in 164.3 min of steps (2.46 s/step); 4 closes; 10772 s in all; card peak 13.9 GB; host RSS 15.1 GB; compiled chain: 5 cache entries (limit 256), 0 limit warnings