Hugging Face's logo

AbstractPhil
/
alephllm-mini-beatrix-training

TensorBoard
English
aleph
mixture-of-experts
linear-attention
byte-level
research-log
Model card Files
xet
 Metrics Community
AbstractPhil commited on
Commit
d9cb037
·
verified ·
1 Parent(s): 492bd1b

mini-beatrix-3: the hub read (her hub blackboards against her stream as conditioning features; caption geometry and the slider read, 10 conditions)

Browse files
Files changed (4) hide show
  1. mini-beatrix-3/reports/v3/hub_read/README.md +69 -0
  2. mini-beatrix-3/reports/v3/hub_read/hub_read_step245674.json +0 -0
  3. mini-beatrix-3/reports/v3/hub_read/hub_read_step245674.md +206 -0
  4. mini-beatrix-3/reports/v3/hub_read/hub_read_step245674_smoke.json +0 -0
mini-beatrix-3/reports/v3/hub_read/README.md ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # The hub read: mini-beatrix-3's hub blackboards against her stream
2
+
3
+ Every block of mini-beatrix-3 carries a hub whose blackboard (4 constellations x 64 slots x 1,024 numbers) holds the block's memory of the text so far. This read asks whether the blackboard is a better conditioning feature than the stream (the block's output, pooled over the caption, at its last byte, or at a closing full stop), measured on the local RTX 4090 with the method of the earlier conditioning probe and the gauge suite, then on the slider instruments of the Anima slider experiments. The decision rule was registered before the read.
4
+
5
+ ## What was measured
6
+
7
+ | part | input | numbers |
8
+ |---|---|---|
9
+ | A, caption geometry | 2,048 COCO captions (a second draw of 2,048 for the noise bar) | against CLIP ViT-B/32: top-10 neighbour overlap and Spearman over 100,000 pairs (the earlier probe's method); against Anima's adapter output, T5-XXL and Qwen3 0.6B: debiased CKA and forward ridge R2 |
10
+ | B, the slider read | 74 mood phrases (60 training, 8 held out, neutral phrases) | the mood axis from the training phrases; leave-one-out signs (of 60), held-out signs (of 8), held-out nearest centre (of 10), separation (mean slider value of the held-out cheerful phrases minus the gloomy ones), balance (the log ratio of the two sides' spreads; 0 is even), a 10-fold probe |
11
+
12
+ ## Conditions (the keys in the ledger)
13
+
14
+ | key | condition |
15
+ |---|---|
16
+ | M0 | the trained model, bare (step 245,674) |
17
+ | M0_draw2 | the same, a second draw of 2,048 captions (the noise bar) |
18
+ | M9_A / M9_B | the nine trained arms mounted (eight stage arms and the caption arm), two arm seeds |
19
+ | M8_A / M8_B | the eight stage arms alone |
20
+ | Mc_A / Mc_B | the caption arm alone |
21
+ | U0 | an untrained copy (random init, seed 0): a control that must fail |
22
+ | A212 | the model at step 212,000, the trunk of the earlier slider experiments (part B only) |
23
+
24
+ ## Part A: the blackboard against the pooled stream (the trained model, bare)
25
+
26
+ | block | against CLIP-B/32, blackboard | against CLIP-B/32, pooled stream | Anima's adapter output, blackboard: CKA (ridge R2) | the adapter output, pooled stream |
27
+ |---|---|---|---|---|
28
+ | 12 | 0.373 / +0.414 | 0.276 / +0.216 | 0.813 (+0.625) | 0.729 (+0.551) |
29
+ | 14 | 0.414 / +0.476 | 0.320 / +0.335 | 0.827 (+0.635) | 0.803 (+0.563) |
30
+ | 22 | 0.335 / +0.367 | 0.250 / +0.288 | 0.814 (+0.636) | 0.753 (+0.556) |
31
+ | 24 | 0.383 / +0.418 | 0.271 / +0.330 | 0.825 (+0.640) | 0.790 (+0.554) |
32
+
33
+ The noise bar (draw 1 against draw 2, every tap and reference): overlap 0.009 mean / 0.030 at most; Spearman 0.017 / 0.057; CKA 0.026 / 0.151; ridge 0.011 / 0.037. Draw 2 repeats the lead: blackboard block 14 0.394 / +0.511 against pooled 0.305 / +0.340.
34
+
35
+ Of the ten blocks read (8-28), the blackboard leads the pooled stream against CLIP-B/32 at 10, in ridge R2 to the adapter output at 10, and in CKA with the adapter output at 6 (every table: hub_read_step245674.md).
36
+
37
+ ## Part B: the decided readings under every condition
38
+
39
+ Each cell: separation; leave-one-out signs (of 60); held-out signs (of 8); balance.
40
+
41
+ | reading | M0 | M9_A | M9_B | M8_A | M8_B | Mc_A | Mc_B | U0 | A212 |
42
+ |---|---|---|---|---|---|---|---|---|---|
43
+ | close/stream/18 | +2.19; 56; 8; 0.41 | +1.82; 52; 8; 0.43 | +1.94; 53; 8; 0.51 | +1.87; 52; 8; 0.64 | +2.02; 53; 7; 0.72 | +1.93; 58; 8; 0.14 | +1.86; 56; 8; 0.36 | +0.15; 42; 6; 0.29 | +1.64; 54; 7; 0.10 |
44
+ | close/hub/22 | +1.83; 55; 8; 0.13 | +1.53; 58; 7; 0.05 | +1.67; 55; 7; 0.00 | +1.67; 56; 7; 0.06 | +1.80; 54; 8; 0.15 | +1.72; 58; 8; 0.29 | +1.67; 58; 8; 0.18 | +0.76; 46; 5; 0.07 | +1.23; 54; 6; 0.40 |
45
+ | close/both/18 | +1.98; 57; 8; 0.33 | +1.69; 58; 8; 0.44 | +1.74; 57; 8; 0.61 | +1.73; 55; 8; 0.58 | +1.84; 56; 8; 0.55 | +1.81; 59; 8; 0.03 | +1.72; 58; 8; 0.49 | +0.43; 48; 6; 0.21 | +1.54; 56; 7; 0.16 |
46
+
47
+ Readings are named form/kind/block: 'close' = the phrase closed by a full stop, read at the stop; 'stream' = the block's output; 'hub' = the blackboard; 'both' = the mean of the hub's and the stream's slider values (descriptive, never a candidate).
48
+
49
+ ## The decision (registered before the read)
50
+
51
+ Eligible readings: held-out signs at least 7 of 8 and leave-one-out at least 50 of 60. The largest separation wins; within .05, the better balance, then the higher CKA with the adapter output. One side is called better when its best reading leads by more than .05 in separation or .10 in balance; otherwise tied.
52
+
53
+ **STREAM BETTER**: the slider feature is close/stream/18 (separation +2.19, balance 0.41). The best hub reading is close/hub/22 (separation +1.83, balance 0.13): the hub is better balanced, the stream separates the moods further.
54
+
55
+ ## The anchor
56
+
57
+ The full run ended at its last check, after every condition was saved: the stored slider features file could not be found offline. The small run (same code, same trunks, 256 captions) reproduced the stored features to 6.0e-06 (trained), 9.2e-07 (random) of the largest stored value.
58
+
59
+ ## Files
60
+
61
+ | file | what |
62
+ |---|---|
63
+ | hub_read_step245674.json | the ledger: every condition's part A and part B, the floor, the rulers, the decision |
64
+ | hub_read_step245674.md | every table |
65
+ | hub_read_step245674_smoke.json | the small run (256 captions; its numbers are a rehearsal, except the anchor) |
66
+
67
+ ## Code
68
+
69
+ github.com/AbstractEyes/alephllm-diffusion-experiments, `src/alephllm_diffusion/hubs/read.py`: the measurement at commit e9828dd; the decision and tables written from the saved ledger with commit 68f8374 (branch feat/hub-read).
mini-beatrix-3/reports/v3/hub_read/hub_read_step245674.json ADDED
The diff for this file is too large to render. See raw diff
 
mini-beatrix-3/reports/v3/hub_read/hub_read_step245674.md ADDED
@@ -0,0 +1,206 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # The hub read at step 245,674 (2048 captions; made 2026-10-07 13:35:04 Pacific Daylight Time)
2
+
3
+ ## Part A: the earlier probe's numbers against CLIP-B/32, the bare trunk (top-10 overlap / Spearman)
4
+
5
+ | block | blackboard | stream pooled | stream last byte |
6
+ |---|---|---|---|
7
+ | 8 | 0.293 / +0.239 | 0.250 / +0.141 | 0.087 / +0.018 |
8
+ | 12 | 0.373 / +0.414 | 0.276 / +0.216 | 0.100 / +0.028 |
9
+ | 14 | 0.414 / +0.476 | 0.320 / +0.335 | 0.118 / +0.074 |
10
+ | 16 | 0.338 / +0.336 | 0.302 / +0.248 | 0.107 / +0.031 |
11
+ | 18 | 0.280 / +0.285 | 0.279 / +0.232 | 0.109 / +0.044 |
12
+ | 20 | 0.287 / +0.312 | 0.256 / +0.228 | 0.109 / +0.054 |
13
+ | 21 | 0.295 / +0.323 | 0.244 / +0.258 | 0.115 / +0.068 |
14
+ | 22 | 0.335 / +0.367 | 0.250 / +0.288 | 0.112 / +0.064 |
15
+ | 24 | 0.383 / +0.418 | 0.271 / +0.330 | 0.110 / +0.062 |
16
+ | 28 | 0.303 / +0.312 | 0.273 / +0.332 | 0.096 / +0.059 |
17
+ | 32 | - | 0.263 / +0.212 | 0.088 / +0.053 |
18
+
19
+ Gaussian floor 0.005 / -0.005. The earlier 2s model (step 52,000): blackboard block 14 .328 / +.407, block 9 .282 / +.257, block 19 .261 / +.268, pooled .249 / +.219, last token .088 / +.012.
20
+
21
+ ## Part A: debiased CKA with Anima's adapter output (the context the sliders push); forward ridge R2 in brackets
22
+
23
+ | block | blackboard bare | stream pooled bare | stream last bare | blackboard nine arms (A / B) | blackboard untrained |
24
+ |---|---|---|---|---|---|
25
+ | 8 | 0.657 (+0.603) | 0.595 (+0.555) | 0.311 (+0.314) | 0.650 (+0.599) / 0.647 (+0.599) | 0.391 (+0.519) |
26
+ | 12 | 0.813 (+0.625) | 0.729 (+0.551) | 0.344 (+0.341) | 0.809 (+0.621) / 0.810 (+0.623) | 0.337 (+0.499) |
27
+ | 14 | 0.827 (+0.635) | 0.803 (+0.563) | 0.428 (+0.392) | 0.829 (+0.633) / 0.827 (+0.635) | 0.342 (+0.492) |
28
+ | 16 | 0.786 (+0.635) | 0.798 (+0.570) | 0.372 (+0.379) | 0.770 (+0.632) / 0.772 (+0.632) | 0.317 (+0.481) |
29
+ | 18 | 0.760 (+0.616) | 0.783 (+0.572) | 0.427 (+0.395) | 0.740 (+0.614) / 0.742 (+0.615) | 0.309 (+0.473) |
30
+ | 20 | 0.762 (+0.614) | 0.764 (+0.567) | 0.473 (+0.392) | 0.741 (+0.613) / 0.753 (+0.613) | 0.308 (+0.469) |
31
+ | 21 | 0.773 (+0.615) | 0.736 (+0.557) | 0.541 (+0.418) | 0.770 (+0.615) / 0.774 (+0.616) | 0.314 (+0.465) |
32
+ | 22 | 0.814 (+0.636) | 0.753 (+0.556) | 0.523 (+0.426) | 0.812 (+0.634) / 0.815 (+0.635) | 0.308 (+0.462) |
33
+ | 24 | 0.825 (+0.640) | 0.790 (+0.554) | 0.496 (+0.428) | 0.820 (+0.637) / 0.822 (+0.638) | 0.307 (+0.458) |
34
+ | 28 | 0.708 (+0.598) | 0.804 (+0.550) | 0.382 (+0.390) | 0.727 (+0.593) / 0.721 (+0.592) | 0.283 (+0.448) |
35
+ | 32 | - | 0.765 (+0.536) | 0.344 (+0.349) | - / - | - |
36
+
37
+ ## Part A: debiased CKA with T5-XXL; forward ridge R2 in brackets
38
+
39
+ | block | blackboard bare | stream pooled bare | stream last bare | blackboard nine arms (A / B) | blackboard untrained |
40
+ |---|---|---|---|---|---|
41
+ | 8 | 0.519 (+0.400) | 0.492 (+0.366) | 0.230 (+0.205) | 0.521 (+0.397) / 0.519 (+0.396) | 0.331 (+0.325) |
42
+ | 12 | 0.614 (+0.421) | 0.586 (+0.379) | 0.255 (+0.219) | 0.614 (+0.419) / 0.614 (+0.420) | 0.285 (+0.305) |
43
+ | 14 | 0.613 (+0.427) | 0.619 (+0.386) | 0.309 (+0.248) | 0.615 (+0.425) / 0.612 (+0.426) | 0.290 (+0.307) |
44
+ | 16 | 0.629 (+0.426) | 0.633 (+0.393) | 0.285 (+0.246) | 0.619 (+0.422) / 0.620 (+0.422) | 0.268 (+0.291) |
45
+ | 18 | 0.648 (+0.422) | 0.659 (+0.398) | 0.349 (+0.271) | 0.624 (+0.417) / 0.627 (+0.418) | 0.268 (+0.301) |
46
+ | 20 | 0.649 (+0.421) | 0.652 (+0.398) | 0.412 (+0.281) | 0.629 (+0.416) / 0.638 (+0.416) | 0.269 (+0.299) |
47
+ | 21 | 0.667 (+0.424) | 0.645 (+0.392) | 0.479 (+0.300) | 0.654 (+0.420) / 0.656 (+0.421) | 0.267 (+0.296) |
48
+ | 22 | 0.682 (+0.437) | 0.654 (+0.392) | 0.463 (+0.304) | 0.672 (+0.433) / 0.674 (+0.434) | 0.265 (+0.299) |
49
+ | 24 | 0.653 (+0.438) | 0.658 (+0.390) | 0.431 (+0.304) | 0.652 (+0.434) / 0.652 (+0.435) | 0.263 (+0.296) |
50
+ | 28 | 0.572 (+0.410) | 0.651 (+0.387) | 0.313 (+0.282) | 0.593 (+0.407) / 0.590 (+0.407) | 0.246 (+0.287) |
51
+ | 32 | - | 0.615 (+0.371) | 0.281 (+0.262) | - / - | - |
52
+
53
+ ## Part A: debiased CKA with Qwen3 0.6B; forward ridge R2 in brackets
54
+
55
+ | block | blackboard bare | stream pooled bare | stream last bare | blackboard nine arms (A / B) | blackboard untrained |
56
+ |---|---|---|---|---|---|
57
+ | 8 | 0.556 (+0.632) | 0.541 (+0.592) | 0.160 (+0.315) | 0.557 (+0.629) / 0.554 (+0.629) | 0.363 (+0.567) |
58
+ | 12 | 0.690 (+0.652) | 0.672 (+0.619) | 0.196 (+0.328) | 0.692 (+0.649) / 0.683 (+0.647) | 0.320 (+0.558) |
59
+ | 14 | 0.698 (+0.674) | 0.733 (+0.635) | 0.261 (+0.381) | 0.698 (+0.671) / 0.689 (+0.669) | 0.335 (+0.553) |
60
+ | 16 | 0.709 (+0.675) | 0.733 (+0.645) | 0.230 (+0.376) | 0.695 (+0.669) / 0.696 (+0.669) | 0.315 (+0.548) |
61
+ | 18 | 0.722 (+0.679) | 0.764 (+0.656) | 0.300 (+0.416) | 0.702 (+0.671) / 0.702 (+0.671) | 0.312 (+0.544) |
62
+ | 20 | 0.730 (+0.685) | 0.763 (+0.663) | 0.363 (+0.433) | 0.705 (+0.678) / 0.717 (+0.677) | 0.308 (+0.539) |
63
+ | 21 | 0.764 (+0.701) | 0.767 (+0.668) | 0.446 (+0.476) | 0.745 (+0.694) / 0.747 (+0.694) | 0.323 (+0.540) |
64
+ | 22 | 0.782 (+0.713) | 0.792 (+0.672) | 0.412 (+0.483) | 0.765 (+0.706) / 0.765 (+0.707) | 0.314 (+0.536) |
65
+ | 24 | 0.757 (+0.716) | 0.818 (+0.678) | 0.372 (+0.483) | 0.756 (+0.709) / 0.755 (+0.710) | 0.315 (+0.535) |
66
+ | 28 | 0.679 (+0.698) | 0.822 (+0.679) | 0.241 (+0.451) | 0.703 (+0.689) / 0.700 (+0.691) | 0.289 (+0.529) |
67
+ | 32 | - | 0.765 (+0.659) | 0.205 (+0.415) | - / - | - |
68
+
69
+ The noise bar (draw 1 against draw 2, every tap and reference): CKA 0.026 mean, 0.151 at most.
70
+
71
+ ## Part B: the slider read, the bare trunk (the text alone; the text closed by a full stop)
72
+
73
+ | feature | leave-one-out signs | held-out signs | held-out nearest | separation | sd up / down | balance | probe accuracy |
74
+ |---|---|---|---|---|---|---|---|
75
+ | last/stream/8 | 43/60 | 5/8 | 6/10 | +0.72 | 0.43 / 1.06 | 0.90 | 0.67 |
76
+ | last/stream/12 | 47/60 | 6/8 | 7/10 | +1.03 | 0.35 / 0.82 | 0.85 | 0.87 |
77
+ | last/stream/14 | 47/60 | 5/8 | 7/10 | +1.26 | 0.33 / 0.74 | 0.82 | 0.97 |
78
+ | last/stream/16 | 55/60 | 8/8 | 10/10 | +1.58 | 0.27 / 0.55 | 0.71 | 0.98 |
79
+ | last/stream/18 | 57/60 | 8/8 | 10/10 | +1.90 | 0.24 / 0.50 | 0.72 | 0.97 |
80
+ | last/stream/20 | 58/60 | 8/8 | 10/10 | +1.83 | 0.28 / 0.44 | 0.47 | 0.98 |
81
+ | last/stream/21 | 57/60 | 8/8 | 10/10 | +1.83 | 0.31 / 0.45 | 0.36 | 0.98 |
82
+ | last/stream/22 | 56/60 | 8/8 | 10/10 | +1.82 | 0.34 / 0.46 | 0.29 | 0.97 |
83
+ | last/stream/24 | 52/60 | 8/8 | 10/10 | +1.84 | 0.36 / 0.51 | 0.35 | 0.97 |
84
+ | last/stream/28 | 52/60 | 8/8 | 10/10 | +1.60 | 0.41 / 0.55 | 0.30 | 0.93 |
85
+ | last/stream/rec | 57/60 | 8/8 | 10/10 | +1.79 | 0.27 / 0.48 | 0.57 | 0.98 |
86
+ | last/hub/8 | 48/60 | 4/8 | 5/10 | -0.04 | 0.54 / 0.55 | 0.02 | 0.70 |
87
+ | last/hub/12 | 56/60 | 8/8 | 10/10 | +1.41 | 0.32 / 0.45 | 0.35 | 0.92 |
88
+ | last/hub/14 | 57/60 | 8/8 | 10/10 | +1.57 | 0.36 / 0.50 | 0.33 | 0.93 |
89
+ | last/hub/16 | 57/60 | 8/8 | 10/10 | +1.23 | 0.28 / 0.42 | 0.39 | 0.93 |
90
+ | last/hub/18 | 57/60 | 8/8 | 10/10 | +1.61 | 0.35 / 0.46 | 0.26 | 0.93 |
91
+ | last/hub/20 | 52/60 | 8/8 | 10/10 | +1.38 | 0.44 / 0.46 | 0.04 | 0.85 |
92
+ | last/hub/21 | 57/60 | 8/8 | 10/10 | +1.50 | 0.39 / 0.44 | 0.11 | 0.93 |
93
+ | last/hub/22 | 54/60 | 8/8 | 10/10 | +1.67 | 0.34 / 0.45 | 0.28 | 0.95 |
94
+ | last/hub/24 | 54/60 | 7/8 | 9/10 | +1.47 | 0.29 / 0.54 | 0.62 | 0.87 |
95
+ | last/hub/28 | 51/60 | 6/8 | 7/10 | +0.59 | 0.39 / 0.46 | 0.18 | 0.70 |
96
+ | last/hub/rec | 56/60 | 8/8 | 10/10 | +1.47 | 0.31 / 0.44 | 0.34 | 0.95 |
97
+ | last/both/8 | 44/60 | 5/8 | - | +0.34 | 0.38 / 0.68 | 0.59 | - |
98
+ | last/both/12 | 51/60 | 8/8 | - | +1.22 | 0.28 / 0.53 | 0.64 | - |
99
+ | last/both/14 | 55/60 | 8/8 | - | +1.41 | 0.25 / 0.45 | 0.58 | - |
100
+ | last/both/16 | 58/60 | 8/8 | - | +1.40 | 0.23 / 0.40 | 0.56 | - |
101
+ | last/both/18 | 58/60 | 8/8 | - | +1.76 | 0.26 / 0.38 | 0.37 | - |
102
+ | last/both/20 | 57/60 | 8/8 | - | +1.60 | 0.31 / 0.33 | 0.09 | - |
103
+ | last/both/21 | 57/60 | 8/8 | - | +1.67 | 0.32 / 0.34 | 0.07 | - |
104
+ | last/both/22 | 56/60 | 8/8 | - | +1.74 | 0.29 / 0.35 | 0.17 | - |
105
+ | last/both/24 | 56/60 | 8/8 | - | +1.65 | 0.27 / 0.43 | 0.46 | - |
106
+ | last/both/28 | 57/60 | 8/8 | - | +1.10 | 0.32 / 0.41 | 0.25 | - |
107
+ | last/both/rec | 58/60 | 8/8 | - | +1.63 | 0.26 / 0.36 | 0.35 | - |
108
+ | close/stream/8 | 48/60 | 6/8 | 7/10 | +0.72 | 0.48 / 0.57 | 0.17 | - |
109
+ | close/stream/12 | 54/60 | 8/8 | 10/10 | +1.29 | 0.38 / 0.37 | 0.02 | - |
110
+ | close/stream/14 | 57/60 | 8/8 | 10/10 | +1.76 | 0.48 / 0.33 | 0.38 | - |
111
+ | close/stream/16 | 56/60 | 8/8 | 10/10 | +2.04 | 0.43 / 0.34 | 0.23 | - |
112
+ | close/stream/18 | 56/60 | 8/8 | 10/10 | +2.19 | 0.54 / 0.36 | 0.41 | - |
113
+ | close/stream/20 | 57/60 | 8/8 | 10/10 | +2.16 | 0.63 / 0.35 | 0.58 | - |
114
+ | close/stream/21 | 57/60 | 8/8 | 10/10 | +2.12 | 0.65 / 0.38 | 0.53 | - |
115
+ | close/stream/22 | 55/60 | 8/8 | 10/10 | +2.05 | 0.64 / 0.41 | 0.46 | - |
116
+ | close/stream/24 | 54/60 | 8/8 | 10/10 | +2.05 | 0.65 / 0.45 | 0.37 | - |
117
+ | close/stream/28 | 45/60 | 8/8 | 10/10 | +1.89 | 0.73 / 0.61 | 0.17 | - |
118
+ | close/stream/rec | 56/60 | 8/8 | 10/10 | +2.11 | 0.55 / 0.35 | 0.46 | - |
119
+ | close/hub/8 | 47/60 | 4/8 | 5/10 | -0.10 | 0.53 / 0.53 | 0.01 | - |
120
+ | close/hub/12 | 56/60 | 8/8 | 10/10 | +1.31 | 0.32 / 0.42 | 0.28 | - |
121
+ | close/hub/14 | 57/60 | 8/8 | 10/10 | +1.55 | 0.36 / 0.46 | 0.25 | - |
122
+ | close/hub/16 | 57/60 | 8/8 | 10/10 | +1.25 | 0.30 / 0.39 | 0.24 | - |
123
+ | close/hub/18 | 58/60 | 8/8 | 10/10 | +1.77 | 0.35 / 0.39 | 0.10 | - |
124
+ | close/hub/20 | 54/60 | 8/8 | 10/10 | +1.58 | 0.43 / 0.40 | 0.09 | - |
125
+ | close/hub/21 | 58/60 | 8/8 | 10/10 | +1.71 | 0.39 / 0.35 | 0.10 | - |
126
+ | close/hub/22 | 55/60 | 8/8 | 10/10 | +1.83 | 0.33 / 0.38 | 0.13 | - |
127
+ | close/hub/24 | 54/60 | 8/8 | 10/10 | +1.56 | 0.29 / 0.50 | 0.55 | - |
128
+ | close/hub/28 | 54/60 | 6/8 | 8/10 | +0.82 | 0.34 / 0.45 | 0.27 | - |
129
+ | close/hub/rec | 57/60 | 8/8 | 10/10 | +1.61 | 0.32 / 0.37 | 0.16 | - |
130
+ | close/both/8 | 49/60 | 7/8 | - | +0.31 | 0.39 / 0.45 | 0.14 | - |
131
+ | close/both/12 | 57/60 | 8/8 | - | +1.30 | 0.32 / 0.34 | 0.06 | - |
132
+ | close/both/14 | 59/60 | 8/8 | - | +1.65 | 0.36 / 0.30 | 0.20 | - |
133
+ | close/both/16 | 59/60 | 8/8 | - | +1.65 | 0.33 / 0.32 | 0.02 | - |
134
+ | close/both/18 | 57/60 | 8/8 | - | +1.98 | 0.42 / 0.30 | 0.33 | - |
135
+ | close/both/20 | 57/60 | 8/8 | - | +1.87 | 0.48 / 0.32 | 0.40 | - |
136
+ | close/both/21 | 57/60 | 8/8 | - | +1.92 | 0.49 / 0.29 | 0.53 | - |
137
+ | close/both/22 | 56/60 | 8/8 | - | +1.94 | 0.44 / 0.31 | 0.36 | - |
138
+ | close/both/24 | 55/60 | 8/8 | - | +1.81 | 0.40 / 0.35 | 0.13 | - |
139
+ | close/both/28 | 55/60 | 7/8 | - | +1.35 | 0.47 / 0.41 | 0.15 | - |
140
+ | close/both/rec | 57/60 | 8/8 | - | +1.86 | 0.40 / 0.29 | 0.34 | - |
141
+
142
+ ('both' = the mean of the hub's and the stream's slider values at that block: descriptive, never a candidate.)
143
+
144
+ ## Part B: the caption frame (descriptive; the bare trunk)
145
+
146
+ | feature | leave-one-out signs | held-out signs | held-out nearest | separation | sd up / down | balance | probe accuracy |
147
+ |---|---|---|---|---|---|---|---|
148
+ | frame/stream/8 | 45/60 | 6/8 | 7/10 | +0.75 | 0.61 / 0.66 | 0.08 | - |
149
+ | frame/stream/12 | 55/60 | 7/8 | 9/10 | +1.29 | 0.39 / 0.40 | 0.04 | - |
150
+ | frame/stream/14 | 56/60 | 8/8 | 10/10 | +1.65 | 0.39 / 0.38 | 0.02 | - |
151
+ | frame/stream/16 | 57/60 | 8/8 | 9/10 | +1.91 | 0.31 / 0.42 | 0.31 | - |
152
+ | frame/stream/18 | 57/60 | 8/8 | 10/10 | +2.01 | 0.30 / 0.52 | 0.55 | - |
153
+ | frame/stream/20 | 57/60 | 8/8 | 10/10 | +1.87 | 0.33 / 0.47 | 0.35 | - |
154
+ | frame/stream/21 | 57/60 | 8/8 | 10/10 | +1.84 | 0.34 / 0.49 | 0.36 | - |
155
+ | frame/stream/22 | 57/60 | 8/8 | 10/10 | +1.75 | 0.38 / 0.48 | 0.24 | - |
156
+ | frame/stream/24 | 57/60 | 8/8 | 10/10 | +1.76 | 0.36 / 0.50 | 0.34 | - |
157
+ | frame/stream/28 | 56/60 | 8/8 | 10/10 | +1.65 | 0.36 / 0.49 | 0.30 | - |
158
+ | frame/stream/rec | 57/60 | 8/8 | 10/10 | +1.89 | 0.31 / 0.47 | 0.43 | - |
159
+ | frame/hub/8 | 45/60 | 2/8 | 4/10 | -0.39 | 0.96 / 1.02 | 0.06 | - |
160
+ | frame/hub/12 | 55/60 | 8/8 | 10/10 | +1.25 | 0.45 / 0.55 | 0.20 | - |
161
+ | frame/hub/14 | 54/60 | 7/8 | 9/10 | +1.72 | 0.45 / 0.67 | 0.39 | - |
162
+ | frame/hub/16 | 56/60 | 7/8 | 9/10 | +1.26 | 0.45 / 0.52 | 0.15 | - |
163
+ | frame/hub/18 | 57/60 | 8/8 | 10/10 | +2.04 | 0.41 / 0.57 | 0.33 | - |
164
+ | frame/hub/20 | 56/60 | 6/8 | 8/10 | +1.50 | 0.47 / 0.50 | 0.06 | - |
165
+ | frame/hub/21 | 57/60 | 8/8 | 10/10 | +2.08 | 0.45 / 0.54 | 0.20 | - |
166
+ | frame/hub/22 | 57/60 | 8/8 | 10/10 | +2.18 | 0.36 / 0.59 | 0.48 | - |
167
+ | frame/hub/24 | 55/60 | 7/8 | 9/10 | +1.74 | 0.59 / 0.60 | 0.01 | - |
168
+ | frame/hub/28 | 48/60 | 6/8 | 6/10 | +1.38 | 0.64 / 0.70 | 0.09 | - |
169
+ | frame/hub/rec | 57/60 | 8/8 | 10/10 | +1.85 | 0.43 / 0.52 | 0.20 | - |
170
+ | frame/both/8 | 49/60 | 5/8 | - | +0.18 | 0.52 / 0.63 | 0.19 | - |
171
+ | frame/both/12 | 56/60 | 8/8 | - | +1.27 | 0.37 / 0.41 | 0.09 | - |
172
+ | frame/both/14 | 58/60 | 7/8 | - | +1.69 | 0.32 / 0.43 | 0.27 | - |
173
+ | frame/both/16 | 58/60 | 8/8 | - | +1.59 | 0.31 / 0.40 | 0.26 | - |
174
+ | frame/both/18 | 57/60 | 8/8 | - | +2.02 | 0.28 / 0.49 | 0.55 | - |
175
+ | frame/both/20 | 56/60 | 8/8 | - | +1.69 | 0.30 / 0.46 | 0.42 | - |
176
+ | frame/both/21 | 57/60 | 8/8 | - | +1.96 | 0.32 / 0.49 | 0.42 | - |
177
+ | frame/both/22 | 57/60 | 8/8 | - | +1.96 | 0.29 / 0.51 | 0.57 | - |
178
+ | frame/both/24 | 57/60 | 8/8 | - | +1.75 | 0.33 / 0.50 | 0.42 | - |
179
+ | frame/both/28 | 54/60 | 7/8 | - | +1.51 | 0.38 / 0.49 | 0.27 | - |
180
+ | frame/both/rec | 57/60 | 8/8 | - | +1.87 | 0.29 / 0.46 | 0.47 | - |
181
+
182
+ ## Part B: the decided features under every condition (separation; leave-one-out signs; held-out signs; balance)
183
+
184
+ | feature | M0 | M9_A | M9_B | M8_A | M8_B | Mc_A | Mc_B | U0 | A212 |
185
+ |---|---|---|---|---|---|---|---|---|---|
186
+ | close/stream/18 | +2.19; 56; 8/8; 0.41 | +1.82; 52; 8/8; 0.43 | +1.94; 53; 8/8; 0.51 | +1.87; 52; 8/8; 0.64 | +2.02; 53; 7/8; 0.72 | +1.93; 58; 8/8; 0.14 | +1.86; 56; 8/8; 0.36 | +0.15; 42; 6/8; 0.29 | +1.64; 54; 7/8; 0.10 |
187
+ | close/hub/22 | +1.83; 55; 8/8; 0.13 | +1.53; 58; 7/8; 0.05 | +1.67; 55; 7/8; 0.00 | +1.67; 56; 7/8; 0.06 | +1.80; 54; 8/8; 0.15 | +1.72; 58; 8/8; 0.29 | +1.67; 58; 8/8; 0.18 | +0.76; 46; 5/8; 0.07 | +1.23; 54; 6/8; 0.40 |
188
+ | close/both/22 | +1.94; 56; 8/8; 0.36 | +1.55; 53; 8/8; 0.25 | +1.76; 54; 8/8; 0.29 | +1.64; 51; 8/8; 0.41 | +1.84; 53; 8/8; 0.48 | +1.72; 58; 8/8; 0.04 | +1.72; 57; 8/8; 0.18 | +0.52; 48; 6/8; 0.24 | +1.46; 53; 7/8; 0.06 |
189
+
190
+ ## The decision
191
+
192
+ **STREAM BETTER**: the slider's feature = close/stream/18 (separation +2.19, balance 0.41)
193
+
194
+ - best hub: close/hub/22 (separation +1.83, balance 0.13, CKA with the adapter output 0.814)
195
+ - best stream: close/stream/18 (separation +2.19, balance 0.41, CKA with the adapter output 0.427)
196
+
197
+ ## The anchor
198
+
199
+ Not read in this run: the stored slider features file was not found in the offline hub cache (an online read earlier the same day moved the cache's main pointer to a newer snapshot without it); the code now searches every cached snapshot, and the decision and tables are saved before the anchor
200
+
201
+ The small run (same code, same trunks) against the stored slider features:
202
+
203
+ | trunk | largest difference | relative to the largest stored value |
204
+ |---|---|---|
205
+ | trained | 3.49e-05 | 6.00e-06 |
206
+ | random | 5.84e-06 | 9.18e-07 |
mini-beatrix-3/reports/v3/hub_read/hub_read_step245674_smoke.json ADDED
The diff for this file is too large to render. See raw diff