thefinalboss commited on
Commit
0d25546
·
verified ·
1 Parent(s): 7385a20

Upload docs/SURGERY_PROBE.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. docs/SURGERY_PROBE.md +64 -15
docs/SURGERY_PROBE.md CHANGED
@@ -1,14 +1,14 @@
1
  # Fractus Post-Surgery Checkpoint Probe
2
 
3
- **Date:** 2026-08-16 18:41 UTC
4
 
5
  ## Unified checkpoint
6
 
7
  - **File:** (4.66 GB)
8
- - **Method:** mean of 4 GPU checkpoints after surgery transforms + 150 steps
9
  - **Surgery:** gate_temp=2.5, omega_scale=4.0, lb_coef=0.02, Kuramoto gradients enabled
10
 
11
- ## Training state (live logs before force-save)
12
 
13
  | GPU | Tokens | CE loss | lb |
14
  |-----|--------|---------|-----|
@@ -17,43 +17,92 @@
17
  | 2 | 187,494,400 | 28.9 | 14.015 |
18
  | 3 | 184,473,600 | 35.8 | 14.027 |
19
 
20
- GPU1 had reached CE ~10.7 in the live surgery process before force-save.
21
 
22
- ## Generation test (unified merge)
23
 
24
- temperature=0.9, top-k=50, max 40 new tokens, tick-by-tick decode.
25
 
26
- **PROMPT:** The meaning of life is
27
 
28
 
29
 
30
- **PROMPT:** Hello, my name is
31
 
32
 
33
 
34
- **PROMPT:** In mathematics, a continuous system
35
 
 
36
 
37
 
38
- **PROMPT:** Fractus thinks
39
 
 
40
 
41
 
42
- **PROMPT:** Once upon a time
43
 
 
44
 
 
45
 
46
- **PROMPT:** The capital of France is
47
 
48
 
 
49
 
50
- **PROMPT:** What is consciousness?
51
 
52
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
53
 
54
  ## Interpretation
55
 
56
  - Unified post-surgery checkpoint **generates**.
57
  - Still **word-level repetition collapse** (not coherent sentences).
58
- - Stable lexical attractors: uphe, Colorado, Fate, Population, commands, Chosen, Towers.
59
- - Surgery reopened routing (LB active, omega diversified); language coherence still requires further digestion.
 
 
 
 
 
 
 
 
1
  # Fractus Post-Surgery Checkpoint Probe
2
 
3
+ **Date:** 2026-08-16 18:45 UTC
4
 
5
  ## Unified checkpoint
6
 
7
  - **File:** (4.66 GB)
8
+ - **Method:** mean of 4 GPU checkpoints after surgery transforms + 150 SGD steps/GPU
9
  - **Surgery:** gate_temp=2.5, omega_scale=4.0, lb_coef=0.02, Kuramoto gradients enabled
10
 
11
+ ## Training state (live surgery logs before force-save)
12
 
13
  | GPU | Tokens | CE loss | lb |
14
  |-----|--------|---------|-----|
 
17
  | 2 | 187,494,400 | 28.9 | 14.015 |
18
  | 3 | 184,473,600 | 35.8 | 14.027 |
19
 
20
+ GPU1 reached CE **10.7** in the live surgery process before force-save.
21
 
22
+ ## Generation results (full raw outputs)
23
 
24
+ Sampling: temperature=0.9, top-k=50, max 40 new tokens, tick-by-tick CTE decode on the unified merge.
25
 
26
+ ### Prompt
27
 
28
 
29
 
30
+ ### Model output
31
 
32
 
33
 
34
+ ---
35
 
36
+ ### Prompt
37
 
38
 
 
39
 
40
+ ### Model output
41
 
42
 
 
43
 
44
+ ---
45
 
46
+ ### Prompt
47
 
 
48
 
49
 
50
+ ### Model output
51
 
 
52
 
53
 
54
+ ---
55
+
56
+ ### Prompt
57
+
58
+
59
+
60
+ ### Model output
61
+
62
+
63
+
64
+ ---
65
+
66
+ ### Prompt
67
+
68
+
69
+
70
+ ### Model output
71
+
72
+
73
+
74
+ ---
75
+
76
+ ### Prompt
77
+
78
+
79
+
80
+ ### Model output
81
+
82
+
83
+
84
+ ---
85
+
86
+ ### Prompt
87
+
88
+
89
+
90
+ ### Model output
91
+
92
+
93
+
94
+ ---
95
 
96
  ## Interpretation
97
 
98
  - Unified post-surgery checkpoint **generates**.
99
  - Still **word-level repetition collapse** (not coherent sentences).
100
+ - Stable lexical attractors observed:
101
+ -
102
+ -
103
+ -
104
+ -
105
+ -
106
+ -
107
+ -
108
+ - Surgery reopened routing (LB active, omega diversified); coherent sentences still need further digestion under the repaired dynamics.