Upload docs/SURGERY_PROBE.md with huggingface_hub
Browse files- 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:
|
| 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
|
| 21 |
|
| 22 |
-
## Generation
|
| 23 |
|
| 24 |
-
temperature=0.9, top-k=50, max 40 new tokens, tick-by-tick decode.
|
| 25 |
|
| 26 |
-
|
| 27 |
|
| 28 |
|
| 29 |
|
| 30 |
-
|
| 31 |
|
| 32 |
|
| 33 |
|
| 34 |
-
|
| 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
|
| 59 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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.
|