scrambled control seed 1, 100k steps (best val 1.5816)

scrambled/README.md

@@ -24,7 +24,7 @@ connectivity tables. FlyGPT's graph is extracted from it deterministically (`bui
 
 This checkpoint's wiring is a degree-preserving scramble of the MaleCNS wiring (control seed 1; edge survival 0.042).
 
-**Trained.** Condition `degree_preserving`, seed 1, step 19000, validation loss 1.7443 nats/char on the fixed Tiny Shakespeare split.
+**Trained.** Condition `degree_preserving`, seed 1, step 94000, validation loss 1.5816 nats/char on the fixed Tiny Shakespeare split.
 
 This is not a biological simulation of a living fly. The "weights" in the MaleCNS release are anatomical synapse
 counts; they are stored here as `graph.synapse_count` and are **not** the model's parameters.
@@ -159,17 +159,18 @@ differences Δ = loss(scrambled) − loss(real) have the same sign **and** the m
 
 | seed | real | degree-preserving scramble | Δ |
 |--:|--:|--:|--:|
-| 1 | 1.7275 | 1.7450 | +0.0175 |
-| 2 | 1.7285 | 1.7410 | +0.0125 |
-| 3 | 1.7490 | 1.7594 | +0.0104 |
-| 4 | 1.7516 | 1.7592 | +0.0076 |
-| 5 | 1.7305 | 1.7377 | +0.0073 |
-| **mean** | **1.7374** | **1.7485** | **+0.0111** |
-
-Validation loss in nats/char at the end of 20,000 steps, same data order, batches, adapter init and edge-value RNG
-stream per seed. Bigram reference on this split: 2.482. All five differences favour the real wiring, but the mean
-gap is below the pre-registered minimum effect, so the verdict is: **no detectable difference at this scale.** The fly connectome learns
-Shakespeare; whether its specific wiring helps, beyond its degree sequence, is not resolved at 5,000 neurons.
+| 1 | 1.6044 | 1.6125 | +0.0081 |
+| 2 | 1.6193 | 1.6299 | +0.0106 |
+| 3 | 1.6031 | 1.6064 | +0.0033 |
+| 4 | 1.6083 | 1.6065 | -0.0018 |
+| 5 | 1.6114 | 1.6073 | -0.0041 |
+| **mean** | **1.6093** | **1.6125** | **+0.0032** |
+
+Validation loss in nats/char at the end of 100,000 steps, same data order, batches, adapter init and edge-value RNG
+stream per seed. Bigram reference on this split: 2.482. The differences are small and not all of the same sign, so the
+verdict is: **no detectable difference at this scale.** The fly connectome learns Shakespeare; at 5,000 neurons its specific wiring does not
+measurably beat a degree-matched scramble. (At 20,000 steps the real wiring led on all five seeds by a mean of
+0.011 nats; with full training the scramble catches up, so that early edge is a learning-speed effect.)
 
 ## Citation
 

scrambled/config.json

@@ -20,6 +20,7 @@
   "degree_normalization": true,
   "init_scale": 1.0,
   "dtype": "bfloat16",
+  "training_steps": 100000,
   "graph": {
     "graph_name": "cb5k",
     "condition": "degree_preserving",
@@ -33,9 +34,9 @@
     "status": "trained",
     "condition": "degree_preserving",
     "seed": 1,
-    "step": 19000,
-    "val_loss": 1.7442553460597991,
-    "checkpoint": "checkpoints/flygpt-v0/cb5k/degree_preserving_seed1.pt",
-    "project": "flygpt-v0"
+    "step": 94000,
+    "val_loss": 1.5815802812576294,
+    "checkpoint": "checkpoints/flygpt-v0-100k/cb5k/degree_preserving_seed1.pt",
+    "project": "flygpt-v0-100k"
   }
 }

scrambled/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d803e1119f8d9e38be5fec284643c8b2bb6d27f5e883631c9e7159ab3a30bdb2
+oid sha256:dbb972ceb161cc7e3d43beed45c39096838dee8a1938b676c7c2abc32a27cd6c
 size 7495610