Update README.md

README.md

@@ -140,15 +140,11 @@ Code and math reached lower loss than English prose, since they're highly struct
 
 Each component removed one at a time, same data order and hyperparameters. The `seed1` row is the baseline with different initial weights, giving a sense of run-to-run noise.
 
-<!-- TODO: paste the compare.py table here -->
-
-| run | mix | web | math | python | cpp | rust | german | russian | japanese |
-|---|---|---|---|---|---|---|---|---|---|
 | baseline | 4.949 | 4.996 | 4.447 | 4.224 | 4.507 | 3.942 | 6.335 | 6.319 | 7.002 |
-| seed1 | | | | | | | | | |
-| no AttnRes | | | | | | | | | |
-| no Engram | | | | | | | | | |
-| no mHC (1 stream) | | | | | | | | | |
+| seed1 | 4.994 | 5.003 | 4.595 | 4.380 | 4.691 | 4.053 | 6.338 | 6.246 | 6.936 |
+| no AttnRes | 4.990 | 5.035 | 4.510 | 4.285 | 4.538 | 4.000 | 6.346 | 6.356 | 6.958 |
+| no Engram | 5.133 | 5.162 | 4.681 | 4.512 | 4.822 | 4.202 | 6.430 | 6.372 | 7.021 |
+| no mHC (1 stream) | 4.989 | 5.024 | 4.525 | 4.322 | 4.614 | 4.026 | 6.316 | 6.309 | 6.939 |
 
 Caveat: with only 4 layers and 20M tokens, depth-routing components (mHC, AttnRes) and memorization components (Engram) have little room to show benefits. A null result here means "not demonstrated at this scale", not "useless".