AbstractPhil commited on
Commit
a4b2d77
·
verified ·
1 Parent(s): b5195fe

the repo card

Browse files
Files changed (1) hide show
  1. README.md +4 -2
README.md CHANGED
@@ -92,8 +92,10 @@ The readout files carry every number; the results files carry every block of eve
92
  |---|---|---|---|
93
  | qwen3/mse_gXA_o0 (seed A, the MSE form) | final, 4,000 steps | .955-.959 at blocks 8-15, .962-.971 at 16-27 (peak .971 at 22), .960 / .954 / .944 at 28-30, .895 at 31; before training .73 -> .32 | .995-.998 at most blocks (.993-.995 at blocks 1, 14, 16, 25, 30; .984 at 31); +.0003 bits per byte on web text, +.0005 on the worst partner's text |
94
  | qwen3/mse_nce_gXA_o0 (seed A, MSE + InfoNCE) | final, 4,000 steps | trails the MSE form at every block (its readout holds the numbers) | as the MSE form, slightly behind |
95
- | qwen3/mse_gXB_o1 (seed B), the shuffled-pairing control, the untrained-copy control, the solo | training | | |
96
- | t5, clip, bert (the family), qwen3/mse_gXA_o0_bytes | after the family's gate (both seeds, the shuffled control failing) | | |
 
 
97
 
98
  The bar for the gap was .85 at every served block from 12 on; the bar for silence .995 at every block beside +.012 bits per byte.
99
  The MSE form is the one carried forward.
 
92
  |---|---|---|---|
93
  | qwen3/mse_gXA_o0 (seed A, the MSE form) | final, 4,000 steps | .955-.959 at blocks 8-15, .962-.971 at 16-27 (peak .971 at 22), .960 / .954 / .944 at 28-30, .895 at 31; before training .73 -> .32 | .995-.998 at most blocks (.993-.995 at blocks 1, 14, 16, 25, 30; .984 at 31); +.0003 bits per byte on web text, +.0005 on the worst partner's text |
94
  | qwen3/mse_nce_gXA_o0 (seed A, MSE + InfoNCE) | final, 4,000 steps | trails the MSE form at every block (its readout holds the numbers) | as the MSE form, slightly behind |
95
+ | qwen3/mse_gXB_o1 (seed B, the MSE form, over the gXB eight) | final, 4,000 steps | .953-.964 at blocks 8-28, .944 / .915 at 30 / 31 (draw 2 within .01); the bar held on both seeds | .994-.997 at most blocks, .985 at 31 (a few middle blocks one to two thousandths under .995); +.0003 bits per byte on web text |
96
+ | qwen3/mse_gXA_o0_shuf (the shuffled-pairing control) | final, 4,000 steps | FAILS as a control must: below its own untrained reading at every block (.587 / .484 / .363 / .201 at blocks 8 / 12 / 20 / 31 against .733 / .640 / .486 / .322 before) | .997-.993, .969 at 31 |
97
+ | qwen3/mse_untrained_o0 (the untrained-copy control), qwen3/mse_solo_o0 (the solo), qwen3/mse_gXA_o0_lamA8 (lambda 8 on the plain rows) | training | | |
98
+ | t5, clip, bert (the family, seeds A and B), qwen3/mse_gXA_o0_bytes | queued on the pod after the controls (the gate passed on both seeds with the control failing) | | |
99
 
100
  The bar for the gap was .85 at every served block from 12 on; the bar for silence .995 at every block beside +.012 bits per byte.
101
  The MSE form is the one carried forward.