the repo card
Browse files
README.md
CHANGED
|
@@ -97,6 +97,7 @@ The readout files carry every number; the results files carry every block of eve
|
|
| 97 |
| qwen3/mse_untrained_o0 (the untrained-copy control: the same recipe on a random-initialization copy of the trunk, seed 0; mounts on such a copy only) | final, 2,000 steps | the random trunk reads the two surfaces alike before any arm (.92-.94); its arm reaches .985: the gauge saturates on an untrained trunk, so the trained trunk's closure is the arm's work | .995-.999 |
|
| 98 |
| qwen3/mse_solo_o0 (the solo: the same recipe with the arm alone on the bare trunk, no stage arms underneath) | final, 4,000 steps | 0.956 at block 31, 0.953-0.977 at blocks 8-28 (the arm over the eight: 0.892; 0.952-0.971); leads the arm over the eight at 28 of 32 blocks | lowest 0.9880 (the arm over the eight 0.9822); +0.0003 bits per byte on web text |
|
| 99 |
| qwen3/mse_gXA_o0_lamA8 (lambda 8 on the plain rows: the whitened-silence fallback, over the gXA eight) | final, 4,000 steps | 0.881 at block 31, 0.949-0.966 at blocks 8-28 (the seed-A arm over the eight at lambda 2: 0.892; mean over the served blocks 0.954 against 0.958) | lowest 0.9899 (block 31; at lambda 2: 0.9822); leads the lambda-2 arm's silence at 30 of 32 blocks; -0.0000 bits per byte on web text |
|
|
|
|
| 100 |
| qwen3/mse_solo_o1_bytes (Qwen3's spelling read at EVERY byte (the segmentation-free variant), seed B, alone on the trunk) | final, 4,000 steps | 0.963 at block 12, 0.947-0.988 at blocks 8-28 (before training 0.69 / 0.50 at blocks 12 / 31) | lowest 0.9955 (block 1); +0.0002 bits per byte on web text |
|
| 101 |
| bert/mse_solo_o1 (BERT's WordPiece spelling (lower case, '##' continuations), seed B, alone on the trunk) | final, 4,000 steps | 0.749 at block 31, 0.854-0.914 at blocks 8-28 (before training 0.87 / 0.67 at blocks 12 / 31) | lowest 0.9796 (block 31); +0.0007 bits per byte on web text |
|
| 102 |
| t5/mse_solo_o0 (T5's sentence-piece spelling, seed A, alone on the trunk) | final, 4,000 steps | 0.940 at block 31, 0.932-0.964 at blocks 8-28 (before training 0.65 / 0.36 at blocks 12 / 31) | lowest 0.9889 (block 31); +0.0001 bits per byte on web text |
|
|
|
|
| 97 |
| qwen3/mse_untrained_o0 (the untrained-copy control: the same recipe on a random-initialization copy of the trunk, seed 0; mounts on such a copy only) | final, 2,000 steps | the random trunk reads the two surfaces alike before any arm (.92-.94); its arm reaches .985: the gauge saturates on an untrained trunk, so the trained trunk's closure is the arm's work | .995-.999 |
|
| 98 |
| qwen3/mse_solo_o0 (the solo: the same recipe with the arm alone on the bare trunk, no stage arms underneath) | final, 4,000 steps | 0.956 at block 31, 0.953-0.977 at blocks 8-28 (the arm over the eight: 0.892; 0.952-0.971); leads the arm over the eight at 28 of 32 blocks | lowest 0.9880 (the arm over the eight 0.9822); +0.0003 bits per byte on web text |
|
| 99 |
| qwen3/mse_gXA_o0_lamA8 (lambda 8 on the plain rows: the whitened-silence fallback, over the gXA eight) | final, 4,000 steps | 0.881 at block 31, 0.949-0.966 at blocks 8-28 (the seed-A arm over the eight at lambda 2: 0.892; mean over the served blocks 0.954 against 0.958) | lowest 0.9899 (block 31; at lambda 2: 0.9822); leads the lambda-2 arm's silence at 30 of 32 blocks; -0.0000 bits per byte on web text |
|
| 100 |
+
| clip/mse_solo_o0_norm (CLIP's spelling (the byte map with word-end marks, lower case), read against A = the tokenizer-normalized text, seed A, alone on the trunk) | final, 4,000 steps | 0.786 at block 12, 0.676-0.920 at blocks 8-28 (before training 0.43 / 0.30 at blocks 12 / 31) | lowest 0.9814 (block 30); +0.0002 bits per byte on web text |
|
| 101 |
| qwen3/mse_solo_o1_bytes (Qwen3's spelling read at EVERY byte (the segmentation-free variant), seed B, alone on the trunk) | final, 4,000 steps | 0.963 at block 12, 0.947-0.988 at blocks 8-28 (before training 0.69 / 0.50 at blocks 12 / 31) | lowest 0.9955 (block 1); +0.0002 bits per byte on web text |
|
| 102 |
| bert/mse_solo_o1 (BERT's WordPiece spelling (lower case, '##' continuations), seed B, alone on the trunk) | final, 4,000 steps | 0.749 at block 31, 0.854-0.914 at blocks 8-28 (before training 0.87 / 0.67 at blocks 12 / 31) | lowest 0.9796 (block 31); +0.0007 bits per byte on web text |
|
| 103 |
| t5/mse_solo_o0 (T5's sentence-piece spelling, seed A, alone on the trunk) | final, 4,000 steps | 0.940 at block 31, 0.932-0.964 at blocks 8-28 (before training 0.65 / 0.36 at blocks 12 / 31) | lowest 0.9889 (block 31); +0.0001 bits per byte on web text |
|