Campaign B v1 PASSED both seeds: two-teacher alloy into mini-beatrix arm — meet +0.22/+0.21, preservation IMPROVED (B9 digit 0.837->0.971), LM +2pts, bit-exact off
Browse files
proto_frame/campaign_b/campaign_b_ledger.json
ADDED
|
@@ -0,0 +1,135 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_env": {
|
| 3 |
+
"transformers": "5.13.0",
|
| 4 |
+
"torch": "2.11.0+cu128",
|
| 5 |
+
"seed": 652978830,
|
| 6 |
+
"spec": "WIDE n32 K64 D4 h256",
|
| 7 |
+
"steps": 6000
|
| 8 |
+
},
|
| 9 |
+
"parity_meet": {
|
| 10 |
+
"bert": 0.296,
|
| 11 |
+
"t5b": 0.3208
|
| 12 |
+
},
|
| 13 |
+
"baselines_P1": {
|
| 14 |
+
"B4_last": {
|
| 15 |
+
"overall": 0.9991,
|
| 16 |
+
"digit": 0.988
|
| 17 |
+
},
|
| 18 |
+
"B9_last": {
|
| 19 |
+
"overall": 0.9874,
|
| 20 |
+
"digit": 0.837
|
| 21 |
+
}
|
| 22 |
+
},
|
| 23 |
+
"B-s0": {
|
| 24 |
+
"descent_gate": {
|
| 25 |
+
"passed": true,
|
| 26 |
+
"first": 0.801062,
|
| 27 |
+
"last": 0.735396,
|
| 28 |
+
"rel_drop": 0.082
|
| 29 |
+
},
|
| 30 |
+
"meet": {
|
| 31 |
+
"bert": 0.5151,
|
| 32 |
+
"t5b": 0.5353
|
| 33 |
+
},
|
| 34 |
+
"id_B4": {
|
| 35 |
+
"overall": 0.9994,
|
| 36 |
+
"digit": 0.9926
|
| 37 |
+
},
|
| 38 |
+
"id_B9": {
|
| 39 |
+
"overall": 0.9978,
|
| 40 |
+
"digit": 0.9713
|
| 41 |
+
},
|
| 42 |
+
"lm": {
|
| 43 |
+
"nextbyte_on": 0.5905,
|
| 44 |
+
"nextbyte_clean": 0.5679,
|
| 45 |
+
"delta": 0.0226,
|
| 46 |
+
"bitexact_off": true
|
| 47 |
+
},
|
| 48 |
+
"alloy_ledger": {
|
| 49 |
+
"mse_bert": {
|
| 50 |
+
"Name": {
|
| 51 |
+
"n": 12,
|
| 52 |
+
"mean": 0.979131
|
| 53 |
+
},
|
| 54 |
+
"digit": {
|
| 55 |
+
"n": 12,
|
| 56 |
+
"mean": 0.874757
|
| 57 |
+
},
|
| 58 |
+
"word": {
|
| 59 |
+
"n": 12,
|
| 60 |
+
"mean": 1.069207
|
| 61 |
+
}
|
| 62 |
+
},
|
| 63 |
+
"mse_t5b": {
|
| 64 |
+
"Name": {
|
| 65 |
+
"n": 12,
|
| 66 |
+
"mean": 0.854603
|
| 67 |
+
},
|
| 68 |
+
"digit": {
|
| 69 |
+
"n": 12,
|
| 70 |
+
"mean": 0.805826
|
| 71 |
+
},
|
| 72 |
+
"word": {
|
| 73 |
+
"n": 12,
|
| 74 |
+
"mean": 1.041047
|
| 75 |
+
}
|
| 76 |
+
}
|
| 77 |
+
}
|
| 78 |
+
},
|
| 79 |
+
"B-s1": {
|
| 80 |
+
"descent_gate": {
|
| 81 |
+
"passed": true,
|
| 82 |
+
"first": 0.801057,
|
| 83 |
+
"last": 0.735587,
|
| 84 |
+
"rel_drop": 0.0817
|
| 85 |
+
},
|
| 86 |
+
"meet": {
|
| 87 |
+
"bert": 0.5151,
|
| 88 |
+
"t5b": 0.5355
|
| 89 |
+
},
|
| 90 |
+
"id_B4": {
|
| 91 |
+
"overall": 0.9994,
|
| 92 |
+
"digit": 0.9926
|
| 93 |
+
},
|
| 94 |
+
"id_B9": {
|
| 95 |
+
"overall": 0.998,
|
| 96 |
+
"digit": 0.9741
|
| 97 |
+
},
|
| 98 |
+
"lm": {
|
| 99 |
+
"nextbyte_on": 0.5875,
|
| 100 |
+
"nextbyte_clean": 0.5679,
|
| 101 |
+
"delta": 0.0197,
|
| 102 |
+
"bitexact_off": true
|
| 103 |
+
},
|
| 104 |
+
"alloy_ledger": {
|
| 105 |
+
"mse_bert": {
|
| 106 |
+
"Name": {
|
| 107 |
+
"n": 12,
|
| 108 |
+
"mean": 0.992707
|
| 109 |
+
},
|
| 110 |
+
"digit": {
|
| 111 |
+
"n": 12,
|
| 112 |
+
"mean": 0.875689
|
| 113 |
+
},
|
| 114 |
+
"word": {
|
| 115 |
+
"n": 12,
|
| 116 |
+
"mean": 1.083938
|
| 117 |
+
}
|
| 118 |
+
},
|
| 119 |
+
"mse_t5b": {
|
| 120 |
+
"Name": {
|
| 121 |
+
"n": 12,
|
| 122 |
+
"mean": 0.855337
|
| 123 |
+
},
|
| 124 |
+
"digit": {
|
| 125 |
+
"n": 12,
|
| 126 |
+
"mean": 0.799736
|
| 127 |
+
},
|
| 128 |
+
"word": {
|
| 129 |
+
"n": 12,
|
| 130 |
+
"mean": 1.050094
|
| 131 |
+
}
|
| 132 |
+
}
|
| 133 |
+
}
|
| 134 |
+
}
|
| 135 |
+
}
|