verify aft_wave_v2/control_matched__agreement/results
Browse files
aft_wave_v2/control_matched__agreement/results/ARTIFACT_MANIFEST.json
ADDED
|
@@ -0,0 +1,447 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"repo": "arcadia-impact/scimt-dispatch-models",
|
| 3 |
+
"remote_prefix": "aft_wave_v2/control_matched__agreement/results",
|
| 4 |
+
"local_folder": "/workspace/wave/results",
|
| 5 |
+
"files": {
|
| 6 |
+
"ARTIFACT_MANIFEST.local.json": {
|
| 7 |
+
"size": 9926,
|
| 8 |
+
"sha256": "a52a2c40240383ca24a51b4a8e462c6faef2499bd6bf7205a3969de0e94a3616"
|
| 9 |
+
},
|
| 10 |
+
"charter_real_4x-baseline/ENDPOINT_DONE.json": {
|
| 11 |
+
"size": 54,
|
| 12 |
+
"sha256": "a5dcb78b1f1adae44ec7312493d22e7b173a89b0486ae7d4592ded8d5fe31508"
|
| 13 |
+
},
|
| 14 |
+
"charter_real_4x-baseline/eval_holdout_adjacent.jsonl": {
|
| 15 |
+
"size": 48444,
|
| 16 |
+
"sha256": "e2a9d5a508f8245dd75472bfa47503c15afe113af9488f06558fd5d4553d3b22"
|
| 17 |
+
},
|
| 18 |
+
"charter_real_4x-baseline/eval_holdout_agreement.jsonl": {
|
| 19 |
+
"size": 92867,
|
| 20 |
+
"sha256": "aad8d42e3428634a14e8d1c93855a723e2b4264c1593002438637d1b9abfefdf"
|
| 21 |
+
},
|
| 22 |
+
"charter_real_4x-baseline/eval_holdout_conflict.jsonl": {
|
| 23 |
+
"size": 92081,
|
| 24 |
+
"sha256": "9d1cc5f654f9cb69b75b1a41e590fd5584640499247f91c6bd94e72807b4409f"
|
| 25 |
+
},
|
| 26 |
+
"charter_real_4x-baseline/eval_trained_adjacent.jsonl": {
|
| 27 |
+
"size": 121156,
|
| 28 |
+
"sha256": "d5facd299c7bd2ac9eb9358b52ccc7a2c45141337f0fe1d85bc365e0f6e9f825"
|
| 29 |
+
},
|
| 30 |
+
"charter_real_4x-baseline/eval_trained_agreement.jsonl": {
|
| 31 |
+
"size": 232224,
|
| 32 |
+
"sha256": "ff462f141b3423b5f68d430c46e4eff12fb8a2b55bf8e38542aa93d65f2429be"
|
| 33 |
+
},
|
| 34 |
+
"charter_real_4x-baseline/eval_trained_conflict.jsonl": {
|
| 35 |
+
"size": 230213,
|
| 36 |
+
"sha256": "5f40ef0f95be19aaa883b34afc2550820ac041eaa0ea6a05c8f41b9198570c37"
|
| 37 |
+
},
|
| 38 |
+
"charter_real_4x-baseline/sanity.jsonl": {
|
| 39 |
+
"size": 6450,
|
| 40 |
+
"sha256": "90acbf15d330eab054bc6195daed5cd1c2960f52dd7b4e949037f6469e242dc9"
|
| 41 |
+
},
|
| 42 |
+
"charter_real_4x-baseline/sanity_prompts.jsonl": {
|
| 43 |
+
"size": 154858,
|
| 44 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 45 |
+
},
|
| 46 |
+
"charter_real_4x__agreement-step128/ENDPOINT_DONE.json": {
|
| 47 |
+
"size": 91,
|
| 48 |
+
"sha256": "3a249d3bb7937708e49c5ef0c8aa63daf9fe6802b733df3e820f23cd94560b5c"
|
| 49 |
+
},
|
| 50 |
+
"charter_real_4x__agreement-step128/eval_holdout_adjacent.jsonl": {
|
| 51 |
+
"size": 48461,
|
| 52 |
+
"sha256": "f5433fd2401a0067c094f20d5610131f1718b80594091a6241bca881800a7f6f"
|
| 53 |
+
},
|
| 54 |
+
"charter_real_4x__agreement-step128/eval_holdout_agreement.jsonl": {
|
| 55 |
+
"size": 92896,
|
| 56 |
+
"sha256": "e61c212893226726f580bed782aceb51474b27133facf44ade5632eeabd3339e"
|
| 57 |
+
},
|
| 58 |
+
"charter_real_4x__agreement-step128/eval_holdout_conflict.jsonl": {
|
| 59 |
+
"size": 92091,
|
| 60 |
+
"sha256": "bed8f71431b9ed73cf353c1f81e18be242a0ea0938612f0ba1addcb1e321833b"
|
| 61 |
+
},
|
| 62 |
+
"charter_real_4x__agreement-step128/eval_trained_adjacent.jsonl": {
|
| 63 |
+
"size": 121162,
|
| 64 |
+
"sha256": "125eb7eec80092bb6f1758d9b33f85aead4b5893627c2d219598ac6229e8ad7f"
|
| 65 |
+
},
|
| 66 |
+
"charter_real_4x__agreement-step128/eval_trained_agreement.jsonl": {
|
| 67 |
+
"size": 232293,
|
| 68 |
+
"sha256": "07b418e1a7690d67890cb64e0be77896b7bf899ef289fc597004de4d7a817507"
|
| 69 |
+
},
|
| 70 |
+
"charter_real_4x__agreement-step128/eval_trained_conflict.jsonl": {
|
| 71 |
+
"size": 230197,
|
| 72 |
+
"sha256": "00e8edcd5aee37502fe997c184597cbeea9aebcdf470f1ab1e9e47135913af7c"
|
| 73 |
+
},
|
| 74 |
+
"charter_real_4x__agreement-step128/sanity.jsonl": {
|
| 75 |
+
"size": 6452,
|
| 76 |
+
"sha256": "d39e1e3a8b85658ca822c7d5209de5b72e68c5cf27ca4092b019de344ea2280a"
|
| 77 |
+
},
|
| 78 |
+
"charter_real_4x__agreement-step128/sanity_prompts.jsonl": {
|
| 79 |
+
"size": 154858,
|
| 80 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 81 |
+
},
|
| 82 |
+
"charter_real_4x__agreement-step256/ENDPOINT_DONE.json": {
|
| 83 |
+
"size": 91,
|
| 84 |
+
"sha256": "3b3a0ccbba4cc213f930df0886f5570b16c0526e3f423bda8f68c344db3c3572"
|
| 85 |
+
},
|
| 86 |
+
"charter_real_4x__agreement-step256/eval_holdout_adjacent.jsonl": {
|
| 87 |
+
"size": 48477,
|
| 88 |
+
"sha256": "f8ef227ecdf4a1325f4915c958ae6a26491cd8c14a85729aba1e112ea155089b"
|
| 89 |
+
},
|
| 90 |
+
"charter_real_4x__agreement-step256/eval_holdout_agreement.jsonl": {
|
| 91 |
+
"size": 92908,
|
| 92 |
+
"sha256": "82831951eee88f34df4a463ac59bcdea16064c6345b89405887f82d78e2531f0"
|
| 93 |
+
},
|
| 94 |
+
"charter_real_4x__agreement-step256/eval_holdout_conflict.jsonl": {
|
| 95 |
+
"size": 92122,
|
| 96 |
+
"sha256": "3f5406824a9bbdde4633b8d65b6ccde3c2b2dc40af42b94676e874ded86f0444"
|
| 97 |
+
},
|
| 98 |
+
"charter_real_4x__agreement-step256/eval_trained_adjacent.jsonl": {
|
| 99 |
+
"size": 121187,
|
| 100 |
+
"sha256": "5025128060d2230a6ae8d9d133c9dd9fb5382fa980e27b8ccb52a424829a268a"
|
| 101 |
+
},
|
| 102 |
+
"charter_real_4x__agreement-step256/eval_trained_agreement.jsonl": {
|
| 103 |
+
"size": 232284,
|
| 104 |
+
"sha256": "e7077ae0f2ab8845a1082dfb2d24b076d7d0d310d7e8aae6486f5cf76976c741"
|
| 105 |
+
},
|
| 106 |
+
"charter_real_4x__agreement-step256/eval_trained_conflict.jsonl": {
|
| 107 |
+
"size": 230269,
|
| 108 |
+
"sha256": "338f72a7aa2f91a44bfc7d7abb2edb57020b4755612433fa4cd75bca0cb591b8"
|
| 109 |
+
},
|
| 110 |
+
"charter_real_4x__agreement-step256/sanity.jsonl": {
|
| 111 |
+
"size": 6452,
|
| 112 |
+
"sha256": "b541447e75b3c1bd57c45db3f1e5a7b1464c0d9c71a64a3d91e762ed0404351b"
|
| 113 |
+
},
|
| 114 |
+
"charter_real_4x__agreement-step256/sanity_prompts.jsonl": {
|
| 115 |
+
"size": 154858,
|
| 116 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 117 |
+
},
|
| 118 |
+
"charter_real_4x__agreement-step32/ENDPOINT_DONE.json": {
|
| 119 |
+
"size": 90,
|
| 120 |
+
"sha256": "bdb45cfaea70521f8a2424a826a5144e3ac4dbe9a7a52f449c6d6e452be970cd"
|
| 121 |
+
},
|
| 122 |
+
"charter_real_4x__agreement-step32/eval_holdout_adjacent.jsonl": {
|
| 123 |
+
"size": 48457,
|
| 124 |
+
"sha256": "a60ffa4e5a2b176bab2b826bacba74cdbe2bc89634f3c5865a9469488be8344c"
|
| 125 |
+
},
|
| 126 |
+
"charter_real_4x__agreement-step32/eval_holdout_agreement.jsonl": {
|
| 127 |
+
"size": 92905,
|
| 128 |
+
"sha256": "471a50f1e0657718cb2c4d018b678b535bd6036f68c07ffc82bdca688cbc6e1b"
|
| 129 |
+
},
|
| 130 |
+
"charter_real_4x__agreement-step32/eval_holdout_conflict.jsonl": {
|
| 131 |
+
"size": 92108,
|
| 132 |
+
"sha256": "f41170f8c63415b2d75cfffcadea4fd9f8c804a06cef863687f9614bfeafc184"
|
| 133 |
+
},
|
| 134 |
+
"charter_real_4x__agreement-step32/eval_trained_adjacent.jsonl": {
|
| 135 |
+
"size": 121167,
|
| 136 |
+
"sha256": "b7dc09338bec9132486ea64d2bc97fc1f4b48dcf32887bc7a8e40d39323808ef"
|
| 137 |
+
},
|
| 138 |
+
"charter_real_4x__agreement-step32/eval_trained_agreement.jsonl": {
|
| 139 |
+
"size": 232291,
|
| 140 |
+
"sha256": "0fecf79778395229f3a67d1b77c5cef2cb66bed6c264a986e5a55f9f442bb6b5"
|
| 141 |
+
},
|
| 142 |
+
"charter_real_4x__agreement-step32/eval_trained_conflict.jsonl": {
|
| 143 |
+
"size": 230169,
|
| 144 |
+
"sha256": "c35307d2dfd9f38f506fdbb0bac8486ed7ffc31d110365e63e0af3e8bb7728f3"
|
| 145 |
+
},
|
| 146 |
+
"charter_real_4x__agreement-step32/sanity.jsonl": {
|
| 147 |
+
"size": 6451,
|
| 148 |
+
"sha256": "ed175282dddf3b15f8cbe7240be52192a6c0cf34dfe6530d1443ac3e6a8c2111"
|
| 149 |
+
},
|
| 150 |
+
"charter_real_4x__agreement-step32/sanity_prompts.jsonl": {
|
| 151 |
+
"size": 154858,
|
| 152 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 153 |
+
},
|
| 154 |
+
"charter_real_4x__agreement-step512/ENDPOINT_DONE.json": {
|
| 155 |
+
"size": 91,
|
| 156 |
+
"sha256": "fcf3f7f279da667f942d00489cd166a956667f9f026531d8077483b31cf4cb7d"
|
| 157 |
+
},
|
| 158 |
+
"charter_real_4x__agreement-step512/eval_holdout_adjacent.jsonl": {
|
| 159 |
+
"size": 48460,
|
| 160 |
+
"sha256": "ef9574aef5c1ddc397281793a4cf5fb9fdc492f2e926606f4691e47288ff8b18"
|
| 161 |
+
},
|
| 162 |
+
"charter_real_4x__agreement-step512/eval_holdout_agreement.jsonl": {
|
| 163 |
+
"size": 92913,
|
| 164 |
+
"sha256": "49becb9121525e172d3b714b4ee60bea976322b17fcff20c1285b30a9d2460e5"
|
| 165 |
+
},
|
| 166 |
+
"charter_real_4x__agreement-step512/eval_holdout_conflict.jsonl": {
|
| 167 |
+
"size": 92094,
|
| 168 |
+
"sha256": "6acbf7443fbcef6610cbc5270e165a5adababf271d9c96617051d4b82e214131"
|
| 169 |
+
},
|
| 170 |
+
"charter_real_4x__agreement-step512/eval_trained_adjacent.jsonl": {
|
| 171 |
+
"size": 121163,
|
| 172 |
+
"sha256": "96e9f943a3f7aeb37c4ee9b575c76425f7eec4e57b8a93964ed868e64eec2468"
|
| 173 |
+
},
|
| 174 |
+
"charter_real_4x__agreement-step512/eval_trained_agreement.jsonl": {
|
| 175 |
+
"size": 232283,
|
| 176 |
+
"sha256": "18e9ac96eca7c2c016a12aab28d91000d5ee38b57e7a81c0ce2ef3d3818157d0"
|
| 177 |
+
},
|
| 178 |
+
"charter_real_4x__agreement-step512/eval_trained_conflict.jsonl": {
|
| 179 |
+
"size": 230238,
|
| 180 |
+
"sha256": "71191a629763543bb9c997d4b0fb534bc9f99757dc3fa47653a73e72580f4df7"
|
| 181 |
+
},
|
| 182 |
+
"charter_real_4x__agreement-step512/sanity.jsonl": {
|
| 183 |
+
"size": 6451,
|
| 184 |
+
"sha256": "8c42a7fa7d11c3aed80009dbe7b2460f125eb2ef9a4a325fa8783d93e3294e46"
|
| 185 |
+
},
|
| 186 |
+
"charter_real_4x__agreement-step512/sanity_prompts.jsonl": {
|
| 187 |
+
"size": 154858,
|
| 188 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 189 |
+
},
|
| 190 |
+
"charter_real_4x__agreement-step64/ENDPOINT_DONE.json": {
|
| 191 |
+
"size": 90,
|
| 192 |
+
"sha256": "9dae59d2f79b016f45df1dec86d571d4c923d01c9d0328a4be5af3a20de32041"
|
| 193 |
+
},
|
| 194 |
+
"charter_real_4x__agreement-step64/eval_holdout_adjacent.jsonl": {
|
| 195 |
+
"size": 48461,
|
| 196 |
+
"sha256": "8b3536018ffe11d4ff6e0a075f5eb03d7ea1570d388b8d4a0d13e6ff221f1b2f"
|
| 197 |
+
},
|
| 198 |
+
"charter_real_4x__agreement-step64/eval_holdout_agreement.jsonl": {
|
| 199 |
+
"size": 92893,
|
| 200 |
+
"sha256": "993b737506f6b8f0409b2d7add648b32a3dfebfaf990132335e480faf64ad70b"
|
| 201 |
+
},
|
| 202 |
+
"charter_real_4x__agreement-step64/eval_holdout_conflict.jsonl": {
|
| 203 |
+
"size": 92098,
|
| 204 |
+
"sha256": "d74a3af78b7b596b79a69fba920e735d07054c5a80975d0058229a594c3ab5fb"
|
| 205 |
+
},
|
| 206 |
+
"charter_real_4x__agreement-step64/eval_trained_adjacent.jsonl": {
|
| 207 |
+
"size": 121153,
|
| 208 |
+
"sha256": "3c66e8cd525db03cbc5890ce2e4d857556d0cbb5564e8b4df8eccfa96a0ef8bf"
|
| 209 |
+
},
|
| 210 |
+
"charter_real_4x__agreement-step64/eval_trained_agreement.jsonl": {
|
| 211 |
+
"size": 232282,
|
| 212 |
+
"sha256": "52141ab3cb1d73b46719cf55021cb9233a1087bd5b481bb8d093ed0353a80391"
|
| 213 |
+
},
|
| 214 |
+
"charter_real_4x__agreement-step64/eval_trained_conflict.jsonl": {
|
| 215 |
+
"size": 230207,
|
| 216 |
+
"sha256": "3282ca4123fea56b2480da59c3a9fa9d927e4518cf86fa2fee82659d656c8d64"
|
| 217 |
+
},
|
| 218 |
+
"charter_real_4x__agreement-step64/sanity.jsonl": {
|
| 219 |
+
"size": 6451,
|
| 220 |
+
"sha256": "289bba0b5c96491534507723e94e1c8ebd81ce3a6cf110b4174c791f591fc6b3"
|
| 221 |
+
},
|
| 222 |
+
"charter_real_4x__agreement-step64/sanity_prompts.jsonl": {
|
| 223 |
+
"size": 154858,
|
| 224 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 225 |
+
},
|
| 226 |
+
"control_matched-baseline/ENDPOINT_DONE.json": {
|
| 227 |
+
"size": 54,
|
| 228 |
+
"sha256": "564c4a3e89230c7c919fe70d978a9e42ec8a58c78bfa34c9141ce527dad86cc1"
|
| 229 |
+
},
|
| 230 |
+
"control_matched-baseline/eval_holdout_adjacent.jsonl": {
|
| 231 |
+
"size": 48445,
|
| 232 |
+
"sha256": "69ef1af71619c9e85a03441734e7dd10d0f23053f6e416c76261e0382d06577a"
|
| 233 |
+
},
|
| 234 |
+
"control_matched-baseline/eval_holdout_agreement.jsonl": {
|
| 235 |
+
"size": 92888,
|
| 236 |
+
"sha256": "491dbb3cfeefc8fa44f73b60d113b88d497999673f3d5ef731a03450027f318c"
|
| 237 |
+
},
|
| 238 |
+
"control_matched-baseline/eval_holdout_conflict.jsonl": {
|
| 239 |
+
"size": 92099,
|
| 240 |
+
"sha256": "6db98ca165851dbe23fd1cda6445f4dd9933c5060d8040f2b4557dbadc37c7ea"
|
| 241 |
+
},
|
| 242 |
+
"control_matched-baseline/eval_trained_adjacent.jsonl": {
|
| 243 |
+
"size": 121168,
|
| 244 |
+
"sha256": "a4f55209b4ba132641a7264394556b84ee1253e137232203111f5a0072ea0970"
|
| 245 |
+
},
|
| 246 |
+
"control_matched-baseline/eval_trained_agreement.jsonl": {
|
| 247 |
+
"size": 232257,
|
| 248 |
+
"sha256": "34637a5645c7f58aa866ca7a419de33c90f93fb98b1ebf2052eaf59cf37cd24e"
|
| 249 |
+
},
|
| 250 |
+
"control_matched-baseline/eval_trained_conflict.jsonl": {
|
| 251 |
+
"size": 230194,
|
| 252 |
+
"sha256": "05c88781d5def6a57afbee3ff41939b86cf69612278eeca54386c2bac3e2471b"
|
| 253 |
+
},
|
| 254 |
+
"control_matched-baseline/sanity.jsonl": {
|
| 255 |
+
"size": 6449,
|
| 256 |
+
"sha256": "22460c69db1727e60049de85103cc5c7826485c3911d9de77323460a1de008a3"
|
| 257 |
+
},
|
| 258 |
+
"control_matched-baseline/sanity_prompts.jsonl": {
|
| 259 |
+
"size": 154858,
|
| 260 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 261 |
+
},
|
| 262 |
+
"control_matched__agreement-step128/ENDPOINT_DONE.json": {
|
| 263 |
+
"size": 91,
|
| 264 |
+
"sha256": "494deed775c103ab71ffda75cc285b270b9d8672e0b65e86cce9387d0f6d2afc"
|
| 265 |
+
},
|
| 266 |
+
"control_matched__agreement-step128/eval_holdout_adjacent.jsonl": {
|
| 267 |
+
"size": 48453,
|
| 268 |
+
"sha256": "727762c4a6f8486ac2a4399d57ab5012d4a2f0283ad85bab13aebd37e6534af8"
|
| 269 |
+
},
|
| 270 |
+
"control_matched__agreement-step128/eval_holdout_agreement.jsonl": {
|
| 271 |
+
"size": 92911,
|
| 272 |
+
"sha256": "c79d5f4bd4f88ddba2a60420bd4d38f3b21d387aabccf556805cd182261ed35a"
|
| 273 |
+
},
|
| 274 |
+
"control_matched__agreement-step128/eval_holdout_conflict.jsonl": {
|
| 275 |
+
"size": 92093,
|
| 276 |
+
"sha256": "2070209c37be0ad35ab8feeece8eb5aeeff664074b64f4c456c04abfbd050fae"
|
| 277 |
+
},
|
| 278 |
+
"control_matched__agreement-step128/eval_trained_adjacent.jsonl": {
|
| 279 |
+
"size": 121171,
|
| 280 |
+
"sha256": "f070ea6894f6cccb2d744ebdaf12cc618bee9c953318460c5b254e8a89bf17ed"
|
| 281 |
+
},
|
| 282 |
+
"control_matched__agreement-step128/eval_trained_agreement.jsonl": {
|
| 283 |
+
"size": 232290,
|
| 284 |
+
"sha256": "80ca85bf9f7dd4af503633d8821ea5ea9ce09a1db5ba4dc02fda01f8e4a48f34"
|
| 285 |
+
},
|
| 286 |
+
"control_matched__agreement-step128/eval_trained_conflict.jsonl": {
|
| 287 |
+
"size": 230187,
|
| 288 |
+
"sha256": "e92e9b09707bfd4b5b8ee643c752a1f8be74cb6a4b5e6c6a9e27e19264e58042"
|
| 289 |
+
},
|
| 290 |
+
"control_matched__agreement-step128/sanity.jsonl": {
|
| 291 |
+
"size": 6451,
|
| 292 |
+
"sha256": "6f9ba8af215cbdc5de0278f950bda7184f98f833200388db36bcaf758f5901a8"
|
| 293 |
+
},
|
| 294 |
+
"control_matched__agreement-step128/sanity_prompts.jsonl": {
|
| 295 |
+
"size": 154858,
|
| 296 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 297 |
+
},
|
| 298 |
+
"control_matched__agreement-step256/ENDPOINT_DONE.json": {
|
| 299 |
+
"size": 91,
|
| 300 |
+
"sha256": "5f21f962d4c83df5d6ea6e3eee9803f8a4060c1b7355a260d2116f5ff864d63b"
|
| 301 |
+
},
|
| 302 |
+
"control_matched__agreement-step256/eval_holdout_adjacent.jsonl": {
|
| 303 |
+
"size": 48473,
|
| 304 |
+
"sha256": "2afdefc54b2ac5f06cd3a8449e797626d3b20a2aa4897471478de85d12370468"
|
| 305 |
+
},
|
| 306 |
+
"control_matched__agreement-step256/eval_holdout_agreement.jsonl": {
|
| 307 |
+
"size": 92921,
|
| 308 |
+
"sha256": "93da1e6a4330ee17a52e17aeca55aa56969f1d1cf0e25b94e8134bb535e10f21"
|
| 309 |
+
},
|
| 310 |
+
"control_matched__agreement-step256/eval_holdout_conflict.jsonl": {
|
| 311 |
+
"size": 92111,
|
| 312 |
+
"sha256": "728d7b574bffa3d4eb03d28d38f67000f3a50f82ea1149f4780c2be172ff7ea0"
|
| 313 |
+
},
|
| 314 |
+
"control_matched__agreement-step256/eval_trained_adjacent.jsonl": {
|
| 315 |
+
"size": 121196,
|
| 316 |
+
"sha256": "4b1d6e1194a799c948cc67c5a4514e557e7d6c89fc6d88ee7d6b5cb295cba96f"
|
| 317 |
+
},
|
| 318 |
+
"control_matched__agreement-step256/eval_trained_agreement.jsonl": {
|
| 319 |
+
"size": 232291,
|
| 320 |
+
"sha256": "115af538c3a4e8f9208cbe76165d89ed5f6a693fb3f05dff98a15afcaa6e09af"
|
| 321 |
+
},
|
| 322 |
+
"control_matched__agreement-step256/eval_trained_conflict.jsonl": {
|
| 323 |
+
"size": 230270,
|
| 324 |
+
"sha256": "1a62e853a0890e028f978135e6972144c73469ca43b286690b8a799f3afff138"
|
| 325 |
+
},
|
| 326 |
+
"control_matched__agreement-step256/sanity.jsonl": {
|
| 327 |
+
"size": 6450,
|
| 328 |
+
"sha256": "e269c54d702f68a4bcb22b0249d1791adfd6f7c3a12786bbe0d63c7077c56320"
|
| 329 |
+
},
|
| 330 |
+
"control_matched__agreement-step256/sanity_prompts.jsonl": {
|
| 331 |
+
"size": 154858,
|
| 332 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 333 |
+
},
|
| 334 |
+
"control_matched__agreement-step32/ENDPOINT_DONE.json": {
|
| 335 |
+
"size": 90,
|
| 336 |
+
"sha256": "c8b1d4316e875552a78e72cdce99c1a21eacda4a6f46a7337fb117eaa5fc778f"
|
| 337 |
+
},
|
| 338 |
+
"control_matched__agreement-step32/eval_holdout_adjacent.jsonl": {
|
| 339 |
+
"size": 48451,
|
| 340 |
+
"sha256": "60b90c6d054e26d70dd3b0460f562b6ca7f3e8dd0d11866de9aed0bd9ea28303"
|
| 341 |
+
},
|
| 342 |
+
"control_matched__agreement-step32/eval_holdout_agreement.jsonl": {
|
| 343 |
+
"size": 92913,
|
| 344 |
+
"sha256": "6edc44bc12d399e95d90443f43d934237a5ff2496ef80655b32296337a484c1f"
|
| 345 |
+
},
|
| 346 |
+
"control_matched__agreement-step32/eval_holdout_conflict.jsonl": {
|
| 347 |
+
"size": 92125,
|
| 348 |
+
"sha256": "7261a66b39083c333dac2a75a5a3baa08668f9ea7b4f8599174e7547a5e5de02"
|
| 349 |
+
},
|
| 350 |
+
"control_matched__agreement-step32/eval_trained_adjacent.jsonl": {
|
| 351 |
+
"size": 121188,
|
| 352 |
+
"sha256": "914a077f0fe4cc5398c29c290638979a43bd37bbd6b388c180d90ad8c4b32bc6"
|
| 353 |
+
},
|
| 354 |
+
"control_matched__agreement-step32/eval_trained_agreement.jsonl": {
|
| 355 |
+
"size": 232301,
|
| 356 |
+
"sha256": "1e7558974d76126567d7e444a50234144880f249b9282be1e736dd63204eb78c"
|
| 357 |
+
},
|
| 358 |
+
"control_matched__agreement-step32/eval_trained_conflict.jsonl": {
|
| 359 |
+
"size": 230212,
|
| 360 |
+
"sha256": "699f2d1dc3b162353d11b2290c1d4f50e8e1dfd8073b554bd8d813e71c23332e"
|
| 361 |
+
},
|
| 362 |
+
"control_matched__agreement-step32/sanity.jsonl": {
|
| 363 |
+
"size": 6453,
|
| 364 |
+
"sha256": "5d331a586bcc5b8c54b755079d5a5d34d67391630eb841027840282af7cdde07"
|
| 365 |
+
},
|
| 366 |
+
"control_matched__agreement-step32/sanity_prompts.jsonl": {
|
| 367 |
+
"size": 154858,
|
| 368 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 369 |
+
},
|
| 370 |
+
"control_matched__agreement-step512/ENDPOINT_DONE.json": {
|
| 371 |
+
"size": 91,
|
| 372 |
+
"sha256": "47c49cb2addcfbcb1ddfa9a83758721644535b6c638b91d3985a35cd08645eff"
|
| 373 |
+
},
|
| 374 |
+
"control_matched__agreement-step512/eval_holdout_adjacent.jsonl": {
|
| 375 |
+
"size": 48456,
|
| 376 |
+
"sha256": "1a7c1cc710f3fb9825764b2a88ace2c70ee4a2a7dfd54b7092c57efc60842bd4"
|
| 377 |
+
},
|
| 378 |
+
"control_matched__agreement-step512/eval_holdout_agreement.jsonl": {
|
| 379 |
+
"size": 92911,
|
| 380 |
+
"sha256": "b97251d5bd3864e0b7dad325ace1235602b882bd50ad781faa5515bfbf9d6e1e"
|
| 381 |
+
},
|
| 382 |
+
"control_matched__agreement-step512/eval_holdout_conflict.jsonl": {
|
| 383 |
+
"size": 92103,
|
| 384 |
+
"sha256": "af561c7a1f733dbb9b89343f2c36ea001bbb47287843a8d91b5ccfa29f233773"
|
| 385 |
+
},
|
| 386 |
+
"control_matched__agreement-step512/eval_trained_adjacent.jsonl": {
|
| 387 |
+
"size": 121163,
|
| 388 |
+
"sha256": "a74df8613a2a609466b0797e0b348f573ae1cbd14701242aff76672888590493"
|
| 389 |
+
},
|
| 390 |
+
"control_matched__agreement-step512/eval_trained_agreement.jsonl": {
|
| 391 |
+
"size": 232287,
|
| 392 |
+
"sha256": "0fcba2b8c54979de5e08d2707b78e68fd0e737d60ef630d31fdfeadf5d21793a"
|
| 393 |
+
},
|
| 394 |
+
"control_matched__agreement-step512/eval_trained_conflict.jsonl": {
|
| 395 |
+
"size": 230203,
|
| 396 |
+
"sha256": "a1b732dd04af149301f93cca6321278c9ecf8f4411bff28a11a04c86001534b1"
|
| 397 |
+
},
|
| 398 |
+
"control_matched__agreement-step512/sanity.jsonl": {
|
| 399 |
+
"size": 6451,
|
| 400 |
+
"sha256": "8c42a7fa7d11c3aed80009dbe7b2460f125eb2ef9a4a325fa8783d93e3294e46"
|
| 401 |
+
},
|
| 402 |
+
"control_matched__agreement-step512/sanity_prompts.jsonl": {
|
| 403 |
+
"size": 154858,
|
| 404 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 405 |
+
},
|
| 406 |
+
"control_matched__agreement-step64/ENDPOINT_DONE.json": {
|
| 407 |
+
"size": 90,
|
| 408 |
+
"sha256": "f01c75b6216c7b16252dde1ff1daff2c44a6c8e2e7ab90f95767b40832349845"
|
| 409 |
+
},
|
| 410 |
+
"control_matched__agreement-step64/eval_holdout_adjacent.jsonl": {
|
| 411 |
+
"size": 48448,
|
| 412 |
+
"sha256": "6a906cbbf4cdc1c98608caa63f8230832e3bdba079339380cc6919e653594bf9"
|
| 413 |
+
},
|
| 414 |
+
"control_matched__agreement-step64/eval_holdout_agreement.jsonl": {
|
| 415 |
+
"size": 92894,
|
| 416 |
+
"sha256": "427140dd2f4c3904f8f4c8ba4474e3a432859f8413d16c244d7af4b84eaffd5c"
|
| 417 |
+
},
|
| 418 |
+
"control_matched__agreement-step64/eval_holdout_conflict.jsonl": {
|
| 419 |
+
"size": 92085,
|
| 420 |
+
"sha256": "8fb55b6d7aaf19a20acda6ab240eafb56ced97ac283777d83ff85eb6e49c9348"
|
| 421 |
+
},
|
| 422 |
+
"control_matched__agreement-step64/eval_trained_adjacent.jsonl": {
|
| 423 |
+
"size": 121137,
|
| 424 |
+
"sha256": "638fc604b9fe5cf5b245b926c563c94ff777619e551ce3a52c55720c47d5ae6a"
|
| 425 |
+
},
|
| 426 |
+
"control_matched__agreement-step64/eval_trained_agreement.jsonl": {
|
| 427 |
+
"size": 232274,
|
| 428 |
+
"sha256": "816991820491702fb3bc56a2b4b27708c90bff76bacca1356bbe3f8252cfb203"
|
| 429 |
+
},
|
| 430 |
+
"control_matched__agreement-step64/eval_trained_conflict.jsonl": {
|
| 431 |
+
"size": 230136,
|
| 432 |
+
"sha256": "cfc64f731c2d2b588ae9680066984f6fd925214d8f15f38f760d8f324ff4baca"
|
| 433 |
+
},
|
| 434 |
+
"control_matched__agreement-step64/sanity.jsonl": {
|
| 435 |
+
"size": 6451,
|
| 436 |
+
"sha256": "ef649bf34e1fca5e2b5d5b59468e7c32842c701deae477f44a4aa704355b7773"
|
| 437 |
+
},
|
| 438 |
+
"control_matched__agreement-step64/sanity_prompts.jsonl": {
|
| 439 |
+
"size": 154858,
|
| 440 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 441 |
+
},
|
| 442 |
+
"sanity_prompts.jsonl": {
|
| 443 |
+
"size": 154858,
|
| 444 |
+
"sha256": "9b5e3d3a725004127c0681dc9abe20650c2177c1fa19941c8630ffa6fa99c82d"
|
| 445 |
+
}
|
| 446 |
+
}
|
| 447 |
+
}
|