Upload eval_data/meta.json with huggingface_hub
Browse files- eval_data/meta.json +278 -0
eval_data/meta.json
ADDED
|
@@ -0,0 +1,278 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"exp": "grid",
|
| 3 |
+
"domain": "grid",
|
| 4 |
+
"n_cells": 9,
|
| 5 |
+
"variant": "intermediate",
|
| 6 |
+
"arm": "correct",
|
| 7 |
+
"seed": 0,
|
| 8 |
+
"composable": [
|
| 9 |
+
"anti_transpose",
|
| 10 |
+
"flip_h",
|
| 11 |
+
"flip_v",
|
| 12 |
+
"rot90",
|
| 13 |
+
"shift_rows"
|
| 14 |
+
],
|
| 15 |
+
"perms": {
|
| 16 |
+
"rot90": [
|
| 17 |
+
6,
|
| 18 |
+
3,
|
| 19 |
+
0,
|
| 20 |
+
7,
|
| 21 |
+
4,
|
| 22 |
+
1,
|
| 23 |
+
8,
|
| 24 |
+
5,
|
| 25 |
+
2
|
| 26 |
+
],
|
| 27 |
+
"rot180": [
|
| 28 |
+
8,
|
| 29 |
+
7,
|
| 30 |
+
6,
|
| 31 |
+
5,
|
| 32 |
+
4,
|
| 33 |
+
3,
|
| 34 |
+
2,
|
| 35 |
+
1,
|
| 36 |
+
0
|
| 37 |
+
],
|
| 38 |
+
"rot270": [
|
| 39 |
+
2,
|
| 40 |
+
5,
|
| 41 |
+
8,
|
| 42 |
+
1,
|
| 43 |
+
4,
|
| 44 |
+
7,
|
| 45 |
+
0,
|
| 46 |
+
3,
|
| 47 |
+
6
|
| 48 |
+
],
|
| 49 |
+
"transpose": [
|
| 50 |
+
0,
|
| 51 |
+
3,
|
| 52 |
+
6,
|
| 53 |
+
1,
|
| 54 |
+
4,
|
| 55 |
+
7,
|
| 56 |
+
2,
|
| 57 |
+
5,
|
| 58 |
+
8
|
| 59 |
+
],
|
| 60 |
+
"anti_transpose": [
|
| 61 |
+
8,
|
| 62 |
+
5,
|
| 63 |
+
2,
|
| 64 |
+
7,
|
| 65 |
+
4,
|
| 66 |
+
1,
|
| 67 |
+
6,
|
| 68 |
+
3,
|
| 69 |
+
0
|
| 70 |
+
],
|
| 71 |
+
"flip_h": [
|
| 72 |
+
2,
|
| 73 |
+
1,
|
| 74 |
+
0,
|
| 75 |
+
5,
|
| 76 |
+
4,
|
| 77 |
+
3,
|
| 78 |
+
8,
|
| 79 |
+
7,
|
| 80 |
+
6
|
| 81 |
+
],
|
| 82 |
+
"flip_v": [
|
| 83 |
+
6,
|
| 84 |
+
7,
|
| 85 |
+
8,
|
| 86 |
+
3,
|
| 87 |
+
4,
|
| 88 |
+
5,
|
| 89 |
+
0,
|
| 90 |
+
1,
|
| 91 |
+
2
|
| 92 |
+
],
|
| 93 |
+
"shift_rows": [
|
| 94 |
+
6,
|
| 95 |
+
7,
|
| 96 |
+
8,
|
| 97 |
+
0,
|
| 98 |
+
1,
|
| 99 |
+
2,
|
| 100 |
+
3,
|
| 101 |
+
4,
|
| 102 |
+
5
|
| 103 |
+
]
|
| 104 |
+
},
|
| 105 |
+
"train_comps": [
|
| 106 |
+
"anti_transpose|rot90",
|
| 107 |
+
"anti_transpose|shift_rows",
|
| 108 |
+
"flip_h|rot90",
|
| 109 |
+
"flip_v|rot90",
|
| 110 |
+
"flip_v|shift_rows",
|
| 111 |
+
"rot90|anti_transpose",
|
| 112 |
+
"rot90|flip_h",
|
| 113 |
+
"rot90|flip_v",
|
| 114 |
+
"rot90|shift_rows",
|
| 115 |
+
"shift_rows|anti_transpose",
|
| 116 |
+
"shift_rows|flip_v",
|
| 117 |
+
"shift_rows|rot90"
|
| 118 |
+
],
|
| 119 |
+
"heldout_comps": [
|
| 120 |
+
"anti_transpose|flip_h",
|
| 121 |
+
"flip_h|anti_transpose",
|
| 122 |
+
"anti_transpose|flip_v",
|
| 123 |
+
"flip_v|anti_transpose"
|
| 124 |
+
],
|
| 125 |
+
"heldout_labels": {
|
| 126 |
+
"anti_transpose|flip_h": "never_paired",
|
| 127 |
+
"flip_h|anti_transpose": "never_paired",
|
| 128 |
+
"anti_transpose|flip_v": "never_paired",
|
| 129 |
+
"flip_v|anti_transpose": "never_paired"
|
| 130 |
+
},
|
| 131 |
+
"n_comp_train_per_pairing": 200,
|
| 132 |
+
"total_comp_train": 2400,
|
| 133 |
+
"compositions": {
|
| 134 |
+
"anti_transpose|rot90": {
|
| 135 |
+
"role": "train",
|
| 136 |
+
"n_train": 200,
|
| 137 |
+
"n_seen_eval": 200
|
| 138 |
+
},
|
| 139 |
+
"anti_transpose|shift_rows": {
|
| 140 |
+
"role": "train",
|
| 141 |
+
"n_train": 200,
|
| 142 |
+
"n_seen_eval": 200
|
| 143 |
+
},
|
| 144 |
+
"flip_h|rot90": {
|
| 145 |
+
"role": "train",
|
| 146 |
+
"n_train": 200,
|
| 147 |
+
"n_seen_eval": 200
|
| 148 |
+
},
|
| 149 |
+
"flip_v|rot90": {
|
| 150 |
+
"role": "train",
|
| 151 |
+
"n_train": 200,
|
| 152 |
+
"n_seen_eval": 200
|
| 153 |
+
},
|
| 154 |
+
"flip_v|shift_rows": {
|
| 155 |
+
"role": "train",
|
| 156 |
+
"n_train": 200,
|
| 157 |
+
"n_seen_eval": 200
|
| 158 |
+
},
|
| 159 |
+
"rot90|anti_transpose": {
|
| 160 |
+
"role": "train",
|
| 161 |
+
"n_train": 200,
|
| 162 |
+
"n_seen_eval": 200
|
| 163 |
+
},
|
| 164 |
+
"rot90|flip_h": {
|
| 165 |
+
"role": "train",
|
| 166 |
+
"n_train": 200,
|
| 167 |
+
"n_seen_eval": 200
|
| 168 |
+
},
|
| 169 |
+
"rot90|flip_v": {
|
| 170 |
+
"role": "train",
|
| 171 |
+
"n_train": 200,
|
| 172 |
+
"n_seen_eval": 200
|
| 173 |
+
},
|
| 174 |
+
"rot90|shift_rows": {
|
| 175 |
+
"role": "train",
|
| 176 |
+
"n_train": 200,
|
| 177 |
+
"n_seen_eval": 200
|
| 178 |
+
},
|
| 179 |
+
"shift_rows|anti_transpose": {
|
| 180 |
+
"role": "train",
|
| 181 |
+
"n_train": 200,
|
| 182 |
+
"n_seen_eval": 200
|
| 183 |
+
},
|
| 184 |
+
"shift_rows|flip_v": {
|
| 185 |
+
"role": "train",
|
| 186 |
+
"n_train": 200,
|
| 187 |
+
"n_seen_eval": 200
|
| 188 |
+
},
|
| 189 |
+
"shift_rows|rot90": {
|
| 190 |
+
"role": "train",
|
| 191 |
+
"n_train": 200,
|
| 192 |
+
"n_seen_eval": 200
|
| 193 |
+
},
|
| 194 |
+
"anti_transpose|flip_h": {
|
| 195 |
+
"role": "heldout",
|
| 196 |
+
"label": "never_paired",
|
| 197 |
+
"n_eval": 200,
|
| 198 |
+
"distinct_answers": 200,
|
| 199 |
+
"top_answer_share": 0.005
|
| 200 |
+
},
|
| 201 |
+
"flip_h|anti_transpose": {
|
| 202 |
+
"role": "heldout",
|
| 203 |
+
"label": "never_paired",
|
| 204 |
+
"n_eval": 200,
|
| 205 |
+
"distinct_answers": 200,
|
| 206 |
+
"top_answer_share": 0.005
|
| 207 |
+
},
|
| 208 |
+
"anti_transpose|flip_v": {
|
| 209 |
+
"role": "heldout",
|
| 210 |
+
"label": "never_paired",
|
| 211 |
+
"n_eval": 200,
|
| 212 |
+
"distinct_answers": 200,
|
| 213 |
+
"top_answer_share": 0.005
|
| 214 |
+
},
|
| 215 |
+
"flip_v|anti_transpose": {
|
| 216 |
+
"role": "heldout",
|
| 217 |
+
"label": "never_paired",
|
| 218 |
+
"n_eval": 200,
|
| 219 |
+
"distinct_answers": 200,
|
| 220 |
+
"top_answer_share": 0.005
|
| 221 |
+
}
|
| 222 |
+
},
|
| 223 |
+
"train_heldout_chain_overlap": [],
|
| 224 |
+
"heldout_instance_collisions": 0,
|
| 225 |
+
"coverage": {
|
| 226 |
+
"train_inner_atoms": [
|
| 227 |
+
"anti_transpose",
|
| 228 |
+
"flip_h",
|
| 229 |
+
"flip_v",
|
| 230 |
+
"rot90",
|
| 231 |
+
"shift_rows"
|
| 232 |
+
],
|
| 233 |
+
"train_outer_atoms": [
|
| 234 |
+
"anti_transpose",
|
| 235 |
+
"flip_h",
|
| 236 |
+
"flip_v",
|
| 237 |
+
"rot90",
|
| 238 |
+
"shift_rows"
|
| 239 |
+
],
|
| 240 |
+
"heldout_inner_atoms": [
|
| 241 |
+
"anti_transpose",
|
| 242 |
+
"flip_h",
|
| 243 |
+
"flip_v"
|
| 244 |
+
],
|
| 245 |
+
"heldout_outer_atoms": [
|
| 246 |
+
"anti_transpose",
|
| 247 |
+
"flip_h",
|
| 248 |
+
"flip_v"
|
| 249 |
+
],
|
| 250 |
+
"heldout_inner_count": 3,
|
| 251 |
+
"heldout_outer_count": 3,
|
| 252 |
+
"per_slot_coverage": {
|
| 253 |
+
"anti_transpose": {
|
| 254 |
+
"train_as_inner": 2,
|
| 255 |
+
"train_as_outer": 2
|
| 256 |
+
},
|
| 257 |
+
"flip_h": {
|
| 258 |
+
"train_as_inner": 1,
|
| 259 |
+
"train_as_outer": 1
|
| 260 |
+
},
|
| 261 |
+
"flip_v": {
|
| 262 |
+
"train_as_inner": 2,
|
| 263 |
+
"train_as_outer": 2
|
| 264 |
+
},
|
| 265 |
+
"rot90": {
|
| 266 |
+
"train_as_inner": 4,
|
| 267 |
+
"train_as_outer": 4
|
| 268 |
+
},
|
| 269 |
+
"shift_rows": {
|
| 270 |
+
"train_as_inner": 3,
|
| 271 |
+
"train_as_outer": 3
|
| 272 |
+
}
|
| 273 |
+
},
|
| 274 |
+
"train_covers_both_slots": true,
|
| 275 |
+
"pairing_collisions": []
|
| 276 |
+
},
|
| 277 |
+
"collisions_total": 0
|
| 278 |
+
}
|