decoding results
Browse files- decoding_results.json +500 -0
- generations.json +0 -0
decoding_results.json
ADDED
|
@@ -0,0 +1,500 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"summary": {
|
| 3 |
+
"reference_ppl_mean": 27.4270831861496,
|
| 4 |
+
"reference_ppl_median": 24.03861141204834,
|
| 5 |
+
"n_samples": 1000,
|
| 6 |
+
"model": "EleutherAI/pythia-160m",
|
| 7 |
+
"max_new_tokens": 70,
|
| 8 |
+
"beam_w1_equals_greedy_frac": 0.998,
|
| 9 |
+
"reference_judge_ppl_median": 16.601261138916016
|
| 10 |
+
},
|
| 11 |
+
"results": [
|
| 12 |
+
{
|
| 13 |
+
"precision": 0.15703556591808848,
|
| 14 |
+
"recall": 0.10770206338672424,
|
| 15 |
+
"f1": 0.11695427749203965,
|
| 16 |
+
"accuracy": 0.013886833130142854,
|
| 17 |
+
"bleu": 0.4602816499134766,
|
| 18 |
+
"self_ppl_mean": 2.129318780740721,
|
| 19 |
+
"self_ppl_median": 1.9777636367502676,
|
| 20 |
+
"distinct_1": 0.037839837248011834,
|
| 21 |
+
"distinct_2": 0.11822546571922413,
|
| 22 |
+
"rep_4gram": 0.3293539897022365,
|
| 23 |
+
"avg_words": 27.035,
|
| 24 |
+
"name": "greedy",
|
| 25 |
+
"strategy": "greedy",
|
| 26 |
+
"params": {
|
| 27 |
+
"name": "greedy",
|
| 28 |
+
"strategy": "greedy"
|
| 29 |
+
},
|
| 30 |
+
"time_s": 24.308775424957275,
|
| 31 |
+
"ms_per_prompt": 24.308775424957275,
|
| 32 |
+
"gen_tokens_mean": 69.995,
|
| 33 |
+
"judge_ppl_median": 3.7137224674224854,
|
| 34 |
+
"judge_ppl_mean": 3.8401719745397567,
|
| 35 |
+
"bertscore_p": 0.8487334251403809,
|
| 36 |
+
"bertscore_r": 0.846333384513855,
|
| 37 |
+
"bertscore_f1": 0.8474068641662598
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"precision": 0.15680400502615022,
|
| 41 |
+
"recall": 0.10761634910100995,
|
| 42 |
+
"f1": 0.11683061986311886,
|
| 43 |
+
"accuracy": 0.013858261701571426,
|
| 44 |
+
"bleu": 0.45979362180213906,
|
| 45 |
+
"self_ppl_mean": 2.129485707215857,
|
| 46 |
+
"self_ppl_median": 1.9777288774584214,
|
| 47 |
+
"distinct_1": 0.03781326236416057,
|
| 48 |
+
"distinct_2": 0.11821601289629231,
|
| 49 |
+
"rep_4gram": 0.3296551592928797,
|
| 50 |
+
"avg_words": 27.054,
|
| 51 |
+
"name": "beam_w1",
|
| 52 |
+
"strategy": "beam",
|
| 53 |
+
"params": {
|
| 54 |
+
"name": "beam_w1",
|
| 55 |
+
"strategy": "beam",
|
| 56 |
+
"width": 1
|
| 57 |
+
},
|
| 58 |
+
"time_s": 98.08044266700745,
|
| 59 |
+
"ms_per_prompt": 98.08044266700745,
|
| 60 |
+
"gen_tokens_mean": 69.995,
|
| 61 |
+
"judge_ppl_median": 3.7137224674224854,
|
| 62 |
+
"judge_ppl_mean": 3.838806267619133,
|
| 63 |
+
"bertscore_p": 0.8487208485603333,
|
| 64 |
+
"bertscore_r": 0.8463250994682312,
|
| 65 |
+
"bertscore_f1": 0.8473963141441345
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"precision": 0.2013900404889916,
|
| 69 |
+
"recall": 0.08663449975312773,
|
| 70 |
+
"f1": 0.11019996943421603,
|
| 71 |
+
"accuracy": 0.010045733414015445,
|
| 72 |
+
"bleu": 0.2953586002637482,
|
| 73 |
+
"self_ppl_mean": 1.8974450017258762,
|
| 74 |
+
"self_ppl_median": 1.8302027765214615,
|
| 75 |
+
"distinct_1": 0.04578023300333643,
|
| 76 |
+
"distinct_2": 0.13660822773823988,
|
| 77 |
+
"rep_4gram": 0.2084343790507725,
|
| 78 |
+
"avg_words": 18.283,
|
| 79 |
+
"name": "beam_w2",
|
| 80 |
+
"strategy": "beam",
|
| 81 |
+
"params": {
|
| 82 |
+
"name": "beam_w2",
|
| 83 |
+
"strategy": "beam",
|
| 84 |
+
"width": 2
|
| 85 |
+
},
|
| 86 |
+
"time_s": 96.86480498313904,
|
| 87 |
+
"ms_per_prompt": 96.86480498313904,
|
| 88 |
+
"gen_tokens_mean": 69.827,
|
| 89 |
+
"judge_ppl_median": 4.084466457366943,
|
| 90 |
+
"judge_ppl_mean": 5.8135339479446415,
|
| 91 |
+
"bertscore_p": 0.8610000014305115,
|
| 92 |
+
"bertscore_r": 0.846572756767273,
|
| 93 |
+
"bertscore_f1": 0.8535789847373962
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"precision": 0.19775710589874798,
|
| 97 |
+
"recall": 0.07414467122910813,
|
| 98 |
+
"f1": 0.09580850458664733,
|
| 99 |
+
"accuracy": 0.008091172722680333,
|
| 100 |
+
"bleu": 0.34056160814471953,
|
| 101 |
+
"self_ppl_mean": 1.7568848200945872,
|
| 102 |
+
"self_ppl_median": 1.739891141883354,
|
| 103 |
+
"distinct_1": 0.04018989048929169,
|
| 104 |
+
"distinct_2": 0.11672463933397959,
|
| 105 |
+
"rep_4gram": 0.23828927173442535,
|
| 106 |
+
"avg_words": 18.537,
|
| 107 |
+
"name": "beam_w4",
|
| 108 |
+
"strategy": "beam",
|
| 109 |
+
"params": {
|
| 110 |
+
"name": "beam_w4",
|
| 111 |
+
"strategy": "beam",
|
| 112 |
+
"width": 4
|
| 113 |
+
},
|
| 114 |
+
"time_s": 105.15537786483765,
|
| 115 |
+
"ms_per_prompt": 105.15537786483765,
|
| 116 |
+
"gen_tokens_mean": 69.264,
|
| 117 |
+
"judge_ppl_median": 3.7111488580703735,
|
| 118 |
+
"judge_ppl_mean": 549.2648122023344,
|
| 119 |
+
"bertscore_p": 0.8595147728919983,
|
| 120 |
+
"bertscore_r": 0.844566285610199,
|
| 121 |
+
"bertscore_f1": 0.8517965078353882
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"precision": 0.20413235594968865,
|
| 125 |
+
"recall": 0.16095103861563514,
|
| 126 |
+
"f1": 0.16238850568174357,
|
| 127 |
+
"accuracy": 0.015531595531122681,
|
| 128 |
+
"bleu": 0.6557075980512005,
|
| 129 |
+
"self_ppl_mean": 5.760568807491463,
|
| 130 |
+
"self_ppl_median": 5.7298509005910425,
|
| 131 |
+
"distinct_1": 0.07298136645962733,
|
| 132 |
+
"distinct_2": 0.35367849194920803,
|
| 133 |
+
"rep_4gram": 0.031188779258196065,
|
| 134 |
+
"avg_words": 31.556,
|
| 135 |
+
"name": "topk_k5",
|
| 136 |
+
"strategy": "top_k",
|
| 137 |
+
"params": {
|
| 138 |
+
"name": "topk_k5",
|
| 139 |
+
"strategy": "top_k",
|
| 140 |
+
"k": 5
|
| 141 |
+
},
|
| 142 |
+
"time_s": 23.9933123588562,
|
| 143 |
+
"ms_per_prompt": 23.9933123588562,
|
| 144 |
+
"gen_tokens_mean": 69.674,
|
| 145 |
+
"judge_ppl_median": 8.357365608215332,
|
| 146 |
+
"judge_ppl_mean": 10.3706097561121,
|
| 147 |
+
"bertscore_p": 0.8579887747764587,
|
| 148 |
+
"bertscore_r": 0.8531241416931152,
|
| 149 |
+
"bertscore_f1": 0.8554333448410034
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"precision": 0.1943437061801331,
|
| 153 |
+
"recall": 0.17018026211031476,
|
| 154 |
+
"f1": 0.16424816130281425,
|
| 155 |
+
"accuracy": 0.013720664726026286,
|
| 156 |
+
"bleu": 0.4768122146961613,
|
| 157 |
+
"self_ppl_mean": 10.232085836050464,
|
| 158 |
+
"self_ppl_median": 9.978154937877655,
|
| 159 |
+
"distinct_1": 0.08890568670775671,
|
| 160 |
+
"distinct_2": 0.47079078959187953,
|
| 161 |
+
"rep_4gram": 0.005152041399905391,
|
| 162 |
+
"avg_words": 34.396,
|
| 163 |
+
"name": "topk_k20",
|
| 164 |
+
"strategy": "top_k",
|
| 165 |
+
"params": {
|
| 166 |
+
"name": "topk_k20",
|
| 167 |
+
"strategy": "top_k",
|
| 168 |
+
"k": 20
|
| 169 |
+
},
|
| 170 |
+
"time_s": 23.697141885757446,
|
| 171 |
+
"ms_per_prompt": 23.697141885757446,
|
| 172 |
+
"gen_tokens_mean": 69.426,
|
| 173 |
+
"judge_ppl_median": 14.711297035217285,
|
| 174 |
+
"judge_ppl_mean": 18.454828331232072
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"precision": 0.18013398164473904,
|
| 178 |
+
"recall": 0.16448516354938555,
|
| 179 |
+
"f1": 0.15464004612237248,
|
| 180 |
+
"accuracy": 0.011155156675690158,
|
| 181 |
+
"bleu": 0.22976691112866304,
|
| 182 |
+
"self_ppl_mean": 14.747323534754505,
|
| 183 |
+
"self_ppl_median": 13.803862584945907,
|
| 184 |
+
"distinct_1": 0.10471868305333183,
|
| 185 |
+
"distinct_2": 0.5331651149975347,
|
| 186 |
+
"rep_4gram": 0.0014281404078550252,
|
| 187 |
+
"avg_words": 35.476,
|
| 188 |
+
"name": "topk_k50",
|
| 189 |
+
"strategy": "top_k",
|
| 190 |
+
"params": {
|
| 191 |
+
"name": "topk_k50",
|
| 192 |
+
"strategy": "top_k",
|
| 193 |
+
"k": 50
|
| 194 |
+
},
|
| 195 |
+
"time_s": 23.64177703857422,
|
| 196 |
+
"ms_per_prompt": 23.64177703857422,
|
| 197 |
+
"gen_tokens_mean": 69.404,
|
| 198 |
+
"judge_ppl_median": 20.432851791381836,
|
| 199 |
+
"judge_ppl_mean": 34.98311813139915
|
| 200 |
+
},
|
| 201 |
+
{
|
| 202 |
+
"precision": 0.20203389091922755,
|
| 203 |
+
"recall": 0.15151844096638248,
|
| 204 |
+
"f1": 0.15594146830403016,
|
| 205 |
+
"accuracy": 0.013403332883272369,
|
| 206 |
+
"bleu": 0.5574700588317076,
|
| 207 |
+
"self_ppl_mean": 7.165241852811464,
|
| 208 |
+
"self_ppl_median": 6.7217895292101915,
|
| 209 |
+
"distinct_1": 0.09238144153398391,
|
| 210 |
+
"distinct_2": 0.43317498581962566,
|
| 211 |
+
"rep_4gram": 0.019412488951243617,
|
| 212 |
+
"avg_words": 29.205,
|
| 213 |
+
"name": "topk_k50_t0.7",
|
| 214 |
+
"strategy": "top_k",
|
| 215 |
+
"params": {
|
| 216 |
+
"name": "topk_k50_t0.7",
|
| 217 |
+
"strategy": "top_k",
|
| 218 |
+
"k": 50,
|
| 219 |
+
"temperature": 0.7
|
| 220 |
+
},
|
| 221 |
+
"time_s": 23.68122911453247,
|
| 222 |
+
"ms_per_prompt": 23.68122911453247,
|
| 223 |
+
"gen_tokens_mean": 69.607,
|
| 224 |
+
"judge_ppl_median": 10.626161575317383,
|
| 225 |
+
"judge_ppl_mean": 17.331569744825362
|
| 226 |
+
},
|
| 227 |
+
{
|
| 228 |
+
"precision": 0.20909230371572832,
|
| 229 |
+
"recall": 0.14641056202688646,
|
| 230 |
+
"f1": 0.15683576868798077,
|
| 231 |
+
"accuracy": 0.015255414154035243,
|
| 232 |
+
"bleu": 0.29849302912542913,
|
| 233 |
+
"self_ppl_mean": 6.067967456491997,
|
| 234 |
+
"self_ppl_median": 5.979495040329027,
|
| 235 |
+
"distinct_1": 0.09776068923777992,
|
| 236 |
+
"distinct_2": 0.4221270626453542,
|
| 237 |
+
"rep_4gram": 0.027907301032833808,
|
| 238 |
+
"avg_words": 28.089,
|
| 239 |
+
"name": "topp_p0.5",
|
| 240 |
+
"strategy": "top_p",
|
| 241 |
+
"params": {
|
| 242 |
+
"name": "topp_p0.5",
|
| 243 |
+
"strategy": "top_p",
|
| 244 |
+
"p": 0.5
|
| 245 |
+
},
|
| 246 |
+
"time_s": 25.245723962783813,
|
| 247 |
+
"ms_per_prompt": 25.245723962783813,
|
| 248 |
+
"gen_tokens_mean": 69.876,
|
| 249 |
+
"judge_ppl_median": 9.52494764328003,
|
| 250 |
+
"judge_ppl_mean": 12.933369012594223
|
| 251 |
+
},
|
| 252 |
+
{
|
| 253 |
+
"precision": 0.1790151623639173,
|
| 254 |
+
"recall": 0.15383583830048134,
|
| 255 |
+
"f1": 0.14916733197352083,
|
| 256 |
+
"accuracy": 0.01169844779025957,
|
| 257 |
+
"bleu": 0.21170796316754,
|
| 258 |
+
"self_ppl_mean": 16.263590068968764,
|
| 259 |
+
"self_ppl_median": 14.61157853686635,
|
| 260 |
+
"distinct_1": 0.1348369399573301,
|
| 261 |
+
"distinct_2": 0.5826313804375157,
|
| 262 |
+
"rep_4gram": 0.002888150162719001,
|
| 263 |
+
"avg_words": 32.81,
|
| 264 |
+
"name": "topp_p0.8",
|
| 265 |
+
"strategy": "top_p",
|
| 266 |
+
"params": {
|
| 267 |
+
"name": "topp_p0.8",
|
| 268 |
+
"strategy": "top_p",
|
| 269 |
+
"p": 0.8
|
| 270 |
+
},
|
| 271 |
+
"time_s": 25.11012601852417,
|
| 272 |
+
"ms_per_prompt": 25.11012601852417,
|
| 273 |
+
"gen_tokens_mean": 69.23,
|
| 274 |
+
"judge_ppl_median": 23.586853981018066,
|
| 275 |
+
"judge_ppl_mean": 8093.695829333544
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"precision": 0.15996633836782667,
|
| 279 |
+
"recall": 0.1436485453422705,
|
| 280 |
+
"f1": 0.1358491321482202,
|
| 281 |
+
"accuracy": 0.00975848890740479,
|
| 282 |
+
"bleu": 0.15980548317421198,
|
| 283 |
+
"self_ppl_mean": 38.624090383453975,
|
| 284 |
+
"self_ppl_median": 32.174860783175475,
|
| 285 |
+
"distinct_1": 0.1859961430054888,
|
| 286 |
+
"distinct_2": 0.6953437891711761,
|
| 287 |
+
"rep_4gram": 0.0004418966228079527,
|
| 288 |
+
"avg_words": 33.705,
|
| 289 |
+
"name": "topp_p0.95",
|
| 290 |
+
"strategy": "top_p",
|
| 291 |
+
"params": {
|
| 292 |
+
"name": "topp_p0.95",
|
| 293 |
+
"strategy": "top_p",
|
| 294 |
+
"p": 0.95
|
| 295 |
+
},
|
| 296 |
+
"time_s": 25.150532722473145,
|
| 297 |
+
"ms_per_prompt": 25.150532722473145,
|
| 298 |
+
"gen_tokens_mean": 68.846,
|
| 299 |
+
"judge_ppl_median": 54.63137245178223,
|
| 300 |
+
"judge_ppl_mean": 3939.534708950758
|
| 301 |
+
},
|
| 302 |
+
{
|
| 303 |
+
"precision": 0.2018115838286414,
|
| 304 |
+
"recall": 0.1429945363511889,
|
| 305 |
+
"f1": 0.15115617299132603,
|
| 306 |
+
"accuracy": 0.013776023575061398,
|
| 307 |
+
"bleu": 0.557218536514725,
|
| 308 |
+
"self_ppl_mean": 5.709627802185269,
|
| 309 |
+
"self_ppl_median": 5.569257784916102,
|
| 310 |
+
"distinct_1": 0.09286149654010255,
|
| 311 |
+
"distinct_2": 0.4019261517867103,
|
| 312 |
+
"rep_4gram": 0.038554351907382906,
|
| 313 |
+
"avg_words": 27.891,
|
| 314 |
+
"name": "topp_p0.9_t0.7",
|
| 315 |
+
"strategy": "top_p",
|
| 316 |
+
"params": {
|
| 317 |
+
"name": "topp_p0.9_t0.7",
|
| 318 |
+
"strategy": "top_p",
|
| 319 |
+
"p": 0.9,
|
| 320 |
+
"temperature": 0.7
|
| 321 |
+
},
|
| 322 |
+
"time_s": 25.12675452232361,
|
| 323 |
+
"ms_per_prompt": 25.12675452232361,
|
| 324 |
+
"gen_tokens_mean": 69.744,
|
| 325 |
+
"judge_ppl_median": 9.09609079360962,
|
| 326 |
+
"judge_ppl_mean": 12.747726317882538
|
| 327 |
+
}
|
| 328 |
+
],
|
| 329 |
+
"beam_timing": [
|
| 330 |
+
{
|
| 331 |
+
"mode": "natural",
|
| 332 |
+
"width": 1,
|
| 333 |
+
"mean_s_per_prompt": 0.7000857973098755,
|
| 334 |
+
"std_s": 0.03165861046389621,
|
| 335 |
+
"mean_ms_per_token": 10.001225675855366,
|
| 336 |
+
"mean_output_tokens": 70.0,
|
| 337 |
+
"total_s": 70.00857973098755,
|
| 338 |
+
"n_prompts": 100
|
| 339 |
+
},
|
| 340 |
+
{
|
| 341 |
+
"mode": "natural",
|
| 342 |
+
"width": 2,
|
| 343 |
+
"mean_s_per_prompt": 0.7082566213607788,
|
| 344 |
+
"std_s": 0.010652223812277732,
|
| 345 |
+
"mean_ms_per_token": 10.138293554772547,
|
| 346 |
+
"mean_output_tokens": 69.86,
|
| 347 |
+
"total_s": 70.82566213607788,
|
| 348 |
+
"n_prompts": 100
|
| 349 |
+
},
|
| 350 |
+
{
|
| 351 |
+
"mode": "natural",
|
| 352 |
+
"width": 4,
|
| 353 |
+
"mean_s_per_prompt": 0.7128961229324341,
|
| 354 |
+
"std_s": 0.02926986205734486,
|
| 355 |
+
"mean_ms_per_token": 10.244703469472553,
|
| 356 |
+
"mean_output_tokens": 69.59,
|
| 357 |
+
"total_s": 71.28961229324341,
|
| 358 |
+
"n_prompts": 100
|
| 359 |
+
},
|
| 360 |
+
{
|
| 361 |
+
"mode": "natural",
|
| 362 |
+
"width": 8,
|
| 363 |
+
"mean_s_per_prompt": 0.7481088328361511,
|
| 364 |
+
"std_s": 0.03363774855082431,
|
| 365 |
+
"mean_ms_per_token": 10.809140646212507,
|
| 366 |
+
"mean_output_tokens": 69.34,
|
| 367 |
+
"total_s": 74.81088328361511,
|
| 368 |
+
"n_prompts": 100
|
| 369 |
+
},
|
| 370 |
+
{
|
| 371 |
+
"mode": "fixed",
|
| 372 |
+
"width": 1,
|
| 373 |
+
"mean_s_per_prompt": 0.693102993965149,
|
| 374 |
+
"std_s": 0.0013031337825710797,
|
| 375 |
+
"mean_ms_per_token": 9.90147134235927,
|
| 376 |
+
"mean_output_tokens": 70.0,
|
| 377 |
+
"total_s": 69.31029939651489,
|
| 378 |
+
"n_prompts": 100
|
| 379 |
+
},
|
| 380 |
+
{
|
| 381 |
+
"mode": "fixed",
|
| 382 |
+
"width": 2,
|
| 383 |
+
"mean_s_per_prompt": 0.7096165156364441,
|
| 384 |
+
"std_s": 0.0011025559183209961,
|
| 385 |
+
"mean_ms_per_token": 10.137378794806345,
|
| 386 |
+
"mean_output_tokens": 70.0,
|
| 387 |
+
"total_s": 70.96165156364441,
|
| 388 |
+
"n_prompts": 100
|
| 389 |
+
},
|
| 390 |
+
{
|
| 391 |
+
"mode": "fixed",
|
| 392 |
+
"width": 4,
|
| 393 |
+
"mean_s_per_prompt": 0.7169742965698243,
|
| 394 |
+
"std_s": 0.0014611031573721494,
|
| 395 |
+
"mean_ms_per_token": 10.24248995099749,
|
| 396 |
+
"mean_output_tokens": 70.0,
|
| 397 |
+
"total_s": 71.69742965698242,
|
| 398 |
+
"n_prompts": 100
|
| 399 |
+
},
|
| 400 |
+
{
|
| 401 |
+
"mode": "fixed",
|
| 402 |
+
"width": 8,
|
| 403 |
+
"mean_s_per_prompt": 0.754322988986969,
|
| 404 |
+
"std_s": 0.0015383658998336079,
|
| 405 |
+
"mean_ms_per_token": 10.776042699813841,
|
| 406 |
+
"mean_output_tokens": 70.0,
|
| 407 |
+
"total_s": 75.4322988986969,
|
| 408 |
+
"n_prompts": 100
|
| 409 |
+
}
|
| 410 |
+
],
|
| 411 |
+
"config": {
|
| 412 |
+
"model_name": "EleutherAI/pythia-160m",
|
| 413 |
+
"dataset": "hamishivi/ROCStories",
|
| 414 |
+
"split": "test",
|
| 415 |
+
"n_samples": 1000,
|
| 416 |
+
"seed": 0,
|
| 417 |
+
"max_new_tokens": 70,
|
| 418 |
+
"batch_size": 32,
|
| 419 |
+
"beam_batch_size": 8,
|
| 420 |
+
"dtype": "float32",
|
| 421 |
+
"length_penalty": 1.0,
|
| 422 |
+
"configs": [
|
| 423 |
+
{
|
| 424 |
+
"name": "greedy",
|
| 425 |
+
"strategy": "greedy"
|
| 426 |
+
},
|
| 427 |
+
{
|
| 428 |
+
"name": "beam_w1",
|
| 429 |
+
"strategy": "beam",
|
| 430 |
+
"width": 1
|
| 431 |
+
},
|
| 432 |
+
{
|
| 433 |
+
"name": "beam_w2",
|
| 434 |
+
"strategy": "beam",
|
| 435 |
+
"width": 2
|
| 436 |
+
},
|
| 437 |
+
{
|
| 438 |
+
"name": "beam_w4",
|
| 439 |
+
"strategy": "beam",
|
| 440 |
+
"width": 4
|
| 441 |
+
},
|
| 442 |
+
{
|
| 443 |
+
"name": "topk_k5",
|
| 444 |
+
"strategy": "top_k",
|
| 445 |
+
"k": 5
|
| 446 |
+
},
|
| 447 |
+
{
|
| 448 |
+
"name": "topk_k20",
|
| 449 |
+
"strategy": "top_k",
|
| 450 |
+
"k": 20
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"name": "topk_k50",
|
| 454 |
+
"strategy": "top_k",
|
| 455 |
+
"k": 50
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"name": "topk_k50_t0.7",
|
| 459 |
+
"strategy": "top_k",
|
| 460 |
+
"k": 50,
|
| 461 |
+
"temperature": 0.7
|
| 462 |
+
},
|
| 463 |
+
{
|
| 464 |
+
"name": "topp_p0.5",
|
| 465 |
+
"strategy": "top_p",
|
| 466 |
+
"p": 0.5
|
| 467 |
+
},
|
| 468 |
+
{
|
| 469 |
+
"name": "topp_p0.8",
|
| 470 |
+
"strategy": "top_p",
|
| 471 |
+
"p": 0.8
|
| 472 |
+
},
|
| 473 |
+
{
|
| 474 |
+
"name": "topp_p0.95",
|
| 475 |
+
"strategy": "top_p",
|
| 476 |
+
"p": 0.95
|
| 477 |
+
},
|
| 478 |
+
{
|
| 479 |
+
"name": "topp_p0.9_t0.7",
|
| 480 |
+
"strategy": "top_p",
|
| 481 |
+
"p": 0.9,
|
| 482 |
+
"temperature": 0.7
|
| 483 |
+
}
|
| 484 |
+
],
|
| 485 |
+
"timing_widths": [
|
| 486 |
+
1,
|
| 487 |
+
2,
|
| 488 |
+
4,
|
| 489 |
+
8
|
| 490 |
+
],
|
| 491 |
+
"timing_samples": 100,
|
| 492 |
+
"judge_model": "EleutherAI/pythia-410m",
|
| 493 |
+
"use_bertscore": true,
|
| 494 |
+
"out_dir": "/home2/ojas.k/a2/outputs/decoding",
|
| 495 |
+
"hf_user": null,
|
| 496 |
+
"repo_name": "anlp-a2-decoding",
|
| 497 |
+
"push_to_hub": true,
|
| 498 |
+
"private": false
|
| 499 |
+
}
|
| 500 |
+
}
|
generations.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|