Ilker Kesen
commited on
Commit
·
6aca973
1
Parent(s):
280375b
include llama-3.2-3b in the leaderboard
Browse files
results/zero-shot/llama-3.2-3b.json
ADDED
|
@@ -0,0 +1,160 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": {
|
| 3 |
+
"model": "meta-llama/Llama-3.2-3B",
|
| 4 |
+
"api": "hf",
|
| 5 |
+
"dtype": "bfloat16",
|
| 6 |
+
"max_length": 131072,
|
| 7 |
+
"architecture": "LlamaForCausalLM",
|
| 8 |
+
"type": "pretrained",
|
| 9 |
+
"num_parameters": "3b"
|
| 10 |
+
},
|
| 11 |
+
"results": [
|
| 12 |
+
{
|
| 13 |
+
"name": "belebele_tr",
|
| 14 |
+
"task": "multiple_choice",
|
| 15 |
+
"acc": 0.47333333333333333,
|
| 16 |
+
"acc_norm": 0.47333333333333333
|
| 17 |
+
},
|
| 18 |
+
{
|
| 19 |
+
"name": "check_worthiness",
|
| 20 |
+
"task": "multiple_choice",
|
| 21 |
+
"acc": 0.37614259597806216,
|
| 22 |
+
"acc_norm": 0.42458866544789764
|
| 23 |
+
},
|
| 24 |
+
{
|
| 25 |
+
"name": "ironytr",
|
| 26 |
+
"task": "text_classification",
|
| 27 |
+
"acc": 0.5,
|
| 28 |
+
"acc_norm": 0.5116666666666667
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"name": "mkqa_tr",
|
| 32 |
+
"task": "extractive_question_answering",
|
| 33 |
+
"exact_match": 0.06007694584196508,
|
| 34 |
+
"f1": 0.10654344736882515
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"name": "mnli_tr",
|
| 38 |
+
"task": "natural_language_inference",
|
| 39 |
+
"acc": 0.3488,
|
| 40 |
+
"acc_norm": 0.3478
|
| 41 |
+
},
|
| 42 |
+
{
|
| 43 |
+
"name": "news_cat",
|
| 44 |
+
"task": "text_classification",
|
| 45 |
+
"acc": 0.664,
|
| 46 |
+
"acc_norm": 0.544
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"name": "offenseval_tr",
|
| 50 |
+
"task": "text_classification",
|
| 51 |
+
"acc": 0.25396825396825395,
|
| 52 |
+
"acc_norm": 0.677437641723356
|
| 53 |
+
},
|
| 54 |
+
{
|
| 55 |
+
"name": "relevance_judgment",
|
| 56 |
+
"task": "multiple_choice",
|
| 57 |
+
"acc": 0.42230347349177333,
|
| 58 |
+
"acc_norm": 0.4227605118829982
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"name": "snli_tr",
|
| 62 |
+
"task": "natural_language_inference",
|
| 63 |
+
"acc": 0.336,
|
| 64 |
+
"acc_norm": 0.3368
|
| 65 |
+
},
|
| 66 |
+
{
|
| 67 |
+
"name": "sts_tr",
|
| 68 |
+
"task": "text_classification",
|
| 69 |
+
"acc": 0.1319796954314721,
|
| 70 |
+
"acc_norm": 0.20667150108774474
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"name": "tquad",
|
| 74 |
+
"task": "extractive_question_answering",
|
| 75 |
+
"exact_match": 0.21188340807174888,
|
| 76 |
+
"f1": 0.4583574684635767
|
| 77 |
+
},
|
| 78 |
+
{
|
| 79 |
+
"name": "turkish_plu_goal_inference",
|
| 80 |
+
"task": "multiple_choice",
|
| 81 |
+
"acc": 0.3906810035842294,
|
| 82 |
+
"acc_norm": 0.3906810035842294
|
| 83 |
+
},
|
| 84 |
+
{
|
| 85 |
+
"name": "turkish_plu_next_event_prediction",
|
| 86 |
+
"task": "multiple_choice",
|
| 87 |
+
"acc": 0.4122137404580153,
|
| 88 |
+
"acc_norm": 0.5389312977099237
|
| 89 |
+
},
|
| 90 |
+
{
|
| 91 |
+
"name": "turkish_plu_step_inference",
|
| 92 |
+
"task": "multiple_choice",
|
| 93 |
+
"acc": 0.30718954248366015,
|
| 94 |
+
"acc_norm": 0.4493464052287582
|
| 95 |
+
},
|
| 96 |
+
{
|
| 97 |
+
"name": "turkish_plu_step_ordering",
|
| 98 |
+
"task": "multiple_choice",
|
| 99 |
+
"acc": 0.5974534769833496,
|
| 100 |
+
"acc_norm": 0.5974534769833496
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"name": "xcopa_tr",
|
| 104 |
+
"task": "multiple_choice",
|
| 105 |
+
"acc": 0.57,
|
| 106 |
+
"acc_norm": 0.57
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"name": "xnli_tr",
|
| 110 |
+
"task": "natural_language_inference",
|
| 111 |
+
"acc": 0.457429718875502,
|
| 112 |
+
"acc_norm": 0.457429718875502
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"name": "xquad_tr",
|
| 116 |
+
"task": "extractive_question_answering",
|
| 117 |
+
"exact_match": 0.15546218487394958,
|
| 118 |
+
"f1": 0.27803412264497246
|
| 119 |
+
},
|
| 120 |
+
{
|
| 121 |
+
"name": "gecturk_generation",
|
| 122 |
+
"task": "grammatical_error_correction",
|
| 123 |
+
"exact_match": 0.004670422263951081
|
| 124 |
+
},
|
| 125 |
+
{
|
| 126 |
+
"name": "mlsum_tr",
|
| 127 |
+
"task": "summarization",
|
| 128 |
+
"rouge1": 0.365518560170642,
|
| 129 |
+
"rouge2": 0.23287817844657932,
|
| 130 |
+
"rougeL": 0.3078744009567239
|
| 131 |
+
},
|
| 132 |
+
{
|
| 133 |
+
"name": "tr-wikihow-summ",
|
| 134 |
+
"task": "summarization",
|
| 135 |
+
"rouge1": 0.21593529633420477,
|
| 136 |
+
"rouge2": 0.06503735021906278,
|
| 137 |
+
"rougeL": 0.15632749796055023
|
| 138 |
+
},
|
| 139 |
+
{
|
| 140 |
+
"name": "wiki_lingua_tr",
|
| 141 |
+
"task": "summarization",
|
| 142 |
+
"rouge1": 0.2043150922082644,
|
| 143 |
+
"rouge2": 0.06428071184104861,
|
| 144 |
+
"rougeL": 0.1502017339873826
|
| 145 |
+
},
|
| 146 |
+
{
|
| 147 |
+
"name": "wmt-tr-en-prompt",
|
| 148 |
+
"task": "machine_translation",
|
| 149 |
+
"wer": 0.9944581931250852,
|
| 150 |
+
"bleu": 0.07893246404648234
|
| 151 |
+
},
|
| 152 |
+
{
|
| 153 |
+
"name": "xlsum_tr",
|
| 154 |
+
"task": "summarization",
|
| 155 |
+
"rouge1": 0.2718989612690207,
|
| 156 |
+
"rouge2": 0.12031087767355472,
|
| 157 |
+
"rougeL": 0.21504760362022454
|
| 158 |
+
}
|
| 159 |
+
]
|
| 160 |
+
}
|