diff --git "a/parse/dev/rBCvMG-JsPd/rBCvMG-JsPd_middle.json" "b/parse/dev/rBCvMG-JsPd/rBCvMG-JsPd_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/dev/rBCvMG-JsPd/rBCvMG-JsPd_middle.json"
@@ -0,0 +1,42817 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 115,
+ 97,
+ 497,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 95,
+ 497,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 95,
+ 497,
+ 120
+ ],
+ "score": 1.0,
+ "content": "Few-Shot Parameter-Efficient Fine-Tuning is Better",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 158,
+ 117,
+ 453,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 158,
+ 117,
+ 453,
+ 140
+ ],
+ "score": 1.0,
+ "content": "and Cheaper than In-Context Learning",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 191,
+ 169,
+ 418,
+ 181
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 191,
+ 168,
+ 420,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 191,
+ 168,
+ 420,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Haokun Liu∗ Derek Tam∗ Mohammed Muqeeth∗",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 172,
+ 187,
+ 435,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 186,
+ 436,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 186,
+ 436,
+ 202
+ ],
+ "score": 1.0,
+ "content": "Jay Mohta Tenghao Huang Mohit Bansal Colin Raffel",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 187,
+ 204,
+ 423,
+ 236
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 237,
+ 203,
+ 374,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 237,
+ 203,
+ 374,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Department of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 217,
+ 214,
+ 393,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 217,
+ 214,
+ 393,
+ 227
+ ],
+ "score": 1.0,
+ "content": "University of North Carolina at Chapel Hill",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 187,
+ 225,
+ 424,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 187,
+ 225,
+ 424,
+ 237
+ ],
+ "score": 1.0,
+ "content": "{haokunl,dtredsox,muqeeth,craffel}@cs.unc.edu",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 283,
+ 250,
+ 328,
+ 263
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 282,
+ 250,
+ 330,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 282,
+ 250,
+ 330,
+ 264
+ ],
+ "score": 1.0,
+ "content": "Abstract",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 277,
+ 469,
+ 474
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 277,
+ 470,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 277,
+ 470,
+ 289
+ ],
+ "score": 1.0,
+ "content": "Few-shot in-context learning (ICL) enables pre-trained language models to per-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 300
+ ],
+ "score": 1.0,
+ "content": "form a previously-unseen task without any gradient-based training by feeding a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 299,
+ 469,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 299,
+ 469,
+ 311
+ ],
+ "score": 1.0,
+ "content": "small number of training examples as part of the input. ICL incurs substantial",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 310,
+ 469,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 310,
+ 469,
+ 322
+ ],
+ "score": 1.0,
+ "content": "computational, memory, and storage costs because it involves processing all of the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 320,
+ 469,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 320,
+ 469,
+ 333
+ ],
+ "score": 1.0,
+ "content": "training examples every time a prediction is made. Parameter-efficient fine-tuning",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 331,
+ 470,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 331,
+ 470,
+ 344
+ ],
+ "score": 1.0,
+ "content": "(PEFT) (e.g. adapter modules, prompt tuning, sparse update methods, etc.) offers",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 354
+ ],
+ "score": 1.0,
+ "content": "an alternative paradigm where a small set of parameters are trained to enable a",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 353,
+ 469,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 353,
+ 469,
+ 365
+ ],
+ "score": 1.0,
+ "content": "model to perform the new task. In this paper, we rigorously compare few-shot",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 364,
+ 470,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 364,
+ 470,
+ 376
+ ],
+ "score": 1.0,
+ "content": "ICL and PEFT and demonstrate that the latter offers better accuracy as well as",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 375,
+ 470,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 375,
+ 470,
+ 388
+ ],
+ "score": 1.0,
+ "content": "dramatically lower computational costs. Along the way, we introduce a new PEFT",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 385,
+ 470,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 385,
+ 203,
+ 399
+ ],
+ "score": 1.0,
+ "content": "method called",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 203,
+ 385,
+ 228,
+ 397
+ ],
+ "score": 0.5,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 228,
+ 385,
+ 470,
+ 399
+ ],
+ "score": 1.0,
+ "content": "that scales activations by learned vectors, attaining stronger",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 397,
+ 471,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 397,
+ 471,
+ 410
+ ],
+ "score": 1.0,
+ "content": "performance while only introducing a relatively tiny amount of new parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 407,
+ 470,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 407,
+ 470,
+ 420
+ ],
+ "score": 1.0,
+ "content": "We also propose a simple recipe based on the T0 model [1] called T-Few that",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 419,
+ 470,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 419,
+ 470,
+ 430
+ ],
+ "score": 1.0,
+ "content": "can be applied to new tasks without task-specific tuning or modifications. We",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 429,
+ 469,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 429,
+ 469,
+ 442
+ ],
+ "score": 1.0,
+ "content": "validate the effectiveness of T-Few on completely unseen tasks by applying it to",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 440,
+ 469,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 440,
+ 469,
+ 452
+ ],
+ "score": 1.0,
+ "content": "the RAFT benchmark [2], attaining super-human performance for the first time",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 451,
+ 470,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 451,
+ 306,
+ 463
+ ],
+ "score": 1.0,
+ "content": "and outperforming the state-of-the-art by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 307,
+ 451,
+ 322,
+ 462
+ ],
+ "score": 0.86,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 451,
+ 470,
+ 463
+ ],
+ "score": 1.0,
+ "content": "absolute. All of the code used in our",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 463,
+ 280,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 463,
+ 280,
+ 474
+ ],
+ "score": 1.0,
+ "content": "experiments is publicly available.1",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 496,
+ 190,
+ 510
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 192,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 192,
+ 512
+ ],
+ "score": 1.0,
+ "content": "1 Introduction",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 522,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "Pre-trained language models have become a cornerstone of natural language processing, thanks",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "to the fact that they can dramatically improve data efficiency on tasks of interest – i.e., using a",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 506,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 506,
+ 556
+ ],
+ "score": 1.0,
+ "content": "pre-trained language model for initialization often produces better results with less labeled data. A",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 555,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 555,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "historically common approach has been gradient-based fine-tuning on a downstream task of interest",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "with pre-trained parameters as the initialization. While fine-tuning has produced many state-of-the-art",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 589
+ ],
+ "score": 1.0,
+ "content": "results [1], it results in a model specialized for a single task with an entirely new set of parameter",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 443,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 443,
+ 600
+ ],
+ "score": 1.0,
+ "content": "values, which can become impractical when fine-tuning on many downstream tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 680
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "An alternative approach popularized by [3, 4] is in-context learning (ICL), which induces a model",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 627
+ ],
+ "score": 1.0,
+ "content": "to perform a downstream task by inputting prompted examples. Few-shot prompting converts a",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 625,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 625,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "small collection of input-target pairs into (typically) human-understandable instructions and examples",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "[3, 4], along with a single unlabeled example for which a prediction is desired. Notably, ICL requires",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 647,
+ 506,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 506,
+ 660
+ ],
+ "score": 1.0,
+ "content": "no gradient-based training and therefore allows a single model to immediately perform a wide variety",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 671
+ ],
+ "score": 1.0,
+ "content": "of tasks. Performing ICL therefore solely relies on the capabilities that a model learned during",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 670,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "pre-training. These characteristics have led to a great deal of recent interest in ICL methods [5–10].",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 119,
+ 691,
+ 278,
+ 711
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 689,
+ 195,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 689,
+ 195,
+ 701
+ ],
+ "score": 1.0,
+ "content": "∗Equal contribution.",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 122,
+ 701,
+ 279,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 701,
+ 279,
+ 712
+ ],
+ "score": 1.0,
+ "content": "1 https://github.com/r-three/t-few",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 731,
+ 385,
+ 741
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 730,
+ 386,
+ 744
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 730,
+ 386,
+ 744
+ ],
+ "score": 1.0,
+ "content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 115,
+ 97,
+ 497,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 95,
+ 497,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 95,
+ 497,
+ 120
+ ],
+ "score": 1.0,
+ "content": "Few-Shot Parameter-Efficient Fine-Tuning is Better",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 158,
+ 117,
+ 453,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 158,
+ 117,
+ 453,
+ 140
+ ],
+ "score": 1.0,
+ "content": "and Cheaper than In-Context Learning",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 191,
+ 169,
+ 418,
+ 181
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 191,
+ 168,
+ 420,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 191,
+ 168,
+ 420,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Haokun Liu∗ Derek Tam∗ Mohammed Muqeeth∗",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 191,
+ 168,
+ 420,
+ 184
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 172,
+ 187,
+ 435,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 186,
+ 436,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 186,
+ 436,
+ 202
+ ],
+ "score": 1.0,
+ "content": "Jay Mohta Tenghao Huang Mohit Bansal Colin Raffel",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 172,
+ 186,
+ 436,
+ 202
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 187,
+ 204,
+ 423,
+ 236
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 237,
+ 203,
+ 374,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 237,
+ 203,
+ 374,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Department of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 217,
+ 214,
+ 393,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 217,
+ 214,
+ 393,
+ 227
+ ],
+ "score": 1.0,
+ "content": "University of North Carolina at Chapel Hill",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 187,
+ 225,
+ 424,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 187,
+ 225,
+ 424,
+ 237
+ ],
+ "score": 1.0,
+ "content": "{haokunl,dtredsox,muqeeth,craffel}@cs.unc.edu",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 187,
+ 203,
+ 424,
+ 237
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 283,
+ 250,
+ 328,
+ 263
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 282,
+ 250,
+ 330,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 282,
+ 250,
+ 330,
+ 264
+ ],
+ "score": 1.0,
+ "content": "Abstract",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 277,
+ 469,
+ 474
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 277,
+ 470,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 277,
+ 470,
+ 289
+ ],
+ "score": 1.0,
+ "content": "Few-shot in-context learning (ICL) enables pre-trained language models to per-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 300
+ ],
+ "score": 1.0,
+ "content": "form a previously-unseen task without any gradient-based training by feeding a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 299,
+ 469,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 299,
+ 469,
+ 311
+ ],
+ "score": 1.0,
+ "content": "small number of training examples as part of the input. ICL incurs substantial",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 310,
+ 469,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 310,
+ 469,
+ 322
+ ],
+ "score": 1.0,
+ "content": "computational, memory, and storage costs because it involves processing all of the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 320,
+ 469,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 320,
+ 469,
+ 333
+ ],
+ "score": 1.0,
+ "content": "training examples every time a prediction is made. Parameter-efficient fine-tuning",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 331,
+ 470,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 331,
+ 470,
+ 344
+ ],
+ "score": 1.0,
+ "content": "(PEFT) (e.g. adapter modules, prompt tuning, sparse update methods, etc.) offers",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 354
+ ],
+ "score": 1.0,
+ "content": "an alternative paradigm where a small set of parameters are trained to enable a",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 353,
+ 469,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 353,
+ 469,
+ 365
+ ],
+ "score": 1.0,
+ "content": "model to perform the new task. In this paper, we rigorously compare few-shot",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 364,
+ 470,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 364,
+ 470,
+ 376
+ ],
+ "score": 1.0,
+ "content": "ICL and PEFT and demonstrate that the latter offers better accuracy as well as",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 375,
+ 470,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 375,
+ 470,
+ 388
+ ],
+ "score": 1.0,
+ "content": "dramatically lower computational costs. Along the way, we introduce a new PEFT",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 385,
+ 470,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 385,
+ 203,
+ 399
+ ],
+ "score": 1.0,
+ "content": "method called",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 203,
+ 385,
+ 228,
+ 397
+ ],
+ "score": 0.5,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 228,
+ 385,
+ 470,
+ 399
+ ],
+ "score": 1.0,
+ "content": "that scales activations by learned vectors, attaining stronger",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 397,
+ 471,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 397,
+ 471,
+ 410
+ ],
+ "score": 1.0,
+ "content": "performance while only introducing a relatively tiny amount of new parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 407,
+ 470,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 407,
+ 470,
+ 420
+ ],
+ "score": 1.0,
+ "content": "We also propose a simple recipe based on the T0 model [1] called T-Few that",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 419,
+ 470,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 419,
+ 470,
+ 430
+ ],
+ "score": 1.0,
+ "content": "can be applied to new tasks without task-specific tuning or modifications. We",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 429,
+ 469,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 429,
+ 469,
+ 442
+ ],
+ "score": 1.0,
+ "content": "validate the effectiveness of T-Few on completely unseen tasks by applying it to",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 440,
+ 469,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 440,
+ 469,
+ 452
+ ],
+ "score": 1.0,
+ "content": "the RAFT benchmark [2], attaining super-human performance for the first time",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 451,
+ 470,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 451,
+ 306,
+ 463
+ ],
+ "score": 1.0,
+ "content": "and outperforming the state-of-the-art by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 307,
+ 451,
+ 322,
+ 462
+ ],
+ "score": 0.86,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 451,
+ 470,
+ 463
+ ],
+ "score": 1.0,
+ "content": "absolute. All of the code used in our",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 463,
+ 280,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 463,
+ 280,
+ 474
+ ],
+ "score": 1.0,
+ "content": "experiments is publicly available.1",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 141,
+ 277,
+ 471,
+ 474
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 496,
+ 190,
+ 510
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 192,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 192,
+ 512
+ ],
+ "score": 1.0,
+ "content": "1 Introduction",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 522,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "Pre-trained language models have become a cornerstone of natural language processing, thanks",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "to the fact that they can dramatically improve data efficiency on tasks of interest – i.e., using a",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 506,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 506,
+ 556
+ ],
+ "score": 1.0,
+ "content": "pre-trained language model for initialization often produces better results with less labeled data. A",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 555,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 555,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "historically common approach has been gradient-based fine-tuning on a downstream task of interest",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "with pre-trained parameters as the initialization. While fine-tuning has produced many state-of-the-art",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 589
+ ],
+ "score": 1.0,
+ "content": "results [1], it results in a model specialized for a single task with an entirely new set of parameter",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 443,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 443,
+ 600
+ ],
+ "score": 1.0,
+ "content": "values, which can become impractical when fine-tuning on many downstream tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 522,
+ 506,
+ 600
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 680
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "An alternative approach popularized by [3, 4] is in-context learning (ICL), which induces a model",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 627
+ ],
+ "score": 1.0,
+ "content": "to perform a downstream task by inputting prompted examples. Few-shot prompting converts a",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 625,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 625,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "small collection of input-target pairs into (typically) human-understandable instructions and examples",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "[3, 4], along with a single unlabeled example for which a prediction is desired. Notably, ICL requires",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 647,
+ 506,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 506,
+ 660
+ ],
+ "score": 1.0,
+ "content": "no gradient-based training and therefore allows a single model to immediately perform a wide variety",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 671
+ ],
+ "score": 1.0,
+ "content": "of tasks. Performing ICL therefore solely relies on the capabilities that a model learned during",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 670,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "pre-training. These characteristics have led to a great deal of recent interest in ICL methods [5–10].",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 105,
+ 603,
+ 506,
+ 681
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 154
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 154
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 154
+ ],
+ "score": 0.954,
+ "type": "image",
+ "image_path": "f3a2decc1db9d1ee4079a3c65d82a4f73bb8450c9cfd81b9e8bca6486d996280.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 98.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 98.0,
+ 506,
+ 126.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 126.0,
+ 506,
+ 154.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 159,
+ 505,
+ 226
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 159,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 159,
+ 194,
+ 172
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Diagram of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 159,
+ 219,
+ 170
+ ],
+ "score": 0.48,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 159,
+ 420,
+ 172
+ ],
+ "score": 1.0,
+ "content": "and the loss terms used in the T-Few recipe. Left:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 159,
+ 446,
+ 171
+ ],
+ "score": 0.35,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 159,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "introduces the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 171,
+ 505,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 167,
+ 182
+ ],
+ "score": 1.0,
+ "content": "learned vectors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 171,
+ 188,
+ 182
+ ],
+ "score": 0.85,
+ "content": "l _ { \\mathrm { k } } , l _ { \\mathrm { v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 171,
+ 207,
+ 182
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 171,
+ 218,
+ 182
+ ],
+ "score": 0.86,
+ "content": "l _ { \\mathrm { f f } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 171,
+ 505,
+ 182
+ ],
+ "score": 1.0,
+ "content": "which respectively rescale (via element-wise multiplication, visualized as",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 116,
+ 192
+ ],
+ "score": 0.72,
+ "content": "\\odot",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 182,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": ") the keys and values in attention mechanisms and the inner activations in position-wise feed-forward",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 191,
+ 506,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 346,
+ 206
+ ],
+ "score": 1.0,
+ "content": "networks. Right: In addition to a standard cross-entropy loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 346,
+ 192,
+ 367,
+ 203
+ ],
+ "score": 0.89,
+ "content": "L _ { \\mathrm { L M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 191,
+ 506,
+ 206
+ ],
+ "score": 1.0,
+ "content": ", we introduce an unlikelihood loss",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 203,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 203,
+ 126,
+ 214
+ ],
+ "score": 0.89,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 126,
+ 203,
+ 429,
+ 216
+ ],
+ "score": 1.0,
+ "content": "that lowers the probability of incorrect outputs and a length-normalized loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 204,
+ 449,
+ 214
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 203,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "that applies a",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 214,
+ 495,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 214,
+ 495,
+ 226
+ ],
+ "score": 1.0,
+ "content": "standard softmax cross-entropy loss to length-normalized log-probabilities of all output choices.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "Despite the practical benefits of ICL, it has several major drawbacks. First, processing all prompted",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "input-target pairs every time the model makes a prediction incurs significant compute costs. Second,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 295
+ ],
+ "score": 1.0,
+ "content": "ICL typically produces inferior performance compared to fine-tuning [4]. Finally, the exact formatting",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 293,
+ 505,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 293,
+ 505,
+ 305
+ ],
+ "score": 1.0,
+ "content": "of the prompt (including the wording [11] and ordering of examples [12]) can have significant and",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 302,
+ 507,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 507,
+ 317
+ ],
+ "score": 1.0,
+ "content": "unpredictable impact on the model’s performance, far beyond inter-run variation of fine-tuning.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 327
+ ],
+ "score": 1.0,
+ "content": "Recent work has also demonstrated that ICL can perform well even when provided with incorrect",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 403,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 403,
+ 338
+ ],
+ "score": 1.0,
+ "content": "labels, raising questions as to how much learning is taking place at all [9].",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 342,
+ 505,
+ 408
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 341,
+ 507,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 341,
+ 507,
+ 354
+ ],
+ "score": 1.0,
+ "content": "An additional paradigm for enabling a model to perform a new task with minimal updates is parameter-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 506,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 353,
+ 506,
+ 366
+ ],
+ "score": 1.0,
+ "content": "efficient fine-tuning (PEFT), where a pre-trained model is fine-tuned by only updating a small number",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "of added or selected parameters. Recent methods have matched the performance of fine-tuning the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 353,
+ 387
+ ],
+ "score": 1.0,
+ "content": "full model while only updating or adding a small fraction (e.g.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 375,
+ 381,
+ 385
+ ],
+ "score": 0.84,
+ "content": "0 . 0 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 382,
+ 374,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": ") of the full model’s parameters",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "[13, 14]. Furthermore, certain PEFT methods allow mixed-task batches where different examples in",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 488,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 488,
+ 409
+ ],
+ "score": 1.0,
+ "content": "a batch are processed differently [14], making both PEFT and ICL viable for multitask models.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 412,
+ 505,
+ 576
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 412,
+ 505,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 412,
+ 505,
+ 425
+ ],
+ "score": 1.0,
+ "content": "While the benefits of PEFT address some shortcomings of fine-tuning (when compared to ICL), there",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 424,
+ 504,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 504,
+ 435
+ ],
+ "score": 1.0,
+ "content": "has been relatively little focus on whether PEFT methods work well when very little labeled data",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 433,
+ 506,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 433,
+ 506,
+ 448
+ ],
+ "score": 1.0,
+ "content": "is available. Our primary goal in this paper is to close this gap by proposing a recipe – i.e., a model, a",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 459
+ ],
+ "score": 1.0,
+ "content": "PEFT method, and a fixed set of hyperparameters – that attains strong performance on novel, unseen",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 456,
+ 505,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 456,
+ 505,
+ 469
+ ],
+ "score": 1.0,
+ "content": "tasks while only updating a tiny fraction of the model’s parameters. Specifically, we base our approach",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "on the T0 model [1], a variant of T5 [15] fine-tuned on a multitask mixture of prompted datasets.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "To improve performance on classification and multiple-choice tasks, we add unlikelihood [16, 17]",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 505,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 505,
+ 501
+ ],
+ "score": 1.0,
+ "content": "and length normalization-based [4] loss terms. In addition, we develop (IA)3, a PEFT method",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 499,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 499,
+ 343,
+ 513
+ ],
+ "score": 1.0,
+ "content": "that multiplies intermediate activations by learned vectors.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 500,
+ 368,
+ 511
+ ],
+ "score": 0.82,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 499,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "attains stronger performance than",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 288,
+ 523
+ ],
+ "score": 1.0,
+ "content": "full-model fine-tuning while updating up to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 511,
+ 325,
+ 522
+ ],
+ "score": 0.89,
+ "content": "1 0 { , } 0 0 0 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 326,
+ 511,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "fewer parameters. Finally, we demonstrate",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 236,
+ 535
+ ],
+ "score": 1.0,
+ "content": "the benefits of pre-training the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 521,
+ 261,
+ 533
+ ],
+ "score": 0.48,
+ "content": "( \\mathtt { I A } ) ^ { \\tilde { 3 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 520,
+ 506,
+ 535
+ ],
+ "score": 1.0,
+ "content": "parameters before fine-tuning [18, 19]. Our overall recipe,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 531,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 423,
+ 546
+ ],
+ "score": 1.0,
+ "content": "which we dub “T-Few”, performs significantly better than ICL (even against",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 533,
+ 442,
+ 544
+ ],
+ "score": 0.87,
+ "content": "1 6 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 443,
+ 531,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "larger models)",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "and outperforms humans for the first time on the real-world few-shot learning benchmark RAFT [2]",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 554,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 554,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "while requiring dramatically less compute and allowing for mixed-task batches during inference. To",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 490,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 490,
+ 578
+ ],
+ "score": 1.0,
+ "content": "facilitate the use of T-Few on new problems and future research on PEFT, we release our code.1",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 582,
+ 505,
+ 615
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 506,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 506,
+ 594
+ ],
+ "score": 1.0,
+ "content": "After providing background on ICL and PEFT in the following section, we discuss the design of",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 592,
+ 507,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 592,
+ 507,
+ 605
+ ],
+ "score": 1.0,
+ "content": "T-Few in section 3. In section 4, we present experiments comparing T-Few to strong ICL baselines.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 603,
+ 392,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 392,
+ 615
+ ],
+ "score": 1.0,
+ "content": "Finally, we discuss related work in section 5 and conclude in section 6.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 644,
+ 189,
+ 658
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 103,
+ 642,
+ 191,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 642,
+ 191,
+ 662
+ ],
+ "score": 1.0,
+ "content": "2 Background",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 677,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "In this section, we provide am verview of ICL and PEFT with a focus on characterizing the com-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "putation, memory, and on-disk storage costs of making a prediction. Real-world costs depend on",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "implementation and hardware, so we report costs in terms of FLOPs for computation and bytes for",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 444,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 444,
+ 722
+ ],
+ "score": 1.0,
+ "content": "memory and storage, respectively. Additional related work is discussed in section 5.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 42.5
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 333,
+ 72,
+ 433,
+ 81
+ ],
+ "lines": []
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 753
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 753
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 154
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 154
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 154
+ ],
+ "score": 0.954,
+ "type": "image",
+ "image_path": "f3a2decc1db9d1ee4079a3c65d82a4f73bb8450c9cfd81b9e8bca6486d996280.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 506,
+ 98.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 98.0,
+ 506,
+ 126.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 126.0,
+ 506,
+ 154.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 159,
+ 505,
+ 226
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 159,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 159,
+ 194,
+ 172
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Diagram of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 159,
+ 219,
+ 170
+ ],
+ "score": 0.48,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 159,
+ 420,
+ 172
+ ],
+ "score": 1.0,
+ "content": "and the loss terms used in the T-Few recipe. Left:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 159,
+ 446,
+ 171
+ ],
+ "score": 0.35,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 159,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "introduces the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 171,
+ 505,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 167,
+ 182
+ ],
+ "score": 1.0,
+ "content": "learned vectors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 171,
+ 188,
+ 182
+ ],
+ "score": 0.85,
+ "content": "l _ { \\mathrm { k } } , l _ { \\mathrm { v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 171,
+ 207,
+ 182
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 171,
+ 218,
+ 182
+ ],
+ "score": 0.86,
+ "content": "l _ { \\mathrm { f f } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 171,
+ 505,
+ 182
+ ],
+ "score": 1.0,
+ "content": "which respectively rescale (via element-wise multiplication, visualized as",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 116,
+ 192
+ ],
+ "score": 0.72,
+ "content": "\\odot",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 182,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": ") the keys and values in attention mechanisms and the inner activations in position-wise feed-forward",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 191,
+ 506,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 346,
+ 206
+ ],
+ "score": 1.0,
+ "content": "networks. Right: In addition to a standard cross-entropy loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 346,
+ 192,
+ 367,
+ 203
+ ],
+ "score": 0.89,
+ "content": "L _ { \\mathrm { L M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 191,
+ 506,
+ 206
+ ],
+ "score": 1.0,
+ "content": ", we introduce an unlikelihood loss",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 203,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 203,
+ 126,
+ 214
+ ],
+ "score": 0.89,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 126,
+ 203,
+ 429,
+ 216
+ ],
+ "score": 1.0,
+ "content": "that lowers the probability of incorrect outputs and a length-normalized loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 204,
+ 449,
+ 214
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 203,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "that applies a",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 214,
+ 495,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 214,
+ 495,
+ 226
+ ],
+ "score": 1.0,
+ "content": "standard softmax cross-entropy loss to length-normalized log-probabilities of all output choices.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "Despite the practical benefits of ICL, it has several major drawbacks. First, processing all prompted",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "input-target pairs every time the model makes a prediction incurs significant compute costs. Second,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 295
+ ],
+ "score": 1.0,
+ "content": "ICL typically produces inferior performance compared to fine-tuning [4]. Finally, the exact formatting",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 293,
+ 505,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 293,
+ 505,
+ 305
+ ],
+ "score": 1.0,
+ "content": "of the prompt (including the wording [11] and ordering of examples [12]) can have significant and",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 302,
+ 507,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 507,
+ 317
+ ],
+ "score": 1.0,
+ "content": "unpredictable impact on the model’s performance, far beyond inter-run variation of fine-tuning.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 327
+ ],
+ "score": 1.0,
+ "content": "Recent work has also demonstrated that ICL can perform well even when provided with incorrect",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 403,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 403,
+ 338
+ ],
+ "score": 1.0,
+ "content": "labels, raising questions as to how much learning is taking place at all [9].",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 260,
+ 507,
+ 338
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 342,
+ 505,
+ 408
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 341,
+ 507,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 341,
+ 507,
+ 354
+ ],
+ "score": 1.0,
+ "content": "An additional paradigm for enabling a model to perform a new task with minimal updates is parameter-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 506,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 353,
+ 506,
+ 366
+ ],
+ "score": 1.0,
+ "content": "efficient fine-tuning (PEFT), where a pre-trained model is fine-tuned by only updating a small number",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "of added or selected parameters. Recent methods have matched the performance of fine-tuning the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 353,
+ 387
+ ],
+ "score": 1.0,
+ "content": "full model while only updating or adding a small fraction (e.g.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 375,
+ 381,
+ 385
+ ],
+ "score": 0.84,
+ "content": "0 . 0 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 382,
+ 374,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": ") of the full model’s parameters",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "[13, 14]. Furthermore, certain PEFT methods allow mixed-task batches where different examples in",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 488,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 488,
+ 409
+ ],
+ "score": 1.0,
+ "content": "a batch are processed differently [14], making both PEFT and ICL viable for multitask models.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 105,
+ 341,
+ 507,
+ 409
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 412,
+ 505,
+ 576
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 412,
+ 505,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 412,
+ 505,
+ 425
+ ],
+ "score": 1.0,
+ "content": "While the benefits of PEFT address some shortcomings of fine-tuning (when compared to ICL), there",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 424,
+ 504,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 504,
+ 435
+ ],
+ "score": 1.0,
+ "content": "has been relatively little focus on whether PEFT methods work well when very little labeled data",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 433,
+ 506,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 433,
+ 506,
+ 448
+ ],
+ "score": 1.0,
+ "content": "is available. Our primary goal in this paper is to close this gap by proposing a recipe – i.e., a model, a",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 459
+ ],
+ "score": 1.0,
+ "content": "PEFT method, and a fixed set of hyperparameters – that attains strong performance on novel, unseen",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 456,
+ 505,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 456,
+ 505,
+ 469
+ ],
+ "score": 1.0,
+ "content": "tasks while only updating a tiny fraction of the model’s parameters. Specifically, we base our approach",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "on the T0 model [1], a variant of T5 [15] fine-tuned on a multitask mixture of prompted datasets.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "To improve performance on classification and multiple-choice tasks, we add unlikelihood [16, 17]",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 505,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 505,
+ 501
+ ],
+ "score": 1.0,
+ "content": "and length normalization-based [4] loss terms. In addition, we develop (IA)3, a PEFT method",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 499,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 499,
+ 343,
+ 513
+ ],
+ "score": 1.0,
+ "content": "that multiplies intermediate activations by learned vectors.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 500,
+ 368,
+ 511
+ ],
+ "score": 0.82,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 499,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "attains stronger performance than",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 288,
+ 523
+ ],
+ "score": 1.0,
+ "content": "full-model fine-tuning while updating up to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 511,
+ 325,
+ 522
+ ],
+ "score": 0.89,
+ "content": "1 0 { , } 0 0 0 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 326,
+ 511,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "fewer parameters. Finally, we demonstrate",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 236,
+ 535
+ ],
+ "score": 1.0,
+ "content": "the benefits of pre-training the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 521,
+ 261,
+ 533
+ ],
+ "score": 0.48,
+ "content": "( \\mathtt { I A } ) ^ { \\tilde { 3 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 520,
+ 506,
+ 535
+ ],
+ "score": 1.0,
+ "content": "parameters before fine-tuning [18, 19]. Our overall recipe,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 531,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 423,
+ 546
+ ],
+ "score": 1.0,
+ "content": "which we dub “T-Few”, performs significantly better than ICL (even against",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 533,
+ 442,
+ 544
+ ],
+ "score": 0.87,
+ "content": "1 6 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 443,
+ 531,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "larger models)",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "and outperforms humans for the first time on the real-world few-shot learning benchmark RAFT [2]",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 554,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 554,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "while requiring dramatically less compute and allowing for mixed-task batches during inference. To",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 490,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 490,
+ 578
+ ],
+ "score": 1.0,
+ "content": "facilitate the use of T-Few on new problems and future research on PEFT, we release our code.1",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 104,
+ 412,
+ 506,
+ 578
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 582,
+ 505,
+ 615
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 506,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 506,
+ 594
+ ],
+ "score": 1.0,
+ "content": "After providing background on ICL and PEFT in the following section, we discuss the design of",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 592,
+ 507,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 592,
+ 507,
+ 605
+ ],
+ "score": 1.0,
+ "content": "T-Few in section 3. In section 4, we present experiments comparing T-Few to strong ICL baselines.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 603,
+ 392,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 392,
+ 615
+ ],
+ "score": 1.0,
+ "content": "Finally, we discuss related work in section 5 and conclude in section 6.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 105,
+ 582,
+ 507,
+ 615
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 644,
+ 189,
+ 658
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 103,
+ 642,
+ 191,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 642,
+ 191,
+ 662
+ ],
+ "score": 1.0,
+ "content": "2 Background",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 677,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "In this section, we provide am verview of ICL and PEFT with a focus on characterizing the com-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "putation, memory, and on-disk storage costs of making a prediction. Real-world costs depend on",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "implementation and hardware, so we report costs in terms of FLOPs for computation and bytes for",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 444,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 444,
+ 722
+ ],
+ "score": 1.0,
+ "content": "memory and storage, respectively. Additional related work is discussed in section 5.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 105,
+ 677,
+ 506,
+ 722
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 73,
+ 278,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 70,
+ 280,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 70,
+ 280,
+ 87
+ ],
+ "score": 1.0,
+ "content": "2.1 Few-shot in-context learning (ICL)",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 96,
+ 505,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 505,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 505,
+ 109
+ ],
+ "score": 1.0,
+ "content": "ICL [3, 4] aims to induce a model to perform a task by feeding in concatenated and prompted",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 108,
+ 505,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 108,
+ 505,
+ 120
+ ],
+ "score": 1.0,
+ "content": "input-target examples (called “shots”) along with an unlabeled query example. Taking the cycled",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 506,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 506,
+ 131
+ ],
+ "score": 1.0,
+ "content": "letter task from Brown et al. [4] as an example, a 4-shot input or context would be “Please",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 495,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 444,
+ 142
+ ],
+ "score": 1.0,
+ "content": "unscramble the letters into a word, and write that word: asinoc",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 131,
+ 454,
+ 139
+ ],
+ "score": 0.79,
+ "content": "=",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 128,
+ 495,
+ 142
+ ],
+ "score": 1.0,
+ "content": "casino,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 140,
+ 505,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 141,
+ 153
+ ],
+ "score": 1.0,
+ "content": "yfrogg",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 142,
+ 151,
+ 150
+ ],
+ "score": 0.77,
+ "content": "=",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 140,
+ 230,
+ 153
+ ],
+ "score": 1.0,
+ "content": "froggy, plesim",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 141,
+ 240,
+ 150
+ ],
+ "score": 0.74,
+ "content": "=",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 240,
+ 140,
+ 324,
+ 153
+ ],
+ "score": 1.0,
+ "content": "simple, iggestb",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 142,
+ 334,
+ 150
+ ],
+ "score": 0.78,
+ "content": "=",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 140,
+ 424,
+ 153
+ ],
+ "score": 1.0,
+ "content": "biggest, astedro",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 141,
+ 436,
+ 150
+ ],
+ "score": 0.66,
+ "content": "= ^ { \\ast }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 437,
+ 140,
+ 505,
+ 153
+ ],
+ "score": 1.0,
+ "content": ", for which the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 151,
+ 505,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 151,
+ 505,
+ 164
+ ],
+ "score": 1.0,
+ "content": "desired output would be “roasted”. ICL induces an autoregressive language model to perform",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 162,
+ 505,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 505,
+ 174
+ ],
+ "score": 1.0,
+ "content": "this task by feeding in the context and sampling from the model. For classification tasks, each",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "score": 1.0,
+ "content": "label is associated with a string (e.g. “positive” and “negative” for sentiment analysis) and",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 182,
+ 507,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 182,
+ 507,
+ 197
+ ],
+ "score": 1.0,
+ "content": "a label is assigned by choosing the label string that the model assigns the highest probability to.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 194,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 194,
+ 310,
+ 208
+ ],
+ "score": 1.0,
+ "content": "For multiple-choice tasks (e.g. choosing between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 310,
+ 195,
+ 321,
+ 204
+ ],
+ "score": 0.77,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 194,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "possible answers to a question), the model’s",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 205,
+ 503,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 503,
+ 219
+ ],
+ "score": 1.0,
+ "content": "prediction is similarly determined by determining which choice is assigned the highest probability.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 222,
+ 505,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 222,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 222,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "The primary advantage of ICL is that it enables a single model to perform many tasks immediately",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "without fine-tuning. This also enables mixed-task batches, where different examples in a batch of data",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 244,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 244,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "correspond to different tasks by using different contexts in the input. ICL is also typically performed",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 506,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 506,
+ 267
+ ],
+ "score": 1.0,
+ "content": "with only a limited number of labeled examples – called few-shot learning – making it data-efficient.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 271,
+ 505,
+ 391
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Despite these advantages, ICL comes with significant practical drawbacks: First, making a prediction",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 282,
+ 505,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 282,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": "is dramatically more expensive because the model needs to process all of the in-context labeled",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "examples. Specifically, ignoring the quadratic complexity of self-attention operations in Transformer",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 304,
+ 506,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 304,
+ 506,
+ 317
+ ],
+ "score": 1.0,
+ "content": "language models (which are typically small compared to the costs of the rest of the model [20]),",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 165,
+ 328
+ ],
+ "score": 1.0,
+ "content": "processing the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 315,
+ 172,
+ 325
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 315,
+ 259,
+ 328
+ ],
+ "score": 1.0,
+ "content": "training examples for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 315,
+ 266,
+ 325
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 266,
+ 315,
+ 506,
+ 328
+ ],
+ "score": 1.0,
+ "content": "-shot ICL increases the computational cost by approximately",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 107,
+ 326,
+ 505,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 326,
+ 131,
+ 336
+ ],
+ "score": 0.89,
+ "content": "k + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 326,
+ 505,
+ 339
+ ],
+ "score": 1.0,
+ "content": "times compared to processing the unlabeled example alone. Memory costs similarly scale",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 218,
+ 350
+ ],
+ "score": 1.0,
+ "content": "approximately linearly with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 337,
+ 225,
+ 347
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 336,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": ", though during inference the memory costs are typically dominated by",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 360
+ ],
+ "score": 1.0,
+ "content": "storing the model’s parameters. Separately, there is a small amount of on-disk storage required for",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": "storing the in-context examples for a given task. For example, storing 32 examples for a task where",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 505,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 505,
+ 382
+ ],
+ "score": 1.0,
+ "content": "the prompted input and target for each example is 512 tokens long would require about 66 kilobytes",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 380,
+ 338,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 235,
+ 393
+ ],
+ "score": 1.0,
+ "content": "of storage on disk (32 examples",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 380,
+ 263,
+ 391
+ ],
+ "score": 0.82,
+ "content": "\\times \\ 5 1 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 380,
+ 292,
+ 393
+ ],
+ "score": 1.0,
+ "content": "tokens",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 292,
+ 380,
+ 314,
+ 391
+ ],
+ "score": 0.82,
+ "content": "\\times \\ 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 380,
+ 338,
+ 393
+ ],
+ "score": 1.0,
+ "content": "bits).",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 396,
+ 505,
+ 440
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 396,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 396,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "Beyond the aforementioned costs, ICL also exhibits unintuitive behavior. Zhao et al. [12] showed",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "that the ordering of examples in the context heavily influences the model’s predictions. Min et al. [9]",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 418,
+ 505,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 418,
+ 505,
+ 431
+ ],
+ "score": 1.0,
+ "content": "showed that ICL can still perform well even if the in-context example labels are swapped (i.e. made",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 507,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 507,
+ 442
+ ],
+ "score": 1.0,
+ "content": "incorrect), which raises questions about whether ICL is really “learning” from the labeled examples.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 446,
+ 505,
+ 577
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 459
+ ],
+ "score": 1.0,
+ "content": "Various approaches have been proposed to mitigate these issues. One way to decrease computational",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": "costs is to cache the key and value vectors for in-context examples. This is possible because decoder-",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 479
+ ],
+ "score": 1.0,
+ "content": "only Transformer language models have a causal masking pattern, so the model’s activations for the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 490
+ ],
+ "score": 1.0,
+ "content": "context do not do not depend on the unlabeled example. In an extreme case, 32-shot ICL with 512",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 501
+ ],
+ "score": 1.0,
+ "content": "tokens per in-context example would result in over 144 gigabytes of cached key and value vectors for",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 230,
+ 513
+ ],
+ "score": 1.0,
+ "content": "the GPT-3 model (32 examples",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 233,
+ 500,
+ 423,
+ 511
+ ],
+ "score": 0.27,
+ "content": "\\times \\ 5 1 2 \\ \\mathrm { t o k e n s \\times 9 6 \\ l a y e r s \\times 1 2 2 8 8 \\mathrm { d _ { m o d e l } \\times 3 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 423,
+ 500,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "bits each for the key",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 374,
+ 524
+ ],
+ "score": 1.0,
+ "content": "and value vectors). Separately, Min et al. [21] proposed ensemble",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 512,
+ 392,
+ 522
+ ],
+ "score": 0.4,
+ "content": "I C L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": ", where instead of using the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 274,
+ 534
+ ],
+ "score": 1.0,
+ "content": "output probability from concatenating the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 274,
+ 523,
+ 281,
+ 532
+ ],
+ "score": 0.8,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 522,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "training examples, the output probabilities of the model",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 335,
+ 545
+ ],
+ "score": 1.0,
+ "content": "on each training example (i.e. 1-shot ICL for each of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 336,
+ 533,
+ 343,
+ 543
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 532,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "examples) are multiplied together. This",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 542,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 542,
+ 325,
+ 557
+ ],
+ "score": 1.0,
+ "content": "lowers the non-parameter memory cost by a factor of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 544,
+ 342,
+ 555
+ ],
+ "score": 0.88,
+ "content": "k / 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 542,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "but increases the computational cost by",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "a factor of 2. In terms of task performance, Min et al. [21] find that ensemble ICL outperforms the",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 566,
+ 231,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 231,
+ 577
+ ],
+ "score": 1.0,
+ "content": "standard concatenative variant.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 599,
+ 261,
+ 611
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 263,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 263,
+ 613
+ ],
+ "score": 1.0,
+ "content": "2.2 Parameter-efficient fine-tuning",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 623,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 636
+ ],
+ "score": 1.0,
+ "content": "While standard fine-tuning updates all parameters of the pre-trained model, it has been demonstrated",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 635,
+ 505,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 635,
+ 505,
+ 646
+ ],
+ "score": 1.0,
+ "content": "that it is possible to instead update or add a relatively small number of parameters. Early methods",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "proposed adding adapters [22–24], which are small trainable feed-forward networks inserted between",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 656,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "the layers in the fixed pre-trained model. Since then, various sophisticated PEFT methods have been",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": "proposed, including methods that choose a sparse subset of parameters to train [25, 26], produce",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "low-rank updates [13], perform optimization in a lower-dimensional subspace [27], add low-rank",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 689,
+ 504,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 504,
+ 701
+ ],
+ "score": 1.0,
+ "content": "adapters using hypercomplex multiplication [28], and more. Relatedly, prompt tuning [14] and prefix",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "tuning [29] concatenate learned continuous embeddings to the model’s input or activations to induce",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "it to perform a task; this can be seen as a PEFT method [30]. State-of-the-art PEFT methods can",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 73,
+ 278,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 70,
+ 280,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 70,
+ 280,
+ 87
+ ],
+ "score": 1.0,
+ "content": "2.1 Few-shot in-context learning (ICL)",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 96,
+ 505,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 505,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 505,
+ 109
+ ],
+ "score": 1.0,
+ "content": "ICL [3, 4] aims to induce a model to perform a task by feeding in concatenated and prompted",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 108,
+ 505,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 108,
+ 505,
+ 120
+ ],
+ "score": 1.0,
+ "content": "input-target examples (called “shots”) along with an unlabeled query example. Taking the cycled",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 506,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 506,
+ 131
+ ],
+ "score": 1.0,
+ "content": "letter task from Brown et al. [4] as an example, a 4-shot input or context would be “Please",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 495,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 444,
+ 142
+ ],
+ "score": 1.0,
+ "content": "unscramble the letters into a word, and write that word: asinoc",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 131,
+ 454,
+ 139
+ ],
+ "score": 0.79,
+ "content": "=",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 128,
+ 495,
+ 142
+ ],
+ "score": 1.0,
+ "content": "casino,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 140,
+ 505,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 141,
+ 153
+ ],
+ "score": 1.0,
+ "content": "yfrogg",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 142,
+ 151,
+ 150
+ ],
+ "score": 0.77,
+ "content": "=",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 140,
+ 230,
+ 153
+ ],
+ "score": 1.0,
+ "content": "froggy, plesim",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 141,
+ 240,
+ 150
+ ],
+ "score": 0.74,
+ "content": "=",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 240,
+ 140,
+ 324,
+ 153
+ ],
+ "score": 1.0,
+ "content": "simple, iggestb",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 142,
+ 334,
+ 150
+ ],
+ "score": 0.78,
+ "content": "=",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 140,
+ 424,
+ 153
+ ],
+ "score": 1.0,
+ "content": "biggest, astedro",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 141,
+ 436,
+ 150
+ ],
+ "score": 0.66,
+ "content": "= ^ { \\ast }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 437,
+ 140,
+ 505,
+ 153
+ ],
+ "score": 1.0,
+ "content": ", for which the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 151,
+ 505,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 151,
+ 505,
+ 164
+ ],
+ "score": 1.0,
+ "content": "desired output would be “roasted”. ICL induces an autoregressive language model to perform",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 162,
+ 505,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 505,
+ 174
+ ],
+ "score": 1.0,
+ "content": "this task by feeding in the context and sampling from the model. For classification tasks, each",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "score": 1.0,
+ "content": "label is associated with a string (e.g. “positive” and “negative” for sentiment analysis) and",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 182,
+ 507,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 182,
+ 507,
+ 197
+ ],
+ "score": 1.0,
+ "content": "a label is assigned by choosing the label string that the model assigns the highest probability to.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 194,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 194,
+ 310,
+ 208
+ ],
+ "score": 1.0,
+ "content": "For multiple-choice tasks (e.g. choosing between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 310,
+ 195,
+ 321,
+ 204
+ ],
+ "score": 0.77,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 194,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "possible answers to a question), the model’s",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 205,
+ 503,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 503,
+ 219
+ ],
+ "score": 1.0,
+ "content": "prediction is similarly determined by determining which choice is assigned the highest probability.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 96,
+ 507,
+ 219
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 222,
+ 505,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 222,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 222,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "The primary advantage of ICL is that it enables a single model to perform many tasks immediately",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "without fine-tuning. This also enables mixed-task batches, where different examples in a batch of data",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 244,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 244,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "correspond to different tasks by using different contexts in the input. ICL is also typically performed",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 506,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 506,
+ 267
+ ],
+ "score": 1.0,
+ "content": "with only a limited number of labeled examples – called few-shot learning – making it data-efficient.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 105,
+ 222,
+ 506,
+ 267
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 271,
+ 505,
+ 391
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Despite these advantages, ICL comes with significant practical drawbacks: First, making a prediction",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 282,
+ 505,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 282,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": "is dramatically more expensive because the model needs to process all of the in-context labeled",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "examples. Specifically, ignoring the quadratic complexity of self-attention operations in Transformer",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 304,
+ 506,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 304,
+ 506,
+ 317
+ ],
+ "score": 1.0,
+ "content": "language models (which are typically small compared to the costs of the rest of the model [20]),",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 165,
+ 328
+ ],
+ "score": 1.0,
+ "content": "processing the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 315,
+ 172,
+ 325
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 315,
+ 259,
+ 328
+ ],
+ "score": 1.0,
+ "content": "training examples for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 315,
+ 266,
+ 325
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 266,
+ 315,
+ 506,
+ 328
+ ],
+ "score": 1.0,
+ "content": "-shot ICL increases the computational cost by approximately",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 107,
+ 326,
+ 505,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 326,
+ 131,
+ 336
+ ],
+ "score": 0.89,
+ "content": "k + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 326,
+ 505,
+ 339
+ ],
+ "score": 1.0,
+ "content": "times compared to processing the unlabeled example alone. Memory costs similarly scale",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 218,
+ 350
+ ],
+ "score": 1.0,
+ "content": "approximately linearly with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 337,
+ 225,
+ 347
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 336,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": ", though during inference the memory costs are typically dominated by",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 360
+ ],
+ "score": 1.0,
+ "content": "storing the model’s parameters. Separately, there is a small amount of on-disk storage required for",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": "storing the in-context examples for a given task. For example, storing 32 examples for a task where",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 505,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 505,
+ 382
+ ],
+ "score": 1.0,
+ "content": "the prompted input and target for each example is 512 tokens long would require about 66 kilobytes",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 380,
+ 338,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 235,
+ 393
+ ],
+ "score": 1.0,
+ "content": "of storage on disk (32 examples",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 380,
+ 263,
+ 391
+ ],
+ "score": 0.82,
+ "content": "\\times \\ 5 1 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 380,
+ 292,
+ 393
+ ],
+ "score": 1.0,
+ "content": "tokens",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 292,
+ 380,
+ 314,
+ 391
+ ],
+ "score": 0.82,
+ "content": "\\times \\ 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 380,
+ 338,
+ 393
+ ],
+ "score": 1.0,
+ "content": "bits).",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 105,
+ 271,
+ 506,
+ 393
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 396,
+ 505,
+ 440
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 396,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 396,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "Beyond the aforementioned costs, ICL also exhibits unintuitive behavior. Zhao et al. [12] showed",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "that the ordering of examples in the context heavily influences the model’s predictions. Min et al. [9]",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 418,
+ 505,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 418,
+ 505,
+ 431
+ ],
+ "score": 1.0,
+ "content": "showed that ICL can still perform well even if the in-context example labels are swapped (i.e. made",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 507,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 507,
+ 442
+ ],
+ "score": 1.0,
+ "content": "incorrect), which raises questions about whether ICL is really “learning” from the labeled examples.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 396,
+ 507,
+ 442
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 446,
+ 505,
+ 577
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 459
+ ],
+ "score": 1.0,
+ "content": "Various approaches have been proposed to mitigate these issues. One way to decrease computational",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": "costs is to cache the key and value vectors for in-context examples. This is possible because decoder-",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 479
+ ],
+ "score": 1.0,
+ "content": "only Transformer language models have a causal masking pattern, so the model’s activations for the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 490
+ ],
+ "score": 1.0,
+ "content": "context do not do not depend on the unlabeled example. In an extreme case, 32-shot ICL with 512",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 501
+ ],
+ "score": 1.0,
+ "content": "tokens per in-context example would result in over 144 gigabytes of cached key and value vectors for",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 230,
+ 513
+ ],
+ "score": 1.0,
+ "content": "the GPT-3 model (32 examples",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 233,
+ 500,
+ 423,
+ 511
+ ],
+ "score": 0.27,
+ "content": "\\times \\ 5 1 2 \\ \\mathrm { t o k e n s \\times 9 6 \\ l a y e r s \\times 1 2 2 8 8 \\mathrm { d _ { m o d e l } \\times 3 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 423,
+ 500,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "bits each for the key",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 374,
+ 524
+ ],
+ "score": 1.0,
+ "content": "and value vectors). Separately, Min et al. [21] proposed ensemble",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 512,
+ 392,
+ 522
+ ],
+ "score": 0.4,
+ "content": "I C L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": ", where instead of using the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 274,
+ 534
+ ],
+ "score": 1.0,
+ "content": "output probability from concatenating the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 274,
+ 523,
+ 281,
+ 532
+ ],
+ "score": 0.8,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 522,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "training examples, the output probabilities of the model",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 335,
+ 545
+ ],
+ "score": 1.0,
+ "content": "on each training example (i.e. 1-shot ICL for each of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 336,
+ 533,
+ 343,
+ 543
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 532,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "examples) are multiplied together. This",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 542,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 542,
+ 325,
+ 557
+ ],
+ "score": 1.0,
+ "content": "lowers the non-parameter memory cost by a factor of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 544,
+ 342,
+ 555
+ ],
+ "score": 0.88,
+ "content": "k / 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 542,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "but increases the computational cost by",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "a factor of 2. In terms of task performance, Min et al. [21] find that ensemble ICL outperforms the",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 566,
+ 231,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 231,
+ 577
+ ],
+ "score": 1.0,
+ "content": "standard concatenative variant.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 104,
+ 446,
+ 506,
+ 577
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 599,
+ 261,
+ 611
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 263,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 263,
+ 613
+ ],
+ "score": 1.0,
+ "content": "2.2 Parameter-efficient fine-tuning",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 623,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 636
+ ],
+ "score": 1.0,
+ "content": "While standard fine-tuning updates all parameters of the pre-trained model, it has been demonstrated",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 635,
+ 505,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 635,
+ 505,
+ 646
+ ],
+ "score": 1.0,
+ "content": "that it is possible to instead update or add a relatively small number of parameters. Early methods",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "proposed adding adapters [22–24], which are small trainable feed-forward networks inserted between",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 656,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "the layers in the fixed pre-trained model. Since then, various sophisticated PEFT methods have been",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": "proposed, including methods that choose a sparse subset of parameters to train [25, 26], produce",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "low-rank updates [13], perform optimization in a lower-dimensional subspace [27], add low-rank",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 689,
+ 504,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 504,
+ 701
+ ],
+ "score": 1.0,
+ "content": "adapters using hypercomplex multiplication [28], and more. Relatedly, prompt tuning [14] and prefix",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "tuning [29] concatenate learned continuous embeddings to the model’s input or activations to induce",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "it to perform a task; this can be seen as a PEFT method [30]. State-of-the-art PEFT methods can",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 72,
+ 504,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 504,
+ 86
+ ],
+ "score": 1.0,
+ "content": "match the performance of fine-tuning all of the model’s parameters while updating only a tiny fraction",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 84,
+ 266,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 84,
+ 126,
+ 96
+ ],
+ "score": 1.0,
+ "content": "(e.g.",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 126,
+ 84,
+ 154,
+ 95
+ ],
+ "score": 0.82,
+ "content": "0 . 0 1 \\%",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 154,
+ 84,
+ 266,
+ 96
+ ],
+ "score": 1.0,
+ "content": ") of the model’s parameters.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 48,
+ "bbox_fs": [
+ 105,
+ 622,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 73,
+ 504,
+ 95
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 504,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 504,
+ 86
+ ],
+ "score": 1.0,
+ "content": "match the performance of fine-tuning all of the model’s parameters while updating only a tiny fraction",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 84,
+ 266,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 84,
+ 126,
+ 96
+ ],
+ "score": 1.0,
+ "content": "(e.g.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 126,
+ 84,
+ 154,
+ 95
+ ],
+ "score": 0.82,
+ "content": "0 . 0 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 154,
+ 84,
+ 266,
+ 96
+ ],
+ "score": 1.0,
+ "content": ") of the model’s parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 100,
+ 505,
+ 221
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 113
+ ],
+ "score": 1.0,
+ "content": "PEFT drastically reduces the memory and storage requirements for training and saving the model. In",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "addition, certain PEFT methods straightforwardly allow mixed-task batches – for example, prompt",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 135
+ ],
+ "score": 1.0,
+ "content": "tuning enables a single model to perform many tasks simply by concatenating different prompt",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "embeddings to each example in the batch [14]. On the other hand, PEFT methods that re-parameterize",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "the model (e.g. [27, 13]) are costly or onerous for mixed-task batches. Separately, different PEFT",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 507,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 507,
+ 167
+ ],
+ "score": 1.0,
+ "content": "methods increase the computation and memory required to perform inference by different amounts.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "For example, adapters effectively add additional (small) layers to the model, resulting in small but",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 188
+ ],
+ "score": 1.0,
+ "content": "non-negligible increases in computational costs and memory. An additional cost incurred by PEFT",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 186,
+ 506,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 186,
+ 506,
+ 199
+ ],
+ "score": 1.0,
+ "content": "is the cost of fine-tuning itself, which must be performed once and is then amortized as the model",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "is used for inference. However, we will show that PEFT can be dramatically more computationally",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "efficient when considering both fine-tuning and inference while achieving better accuracy than ICL.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 236,
+ 267,
+ 250
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 234,
+ 269,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 234,
+ 269,
+ 254
+ ],
+ "score": 1.0,
+ "content": "3 Designing the T-Few Recipe",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 261,
+ 505,
+ 349
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 505,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 505,
+ 274
+ ],
+ "score": 1.0,
+ "content": "Given that PEFT allows a model to be adapted to a new task with relatively small storage requirements",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 273,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "and computational cost, we argue that PEFT presents a promising alternative to ICL. Our goal",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "is therefore to develop a recipe that allows a model to attain high accuracy on new tasks with",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 295,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "limited labeled examples while allowing mixed-task batches during inference and incurring minimal",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 304,
+ 505,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 304,
+ 505,
+ 319
+ ],
+ "score": 1.0,
+ "content": "computational and storage costs. By recipe, we mean a specific model and hyperparameter setting",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "that provides strong performance on any new task without manual tuning or per-task adjustments.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 506,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 506,
+ 340
+ ],
+ "score": 1.0,
+ "content": "In this way, we can ensure that our approach is a realistic option in few-shot settings where limited",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 338,
+ 298,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 298,
+ 350
+ ],
+ "score": 1.0,
+ "content": "labeled data is available for evaluation [31, 32].",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 363,
+ 214,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 363,
+ 216,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 363,
+ 216,
+ 376
+ ],
+ "score": 1.0,
+ "content": "3.1 Model and Datasets",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 384,
+ 505,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "As a first step, we must choose a pre-trained model. Ideally, the model should attain high performance",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 407
+ ],
+ "score": 1.0,
+ "content": "on new tasks after fine-tuning on a limited number of labeled examples. In preliminary experiments",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 506,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 506,
+ 417
+ ],
+ "score": 1.0,
+ "content": "applying PEFT methods to different pre-trained models, we attained the best performance with T0",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "[1]. T0 is based on T5 [15], an encoder-decoder Transformer model [33] that was pre-trained via a",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "masked language modeling objective [34] on a large corpus of unlabeled text data. T0 was created by",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 450
+ ],
+ "score": 1.0,
+ "content": "fine-tuning T5 on a multitask mixture of datasets in order to enable zero-shot generalization, i.e. the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 450,
+ 505,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 505,
+ 461
+ ],
+ "score": 1.0,
+ "content": "ability to perform tasks without any additional gradient-based training. Examples in the datasets used",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 472
+ ],
+ "score": 1.0,
+ "content": "to train T0 were prompted by applying the prompt templates from the Public Pool of Prompts (P3",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 483
+ ],
+ "score": 1.0,
+ "content": "[35]), which convert each example in each dataset to a prompted text-to-text format where each label",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": "corresponds to a different string. For brevity, we omit a detailed description of T0 and T5; interested",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 505
+ ],
+ "score": 1.0,
+ "content": "readers can refer to Sanh et al. [1] and Raffel et al. [15]. T0 was released in three billion and eleven",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "billion parameter variants, referred to as “T0-3B” and simply “T0” respectively. In this section (where",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 514,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "our goal is to design the T-Few recipe through extensive experimentation), we use T0-3B to reduce",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 488,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 488,
+ 538
+ ],
+ "score": 1.0,
+ "content": "computational costs. For all models and experiments, we use Hugging Face Transformers [36].",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 673
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 506,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 506,
+ 555
+ ],
+ "score": 1.0,
+ "content": "While T0 was designed for zero-shot generalization, we will demonstrate that it also attains strong",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 553,
+ 506,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 506,
+ 565
+ ],
+ "score": 1.0,
+ "content": "performance after fine-tuning with only a few labeled examples. To test T0’s generalization, Sanh et al.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "[1] chose a set of tasks (and corresponding datasets) to hold out from the multitask training mixture",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 587
+ ],
+ "score": 1.0,
+ "content": "– specifically, sentence completion (COPA [37], H-SWAG [38], and Story Cloze [39] datasets),",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "natural language inference (ANLI [40], CB [41], and RTE [42]), coreference resolution (WSC [43]",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 608
+ ],
+ "score": 1.0,
+ "content": "and Winogrande [44]), and word sense disambiguation (WiC [45]). Evaluation of generalization",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 608,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "capabilities can then be straightforwardly done by measuring performance on these held-out datasets.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 630
+ ],
+ "score": 1.0,
+ "content": "We also will later test T-Few’s abilities in the RAFT benchmark [2] in section 4.3, a collection of",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 629,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 629,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "unseen “real-world” few-shot tasks with no validation set and a held-out test set. ANLI, WiC, WSC",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 652
+ ],
+ "score": 1.0,
+ "content": "is licensed under a Creative Commons License. Winogrande is licensed under an Apache license.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "COPA is under a BSD-2 Clause license. We could not find the license of RTE and CB but they are",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 461,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 461,
+ 674
+ ],
+ "score": 1.0,
+ "content": "part of SuperGLUE which mentions the datasets are allowed for use in research context.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 678,
+ 504,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "To ease comparison, we use the same number of few-shot training examples for each dataset as Brown",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "et al. [4], which varies from 20 to 70. Unfortunately, the few-shot dataset subsets used by Brown",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "et al. [4] have not been publicly disclosed. To allow for a more robust comparison, we therefore",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "constructed five few-shot datasets by sampling subsets with different seeds and report the median and",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 50.5
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 11,
+ "width": 9
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 73,
+ 504,
+ 95
+ ],
+ "lines": [],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 72,
+ 504,
+ 96
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 100,
+ 505,
+ 221
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 113
+ ],
+ "score": 1.0,
+ "content": "PEFT drastically reduces the memory and storage requirements for training and saving the model. In",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "addition, certain PEFT methods straightforwardly allow mixed-task batches – for example, prompt",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 135
+ ],
+ "score": 1.0,
+ "content": "tuning enables a single model to perform many tasks simply by concatenating different prompt",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "embeddings to each example in the batch [14]. On the other hand, PEFT methods that re-parameterize",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "the model (e.g. [27, 13]) are costly or onerous for mixed-task batches. Separately, different PEFT",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 507,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 507,
+ 167
+ ],
+ "score": 1.0,
+ "content": "methods increase the computation and memory required to perform inference by different amounts.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "For example, adapters effectively add additional (small) layers to the model, resulting in small but",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 188
+ ],
+ "score": 1.0,
+ "content": "non-negligible increases in computational costs and memory. An additional cost incurred by PEFT",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 186,
+ 506,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 186,
+ 506,
+ 199
+ ],
+ "score": 1.0,
+ "content": "is the cost of fine-tuning itself, which must be performed once and is then amortized as the model",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "is used for inference. However, we will show that PEFT can be dramatically more computationally",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "efficient when considering both fine-tuning and inference while achieving better accuracy than ICL.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 105,
+ 100,
+ 507,
+ 222
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 236,
+ 267,
+ 250
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 234,
+ 269,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 234,
+ 269,
+ 254
+ ],
+ "score": 1.0,
+ "content": "3 Designing the T-Few Recipe",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 261,
+ 505,
+ 349
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 505,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 505,
+ 274
+ ],
+ "score": 1.0,
+ "content": "Given that PEFT allows a model to be adapted to a new task with relatively small storage requirements",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 273,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "and computational cost, we argue that PEFT presents a promising alternative to ICL. Our goal",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "is therefore to develop a recipe that allows a model to attain high accuracy on new tasks with",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 295,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "limited labeled examples while allowing mixed-task batches during inference and incurring minimal",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 304,
+ 505,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 304,
+ 505,
+ 319
+ ],
+ "score": 1.0,
+ "content": "computational and storage costs. By recipe, we mean a specific model and hyperparameter setting",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "that provides strong performance on any new task without manual tuning or per-task adjustments.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 506,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 506,
+ 340
+ ],
+ "score": 1.0,
+ "content": "In this way, we can ensure that our approach is a realistic option in few-shot settings where limited",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 338,
+ 298,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 298,
+ 350
+ ],
+ "score": 1.0,
+ "content": "labeled data is available for evaluation [31, 32].",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 261,
+ 506,
+ 350
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 363,
+ 214,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 363,
+ 216,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 363,
+ 216,
+ 376
+ ],
+ "score": 1.0,
+ "content": "3.1 Model and Datasets",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 384,
+ 505,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "As a first step, we must choose a pre-trained model. Ideally, the model should attain high performance",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 407
+ ],
+ "score": 1.0,
+ "content": "on new tasks after fine-tuning on a limited number of labeled examples. In preliminary experiments",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 506,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 506,
+ 417
+ ],
+ "score": 1.0,
+ "content": "applying PEFT methods to different pre-trained models, we attained the best performance with T0",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "[1]. T0 is based on T5 [15], an encoder-decoder Transformer model [33] that was pre-trained via a",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "masked language modeling objective [34] on a large corpus of unlabeled text data. T0 was created by",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 450
+ ],
+ "score": 1.0,
+ "content": "fine-tuning T5 on a multitask mixture of datasets in order to enable zero-shot generalization, i.e. the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 450,
+ 505,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 505,
+ 461
+ ],
+ "score": 1.0,
+ "content": "ability to perform tasks without any additional gradient-based training. Examples in the datasets used",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 472
+ ],
+ "score": 1.0,
+ "content": "to train T0 were prompted by applying the prompt templates from the Public Pool of Prompts (P3",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 483
+ ],
+ "score": 1.0,
+ "content": "[35]), which convert each example in each dataset to a prompted text-to-text format where each label",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": "corresponds to a different string. For brevity, we omit a detailed description of T0 and T5; interested",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 505
+ ],
+ "score": 1.0,
+ "content": "readers can refer to Sanh et al. [1] and Raffel et al. [15]. T0 was released in three billion and eleven",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "billion parameter variants, referred to as “T0-3B” and simply “T0” respectively. In this section (where",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 514,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "our goal is to design the T-Few recipe through extensive experimentation), we use T0-3B to reduce",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 488,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 488,
+ 538
+ ],
+ "score": 1.0,
+ "content": "computational costs. For all models and experiments, we use Hugging Face Transformers [36].",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 383,
+ 506,
+ 538
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 673
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 506,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 506,
+ 555
+ ],
+ "score": 1.0,
+ "content": "While T0 was designed for zero-shot generalization, we will demonstrate that it also attains strong",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 553,
+ 506,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 506,
+ 565
+ ],
+ "score": 1.0,
+ "content": "performance after fine-tuning with only a few labeled examples. To test T0’s generalization, Sanh et al.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "[1] chose a set of tasks (and corresponding datasets) to hold out from the multitask training mixture",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 587
+ ],
+ "score": 1.0,
+ "content": "– specifically, sentence completion (COPA [37], H-SWAG [38], and Story Cloze [39] datasets),",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "natural language inference (ANLI [40], CB [41], and RTE [42]), coreference resolution (WSC [43]",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 608
+ ],
+ "score": 1.0,
+ "content": "and Winogrande [44]), and word sense disambiguation (WiC [45]). Evaluation of generalization",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 608,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "capabilities can then be straightforwardly done by measuring performance on these held-out datasets.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 630
+ ],
+ "score": 1.0,
+ "content": "We also will later test T-Few’s abilities in the RAFT benchmark [2] in section 4.3, a collection of",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 629,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 629,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "unseen “real-world” few-shot tasks with no validation set and a held-out test set. ANLI, WiC, WSC",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 652
+ ],
+ "score": 1.0,
+ "content": "is licensed under a Creative Commons License. Winogrande is licensed under an Apache license.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "COPA is under a BSD-2 Clause license. We could not find the license of RTE and CB but they are",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 461,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 461,
+ 674
+ ],
+ "score": 1.0,
+ "content": "part of SuperGLUE which mentions the datasets are allowed for use in research context.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 105,
+ 540,
+ 506,
+ 674
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 678,
+ 504,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "To ease comparison, we use the same number of few-shot training examples for each dataset as Brown",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "et al. [4], which varies from 20 to 70. Unfortunately, the few-shot dataset subsets used by Brown",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "et al. [4] have not been publicly disclosed. To allow for a more robust comparison, we therefore",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "constructed five few-shot datasets by sampling subsets with different seeds and report the median and",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "interquartile range. We prompt examples using a randomly-sampled prompt template from P3 Bach",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 84,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "et al. [35] for each example at each step. Unless otherwise stated, we train our model for 1K steps",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 94,
+ 378,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 94,
+ 378,
+ 107
+ ],
+ "score": 1.0,
+ "content": "with a batch size of 8 and report performance at the end of training.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 50.5,
+ "bbox_fs": [
+ 105,
+ 678,
+ 505,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 72,
+ 505,
+ 106
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "interquartile range. We prompt examples using a randomly-sampled prompt template from P3 Bach",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 84,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "et al. [35] for each example at each step. Unless otherwise stated, we train our model for 1K steps",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 94,
+ 378,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 94,
+ 378,
+ 107
+ ],
+ "score": 1.0,
+ "content": "with a batch size of 8 and report performance at the end of training.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 110,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 110,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 110,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "For evaluation, we use “rank classification”, where the model’s log-probabilities for all possible label",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 133
+ ],
+ "score": 1.0,
+ "content": "strings are ranked and the model’s prediction is considered correct if the highest-ranked choice is the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 133,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 133,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "correct answer. Rank classification evaluation is compatible with both classification and multiple-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 144,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "choice tasks. Since model performance can vary significantly depending on the prompt template used,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 155,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 155,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": "we report the median accuracy across all prompt templates from P3 and across few-shot data subsets",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 178
+ ],
+ "score": 1.0,
+ "content": "for each dataset. For all datasets, we report the accuracy on the test set or validation set when the test",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "labels are not public (e.g. SuperGLUE datasets). In the main text, we report median accuracy across",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 504,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 504,
+ 200
+ ],
+ "score": 1.0,
+ "content": "the nine datasets mentioned above. Detailed results on each dataset are provided in the appendices.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 212,
+ 341,
+ 224
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 340,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 340,
+ 226
+ ],
+ "score": 1.0,
+ "content": "3.2 Unlikelihood Training and Length Normalization",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 277
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 244
+ ],
+ "score": 1.0,
+ "content": "Before investigating PEFT methods, we first explore two additional loss terms to improve the",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "performance of few-shot fine-tuning of language models. Language models are normally trained",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 202,
+ 267
+ ],
+ "score": 1.0,
+ "content": "with cross-entropy loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 253,
+ 333,
+ 267
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { L _ { \\mathrm { L M } } = - \\frac { 1 } { T } \\sum _ { t } \\log \\bar { p } ( y _ { t } | \\mathbf { x } , y _ { < t } ) } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 254,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "where the model is trained to increase the",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 470,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 272,
+ 279
+ ],
+ "score": 1.0,
+ "content": "probability of the correct target sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 266,
+ 357,
+ 277
+ ],
+ "score": 0.88,
+ "content": "\\mathbf { y } = ( y _ { 1 } , y _ { 2 } , \\dots y _ { T } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 265,
+ 459,
+ 279
+ ],
+ "score": 1.0,
+ "content": "given the input sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 267,
+ 467,
+ 275
+ ],
+ "score": 0.41,
+ "content": "\\mathbf { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 467,
+ 265,
+ 470,
+ 279
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 281,
+ 504,
+ 315
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "For evaluation, we use rank classification (described in section 3.1) which depends on both the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 292,
+ 506,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 292,
+ 506,
+ 304
+ ],
+ "score": 1.0,
+ "content": "probabilities that the model assigns to the correct choice as well as the incorrect choices. To account",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 303,
+ 352,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 352,
+ 315
+ ],
+ "score": 1.0,
+ "content": "for this during training, we add an unlikelihood loss [16, 17]:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 206,
+ 320,
+ 404,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 206,
+ 320,
+ 404,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 206,
+ 320,
+ 404,
+ 355
+ ],
+ "score": 0.93,
+ "content": "L _ { \\mathrm { U L } } = - \\frac { \\sum _ { n = 1 } ^ { N } \\sum _ { t = 1 } ^ { T ^ { ( n ) } } \\log ( 1 - p ( \\hat { y } _ { i } ^ { ( n ) } | \\mathbf { x } , \\hat { y } _ { < t } ^ { ( n ) } ) ) } { \\sum _ { n = 1 } ^ { N } T ^ { ( n ) } }",
+ "type": "interline_equation",
+ "image_path": "09597e6adfa17b86f3b9c5cf1f08d608571afbd01f372ddf3b09c30cd5f1d152.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 206,
+ 320,
+ 404,
+ 337.5
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 206,
+ 337.5,
+ 404,
+ 355.0
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 405
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 360,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 360,
+ 474,
+ 374
+ ],
+ "score": 1.0,
+ "content": "which discourages the model from predicting tokens from incorrect target sequences, where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 360,
+ 505,
+ 372
+ ],
+ "score": 0.89,
+ "content": "\\hat { \\mathbf { y } } ^ { ( n ) } =",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 181,
+ 384
+ ],
+ "score": 0.9,
+ "content": "\\left( \\hat { y } _ { 1 } , \\hat { y } _ { 2 } , \\dots , \\hat { y } _ { T ^ { ( n ) } } \\right)",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 372,
+ 205,
+ 385
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 206,
+ 374,
+ 213,
+ 382
+ ],
+ "score": 0.77,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 372,
+ 236,
+ 385
+ ],
+ "score": 1.0,
+ "content": "-th of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 372,
+ 246,
+ 382
+ ],
+ "score": 0.84,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 372,
+ 466,
+ 385
+ ],
+ "score": 1.0,
+ "content": "incorrect target sequences. We hypothesize that adding",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 467,
+ 372,
+ 486,
+ 383
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 372,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "will",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 395
+ ],
+ "score": 1.0,
+ "content": "improve results on rank classification because the model will be trained to assign lower probabilities",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 473,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 473,
+ 406
+ ],
+ "score": 1.0,
+ "content": "to incorrect choices, thereby improving the chance that the correct choice is ranked highest.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 423
+ ],
+ "score": 1.0,
+ "content": "The possible target sequences for a given training example can have significantly different lengths,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 507,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 507,
+ 434
+ ],
+ "score": 1.0,
+ "content": "especially in multiple-choice tasks. Ranking each choice based on probability can therefore “favor”",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 506,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 416,
+ 445
+ ],
+ "score": 1.0,
+ "content": "shorter choices because the model’s assigned probability to each token is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 433,
+ 435,
+ 443
+ ],
+ "score": 0.87,
+ "content": "\\leq 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 432,
+ 506,
+ 445
+ ],
+ "score": 1.0,
+ "content": ". To rectify this,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 443,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "we consider using length normalization when performing rank classification, which divides the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 454,
+ 506,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 506,
+ 466
+ ],
+ "score": 1.0,
+ "content": "model’s score on each possible answer choice by the number of tokens in the choice (as used in",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 465,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "GPT-3 [4]). When using length normalization during evaluation, we introduce an additional loss",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 476,
+ 505,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 476,
+ 505,
+ 487
+ ],
+ "score": 1.0,
+ "content": "term during training that more closely reflects length-normalized evaluation. First, we compute the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 103,
+ 483,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 483,
+ 358,
+ 506
+ ],
+ "score": 1.0,
+ "content": "length-normalized log probability of a given output sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 486,
+ 503,
+ 502
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { \\beta ( \\mathbf { x } , \\mathbf { y } ) = \\frac { 1 } { T } \\sum _ { t = 1 } ^ { T } \\log p ( y _ { t } | \\mathbf { \\bar { x } } , y _ { < t } ) } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 504,
+ 483,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 498,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 498,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "Then, we maximize the length-normalized log probability of the correct answer choice by minimizing",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 510,
+ 235,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 235,
+ 524
+ ],
+ "score": 1.0,
+ "content": "the softmax cross-entropy loss:",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 195,
+ 527,
+ 415,
+ 558
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 195,
+ 527,
+ 415,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 195,
+ 527,
+ 415,
+ 558
+ ],
+ "score": 0.94,
+ "content": "L _ { \\mathrm { L N } } = - \\log \\frac { \\exp ( \\beta ( \\mathbf { x } , \\mathbf { y } ) ) } { \\exp ( \\beta ( \\mathbf { x } , \\mathbf { y } ) ) + \\sum _ { n = 1 } ^ { N } \\exp ( \\beta ( \\mathbf { x } , \\hat { \\mathbf { y } } ^ { ( n ) } ) ) }",
+ "type": "interline_equation",
+ "image_path": "6bebee14ad649a0aab42ace70b0e9e373d5e77b5b6427bc54162cd106dc05588.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 35.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 195,
+ 527,
+ 415,
+ 542.5
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 195,
+ 542.5,
+ 415,
+ 558.0
+ ],
+ "spans": [],
+ "index": 36
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 562,
+ 504,
+ 596
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 218,
+ 576
+ ],
+ "score": 1.0,
+ "content": "When training a model with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 562,
+ 263,
+ 573
+ ],
+ "score": 0.29,
+ "content": "L _ { \\mathrm { L M } } , L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 560,
+ 283,
+ 576
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 283,
+ 562,
+ 302,
+ 573
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 560,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": ", we simply sum them. This avoids introducing any",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "hyperparameters that would be problematic to tune in the few-shot setting (where realistically-sized",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 584,
+ 289,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 584,
+ 289,
+ 596
+ ],
+ "score": 1.0,
+ "content": "validation sets are tiny by necessity [31, 32]).",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 600,
+ 505,
+ 655
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 613
+ ],
+ "score": 1.0,
+ "content": "We report the results of fine-tuning all of T0-3B’s parameters with and without length normalization",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 303,
+ 625
+ ],
+ "score": 1.0,
+ "content": "on all datasets in appendix B. We find that adding",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 611,
+ 323,
+ 622
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 609,
+ 435,
+ 625
+ ],
+ "score": 1.0,
+ "content": "improves the accuracy from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 435,
+ 611,
+ 462,
+ 622
+ ],
+ "score": 0.86,
+ "content": "6 0 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 609,
+ 473,
+ 625
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 611,
+ 505,
+ 622
+ ],
+ "score": 0.86,
+ "content": "6 2 . 7 1 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 184,
+ 635
+ ],
+ "score": 1.0,
+ "content": "and including both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 622,
+ 204,
+ 633
+ ],
+ "score": 0.91,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 622,
+ 222,
+ 635
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 622,
+ 241,
+ 633
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 241,
+ 622,
+ 382,
+ 635
+ ],
+ "score": 1.0,
+ "content": "provides a further improvement to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 622,
+ 410,
+ 633
+ ],
+ "score": 0.86,
+ "content": "6 3 . 3 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 410,
+ 622,
+ 506,
+ 635
+ ],
+ "score": 1.0,
+ "content": ". Since these loss terms",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "improve performance without introducing any additional hyperparameters, we include them in our",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 305,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 305,
+ 657
+ ],
+ "score": 1.0,
+ "content": "recipe and use them in all following experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 668,
+ 310,
+ 680
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 311,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 311,
+ 682
+ ],
+ "score": 1.0,
+ "content": "3.3 Parameter-efficient fine-tuning with (IA)3",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 45
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 688,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "In order to compare favorably to few-shot ICL, we need a PEFT method that has the following",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "properties: First, it must add or update as few parameters as possible to avoid incurring storage",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 507,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 507,
+ 723
+ ],
+ "score": 1.0,
+ "content": "and memory costs. Second, it should achieve strong accuracy after few-shot training on new tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 740,
+ 309,
+ 753
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 740,
+ 309,
+ 753
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 72,
+ 505,
+ 106
+ ],
+ "lines": [],
+ "index": 1,
+ "bbox_fs": [
+ 105,
+ 73,
+ 505,
+ 107
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 110,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 110,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 110,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "For evaluation, we use “rank classification”, where the model’s log-probabilities for all possible label",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 133
+ ],
+ "score": 1.0,
+ "content": "strings are ranked and the model’s prediction is considered correct if the highest-ranked choice is the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 133,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 133,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "correct answer. Rank classification evaluation is compatible with both classification and multiple-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 144,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "choice tasks. Since model performance can vary significantly depending on the prompt template used,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 155,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 155,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": "we report the median accuracy across all prompt templates from P3 and across few-shot data subsets",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 178
+ ],
+ "score": 1.0,
+ "content": "for each dataset. For all datasets, we report the accuracy on the test set or validation set when the test",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "labels are not public (e.g. SuperGLUE datasets). In the main text, we report median accuracy across",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 504,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 504,
+ 200
+ ],
+ "score": 1.0,
+ "content": "the nine datasets mentioned above. Detailed results on each dataset are provided in the appendices.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 110,
+ 506,
+ 200
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 212,
+ 341,
+ 224
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 340,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 340,
+ 226
+ ],
+ "score": 1.0,
+ "content": "3.2 Unlikelihood Training and Length Normalization",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 277
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 244
+ ],
+ "score": 1.0,
+ "content": "Before investigating PEFT methods, we first explore two additional loss terms to improve the",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "performance of few-shot fine-tuning of language models. Language models are normally trained",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 202,
+ 267
+ ],
+ "score": 1.0,
+ "content": "with cross-entropy loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 253,
+ 333,
+ 267
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { L _ { \\mathrm { L M } } = - \\frac { 1 } { T } \\sum _ { t } \\log \\bar { p } ( y _ { t } | \\mathbf { x } , y _ { < t } ) } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 254,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "where the model is trained to increase the",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 470,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 272,
+ 279
+ ],
+ "score": 1.0,
+ "content": "probability of the correct target sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 266,
+ 357,
+ 277
+ ],
+ "score": 0.88,
+ "content": "\\mathbf { y } = ( y _ { 1 } , y _ { 2 } , \\dots y _ { T } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 265,
+ 459,
+ 279
+ ],
+ "score": 1.0,
+ "content": "given the input sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 267,
+ 467,
+ 275
+ ],
+ "score": 0.41,
+ "content": "\\mathbf { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 467,
+ 265,
+ 470,
+ 279
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 105,
+ 232,
+ 505,
+ 279
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 281,
+ 504,
+ 315
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "For evaluation, we use rank classification (described in section 3.1) which depends on both the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 292,
+ 506,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 292,
+ 506,
+ 304
+ ],
+ "score": 1.0,
+ "content": "probabilities that the model assigns to the correct choice as well as the incorrect choices. To account",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 303,
+ 352,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 352,
+ 315
+ ],
+ "score": 1.0,
+ "content": "for this during training, we add an unlikelihood loss [16, 17]:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 105,
+ 280,
+ 506,
+ 315
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 206,
+ 320,
+ 404,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 206,
+ 320,
+ 404,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 206,
+ 320,
+ 404,
+ 355
+ ],
+ "score": 0.93,
+ "content": "L _ { \\mathrm { U L } } = - \\frac { \\sum _ { n = 1 } ^ { N } \\sum _ { t = 1 } ^ { T ^ { ( n ) } } \\log ( 1 - p ( \\hat { y } _ { i } ^ { ( n ) } | \\mathbf { x } , \\hat { y } _ { < t } ^ { ( n ) } ) ) } { \\sum _ { n = 1 } ^ { N } T ^ { ( n ) } }",
+ "type": "interline_equation",
+ "image_path": "09597e6adfa17b86f3b9c5cf1f08d608571afbd01f372ddf3b09c30cd5f1d152.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 206,
+ 320,
+ 404,
+ 337.5
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 206,
+ 337.5,
+ 404,
+ 355.0
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 405
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 360,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 360,
+ 474,
+ 374
+ ],
+ "score": 1.0,
+ "content": "which discourages the model from predicting tokens from incorrect target sequences, where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 360,
+ 505,
+ 372
+ ],
+ "score": 0.89,
+ "content": "\\hat { \\mathbf { y } } ^ { ( n ) } =",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 181,
+ 384
+ ],
+ "score": 0.9,
+ "content": "\\left( \\hat { y } _ { 1 } , \\hat { y } _ { 2 } , \\dots , \\hat { y } _ { T ^ { ( n ) } } \\right)",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 372,
+ 205,
+ 385
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 206,
+ 374,
+ 213,
+ 382
+ ],
+ "score": 0.77,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 372,
+ 236,
+ 385
+ ],
+ "score": 1.0,
+ "content": "-th of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 372,
+ 246,
+ 382
+ ],
+ "score": 0.84,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 372,
+ 466,
+ 385
+ ],
+ "score": 1.0,
+ "content": "incorrect target sequences. We hypothesize that adding",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 467,
+ 372,
+ 486,
+ 383
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 372,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "will",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 395
+ ],
+ "score": 1.0,
+ "content": "improve results on rank classification because the model will be trained to assign lower probabilities",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 473,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 473,
+ 406
+ ],
+ "score": 1.0,
+ "content": "to incorrect choices, thereby improving the chance that the correct choice is ranked highest.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 104,
+ 360,
+ 506,
+ 406
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 506,
+ 423
+ ],
+ "score": 1.0,
+ "content": "The possible target sequences for a given training example can have significantly different lengths,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 507,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 507,
+ 434
+ ],
+ "score": 1.0,
+ "content": "especially in multiple-choice tasks. Ranking each choice based on probability can therefore “favor”",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 506,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 416,
+ 445
+ ],
+ "score": 1.0,
+ "content": "shorter choices because the model’s assigned probability to each token is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 433,
+ 435,
+ 443
+ ],
+ "score": 0.87,
+ "content": "\\leq 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 432,
+ 506,
+ 445
+ ],
+ "score": 1.0,
+ "content": ". To rectify this,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 443,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "we consider using length normalization when performing rank classification, which divides the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 454,
+ 506,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 506,
+ 466
+ ],
+ "score": 1.0,
+ "content": "model’s score on each possible answer choice by the number of tokens in the choice (as used in",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 465,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "GPT-3 [4]). When using length normalization during evaluation, we introduce an additional loss",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 476,
+ 505,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 476,
+ 505,
+ 487
+ ],
+ "score": 1.0,
+ "content": "term during training that more closely reflects length-normalized evaluation. First, we compute the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 103,
+ 483,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 483,
+ 358,
+ 506
+ ],
+ "score": 1.0,
+ "content": "length-normalized log probability of a given output sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 486,
+ 503,
+ 502
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { \\beta ( \\mathbf { x } , \\mathbf { y } ) = \\frac { 1 } { T } \\sum _ { t = 1 } ^ { T } \\log p ( y _ { t } | \\mathbf { \\bar { x } } , y _ { < t } ) } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 504,
+ 483,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 498,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 498,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "Then, we maximize the length-normalized log probability of the correct answer choice by minimizing",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 510,
+ 235,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 235,
+ 524
+ ],
+ "score": 1.0,
+ "content": "the softmax cross-entropy loss:",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 103,
+ 410,
+ 507,
+ 524
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 195,
+ 527,
+ 415,
+ 558
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 195,
+ 527,
+ 415,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 195,
+ 527,
+ 415,
+ 558
+ ],
+ "score": 0.94,
+ "content": "L _ { \\mathrm { L N } } = - \\log \\frac { \\exp ( \\beta ( \\mathbf { x } , \\mathbf { y } ) ) } { \\exp ( \\beta ( \\mathbf { x } , \\mathbf { y } ) ) + \\sum _ { n = 1 } ^ { N } \\exp ( \\beta ( \\mathbf { x } , \\hat { \\mathbf { y } } ^ { ( n ) } ) ) }",
+ "type": "interline_equation",
+ "image_path": "6bebee14ad649a0aab42ace70b0e9e373d5e77b5b6427bc54162cd106dc05588.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 35.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 195,
+ 527,
+ 415,
+ 542.5
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 195,
+ 542.5,
+ 415,
+ 558.0
+ ],
+ "spans": [],
+ "index": 36
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 562,
+ 504,
+ 596
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 218,
+ 576
+ ],
+ "score": 1.0,
+ "content": "When training a model with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 562,
+ 263,
+ 573
+ ],
+ "score": 0.29,
+ "content": "L _ { \\mathrm { L M } } , L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 560,
+ 283,
+ 576
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 283,
+ 562,
+ 302,
+ 573
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 560,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": ", we simply sum them. This avoids introducing any",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "hyperparameters that would be problematic to tune in the few-shot setting (where realistically-sized",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 584,
+ 289,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 584,
+ 289,
+ 596
+ ],
+ "score": 1.0,
+ "content": "validation sets are tiny by necessity [31, 32]).",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 105,
+ 560,
+ 506,
+ 596
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 600,
+ 505,
+ 655
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 613
+ ],
+ "score": 1.0,
+ "content": "We report the results of fine-tuning all of T0-3B’s parameters with and without length normalization",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 303,
+ 625
+ ],
+ "score": 1.0,
+ "content": "on all datasets in appendix B. We find that adding",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 611,
+ 323,
+ 622
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 609,
+ 435,
+ 625
+ ],
+ "score": 1.0,
+ "content": "improves the accuracy from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 435,
+ 611,
+ 462,
+ 622
+ ],
+ "score": 0.86,
+ "content": "6 0 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 609,
+ 473,
+ 625
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 611,
+ 505,
+ 622
+ ],
+ "score": 0.86,
+ "content": "6 2 . 7 1 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 184,
+ 635
+ ],
+ "score": 1.0,
+ "content": "and including both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 622,
+ 204,
+ 633
+ ],
+ "score": 0.91,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 622,
+ 222,
+ 635
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 622,
+ 241,
+ 633
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 241,
+ 622,
+ 382,
+ 635
+ ],
+ "score": 1.0,
+ "content": "provides a further improvement to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 622,
+ 410,
+ 633
+ ],
+ "score": 0.86,
+ "content": "6 3 . 3 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 410,
+ 622,
+ 506,
+ 635
+ ],
+ "score": 1.0,
+ "content": ". Since these loss terms",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "improve performance without introducing any additional hyperparameters, we include them in our",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 305,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 305,
+ 657
+ ],
+ "score": 1.0,
+ "content": "recipe and use them in all following experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 105,
+ 600,
+ 506,
+ 657
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 668,
+ 310,
+ 680
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 311,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 311,
+ 682
+ ],
+ "score": 1.0,
+ "content": "3.3 Parameter-efficient fine-tuning with (IA)3",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 45
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 688,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "In order to compare favorably to few-shot ICL, we need a PEFT method that has the following",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "properties: First, it must add or update as few parameters as possible to avoid incurring storage",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 507,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 507,
+ 723
+ ],
+ "score": 1.0,
+ "content": "and memory costs. Second, it should achieve strong accuracy after few-shot training on new tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 47,
+ "bbox_fs": [
+ 105,
+ 687,
+ 507,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 72,
+ 505,
+ 193
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "Finally, it must allow for mixed-task batches, since that is a capability of ICL. In order to easily",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "enable mixed-task batches, a PEFT method should ideally not modify the model itself. Otherwise,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 94,
+ 506,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 94,
+ 506,
+ 107
+ ],
+ "score": 1.0,
+ "content": "each example in a batch would effectively need to be processed by a different model or computational",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "graph. A more convenient alternative is provided by methods that directly modify the activations of",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 130
+ ],
+ "score": 1.0,
+ "content": "the model since this can be done independently and cheaply to each example in the batch according",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "to which task the example corresponds to. Prompt tuning and prefix tuning methods [14, 29] work by",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 506,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 506,
+ 150
+ ],
+ "score": 1.0,
+ "content": "concatenating learned vectors to activation or embedding sequences and are therefore examples of",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "activation-modifying PEFT methods that allow for mixed-task batches. However, as we will discuss",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 160,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 160,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "later, we were unable to attain reasonable accuracy with prompt tuning and found that the more",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 505,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 505,
+ 183
+ ],
+ "score": 1.0,
+ "content": "performant PEFT methods did not allow for mixed-task batches. We therefore developed a new PEFT",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 182,
+ 243,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 243,
+ 194
+ ],
+ "score": 1.0,
+ "content": "method that meets our desiderata.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 198,
+ 505,
+ 308
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "As an alternative, we explored element-wise multiplication (i.e. rescaling) of the model’s activations",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 406,
+ 222
+ ],
+ "score": 1.0,
+ "content": "against a learned vector. Specifically, we consider adaptation of the form",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 209,
+ 429,
+ 219
+ ],
+ "score": 0.86,
+ "content": "l \\odot x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 208,
+ 457,
+ 222
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 458,
+ 208,
+ 487,
+ 219
+ ],
+ "score": 0.91,
+ "content": "l \\in \\mathbb { R } ^ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 208,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "is a",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 104,
+ 217,
+ 504,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 217,
+ 220,
+ 234
+ ],
+ "score": 1.0,
+ "content": "learned task-specific vector,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 220,
+ 221,
+ 229,
+ 230
+ ],
+ "score": 0.81,
+ "content": "\\odot",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 217,
+ 406,
+ 234
+ ],
+ "score": 1.0,
+ "content": "represents element-wise multiplication, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 407,
+ 219,
+ 450,
+ 230
+ ],
+ "score": 0.92,
+ "content": "\\boldsymbol { x } \\in \\mathbb { R } ^ { T \\times d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 217,
+ 495,
+ 234
+ ],
+ "score": 1.0,
+ "content": "is a length-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 221,
+ 504,
+ 230
+ ],
+ "score": 0.8,
+ "content": "T",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 103,
+ 227,
+ 503,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 227,
+ 388,
+ 247
+ ],
+ "score": 1.0,
+ "content": "sequence of activations. We use “broadcasting notation” [46] so that the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 231,
+ 416,
+ 243
+ ],
+ "score": 0.93,
+ "content": "( i , j ) ^ { \\mathrm { t h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 227,
+ 450,
+ 247
+ ],
+ "score": 1.0,
+ "content": "entry of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 231,
+ 470,
+ 242
+ ],
+ "score": 0.89,
+ "content": "l \\odot x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 227,
+ 479,
+ 247
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 231,
+ 503,
+ 243
+ ],
+ "score": 0.91,
+ "content": "l _ { j } x _ { i , j }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 242,
+ 506,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 242,
+ 506,
+ 254
+ ],
+ "score": 1.0,
+ "content": "In preliminary experiments, we found it was not necessary to introduce a learned rescaling vector",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "for each set of activations in the Transformer model. Instead, we found it was sufficient to introduce",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 263,
+ 505,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 263,
+ 505,
+ 276
+ ],
+ "score": 1.0,
+ "content": "rescaling vectors on the keys and values in self-attention and encoder-decoder attention mechanisms",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 273,
+ 506,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 506,
+ 288
+ ],
+ "score": 1.0,
+ "content": "and on the intermediate activation of the position-wise feed-forward networks. Specifically, using",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 284,
+ 506,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 405,
+ 298
+ ],
+ "score": 1.0,
+ "content": "the notation from Vaswani et al. [33], we introduce three learned vectors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 285,
+ 483,
+ 297
+ ],
+ "score": 0.92,
+ "content": "l _ { \\mathbf { k } } \\in \\mathbb { R } ^ { \\hat { d } _ { \\mathbf { k } } } , l _ { \\mathbf { v } } \\in \\mathbb { R } ^ { \\hat { d } _ { \\mathbf { v } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 284,
+ 506,
+ 298
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 107,
+ 295,
+ 372,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 296,
+ 145,
+ 307
+ ],
+ "score": 0.92,
+ "content": "l _ { \\mathrm { f f } } \\in \\mathbb { R } ^ { d _ { \\mathrm { f f } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 295,
+ 372,
+ 309
+ ],
+ "score": 1.0,
+ "content": ", which are introduced into the attention mechanisms as:",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 234,
+ 312,
+ 377,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 234,
+ 312,
+ 377,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 234,
+ 312,
+ 377,
+ 340
+ ],
+ "score": 0.95,
+ "content": "\\mathrm { s o f t m a x } \\left( \\frac { Q ( l _ { \\mathrm { k } } \\odot K ^ { T } ) } { \\sqrt { d _ { k } } } \\right) ( l _ { \\mathrm { v } } \\odot V )",
+ "type": "interline_equation",
+ "image_path": "ea464786d6df4a5cf212c9add5baf0056ab43c59bad7c04cf8d7096f728c58ee.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 234,
+ 312,
+ 377,
+ 326.0
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 234,
+ 326.0,
+ 377,
+ 340.0
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 345,
+ 505,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 311,
+ 358
+ ],
+ "score": 1.0,
+ "content": "and in the position-wise feed-forward networks as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 345,
+ 387,
+ 357
+ ],
+ "score": 0.93,
+ "content": "( l _ { \\mathrm { f f } } \\odot \\gamma ( W _ { 1 } x ) ) W _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 387,
+ 344,
+ 417,
+ 358
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 418,
+ 347,
+ 425,
+ 357
+ ],
+ "score": 0.83,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 344,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "is the feed-forward",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 316,
+ 369
+ ],
+ "score": 1.0,
+ "content": "network nonlinearity. We introduce a separate set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 357,
+ 337,
+ 367
+ ],
+ "score": 0.9,
+ "content": "l _ { \\mathrm { k } } , l _ { \\mathrm { v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 337,
+ 356,
+ 358,
+ 369
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 357,
+ 368,
+ 367
+ ],
+ "score": 0.87,
+ "content": "l _ { \\mathrm { f f } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 368,
+ 356,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "vectors in each Transformer layer",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 104,
+ 366,
+ 506,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 366,
+ 211,
+ 380
+ ],
+ "score": 1.0,
+ "content": "block. This adds a total of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 367,
+ 279,
+ 379
+ ],
+ "score": 0.92,
+ "content": "L ( d _ { k } + d _ { v } + \\bar { d } _ { \\mathrm { f f } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 279,
+ 366,
+ 364,
+ 380
+ ],
+ "score": 1.0,
+ "content": "new parameters for a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 368,
+ 372,
+ 377
+ ],
+ "score": 0.82,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 366,
+ 506,
+ 380
+ ],
+ "score": 1.0,
+ "content": "-layer-block Transformer encoder",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 378,
+ 505,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 378,
+ 124,
+ 390
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 124,
+ 378,
+ 207,
+ 390
+ ],
+ "score": 0.91,
+ "content": "L ( 2 d _ { k } + 2 d _ { v } + d _ { \\mathrm { f f } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 378,
+ 505,
+ 390
+ ],
+ "score": 1.0,
+ "content": "(with factors of 2 accounting for the presence of both self-attention and",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 388,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 388,
+ 236,
+ 401
+ ],
+ "score": 1.0,
+ "content": "encoder-decoder attention) for a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 389,
+ 245,
+ 399
+ ],
+ "score": 0.78,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 388,
+ 330,
+ 401
+ ],
+ "score": 1.0,
+ "content": "-layer-block decoder.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 389,
+ 352,
+ 400
+ ],
+ "score": 0.91,
+ "content": "l _ { \\mathrm { k } } , l _ { \\mathrm { v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 352,
+ 388,
+ 372,
+ 401
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 372,
+ 389,
+ 382,
+ 400
+ ],
+ "score": 0.86,
+ "content": "l _ { \\mathrm { f f } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 388,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "are all initialized with ones so",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "that the overall function computed by the model does not change when they are added. We call our",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 506,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 140,
+ 423
+ ],
+ "score": 1.0,
+ "content": "method",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 141,
+ 410,
+ 165,
+ 421
+ ],
+ "score": 0.62,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 165,
+ 410,
+ 506,
+ 423
+ ],
+ "score": 1.0,
+ "content": ", which stands for “Infused Adapter by Inhibiting and Amplifying Inner Activations”.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 426,
+ 505,
+ 504
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 426,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 426,
+ 132,
+ 438
+ ],
+ "score": 0.33,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 426,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "makes mixed-task batches possible because each sequence of activations in the batch can be",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 450
+ ],
+ "score": 1.0,
+ "content": "separately and cheaply multiplied by its associated learned task vector. We also note that, in the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 104,
+ 448,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 448,
+ 462,
+ 462
+ ],
+ "score": 1.0,
+ "content": "event that a model will only be used on a single task, the modifications introduced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 462,
+ 448,
+ 487,
+ 460
+ ],
+ "score": 0.68,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 448,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "can",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 472
+ ],
+ "score": 1.0,
+ "content": "also be applied to weight matrices permanently so that no elementwise multiplication is required and",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 470,
+ 505,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 470,
+ 505,
+ 483
+ ],
+ "score": 1.0,
+ "content": "the model’s architecture remains unchanged. This possible because element-wise multiplications",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 162,
+ 494
+ ],
+ "score": 1.0,
+ "content": "performed in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 481,
+ 187,
+ 492
+ ],
+ "score": 0.32,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 481,
+ 387,
+ 494
+ ],
+ "score": 1.0,
+ "content": "always co-occur with a matrix multiplication, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 481,
+ 473,
+ 493
+ ],
+ "score": 0.91,
+ "content": "l \\odot W x = ( l \\odot W ) x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 474,
+ 481,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": ". In this",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 466,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 466,
+ 505
+ ],
+ "score": 1.0,
+ "content": "case, our method incurs no additional computational cost compared to the original model.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 508,
+ 506,
+ 662
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 506,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 506,
+ 522
+ ],
+ "score": 1.0,
+ "content": "To validate (IA)3, we compare it to a large variety of existing adaptation methods in our setting of",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 519,
+ 506,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 519,
+ 506,
+ 533
+ ],
+ "score": 1.0,
+ "content": "fine-tuning T0-3B on few-shot datasets from held-out tasks. Specifically, we compare with 9 strong",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 530,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 530,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "PEFT methods: BitFit [47] which updates only the bias parameters; Adapters [23] which introduce",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "task-specific layers after the self-attention and position-wise feed-forward networks; Compacter and",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 552,
+ 506,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 150,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Compacter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 554,
+ 163,
+ 563
+ ],
+ "score": 0.49,
+ "content": "^ { + + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 163,
+ 552,
+ 506,
+ 565
+ ],
+ "score": 1.0,
+ "content": "[28] which improve upon adapters by using low-rank matrices and hypercomplex mul-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "tiplication; prompt tuning [14] which learns task-specific prompt embeddings that are concatenated to",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 507,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 507,
+ 588
+ ],
+ "score": 1.0,
+ "content": "the model’s input; FISH Mask [26] which chooses a subset of parameters to update based on their ap-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 506,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 506,
+ 597
+ ],
+ "score": 1.0,
+ "content": "proximate Fisher information; Intrinsic SAID [27] which performs optimization in a low-dimensional",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "subspace; prefix-tuning [29] which learns task-specific vectors that are concatenated to the model’s",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "activations; and LoRA [13] which assigns low-rank updates to parameter matrices. Additionally, we",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 507,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 507,
+ 631
+ ],
+ "score": 1.0,
+ "content": "include the baselines of full-model fine-tuning and updating only the layer normalization parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 506,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 506,
+ 643
+ ],
+ "score": 1.0,
+ "content": "For certain methods that allow changing the parameter efficiency, we report results for different",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 144,
+ 653
+ ],
+ "score": 1.0,
+ "content": "budgets:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 640,
+ 167,
+ 650
+ ],
+ "score": 0.87,
+ "content": "0 . 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 639,
+ 185,
+ 653
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 640,
+ 213,
+ 650
+ ],
+ "score": 0.88,
+ "content": "0 . 0 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 639,
+ 506,
+ 653
+ ],
+ "score": 1.0,
+ "content": "sparsity for FISH Mask, 10 and 100 learned prompt vectors for prompt",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 405,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 405,
+ 663
+ ],
+ "score": 1.0,
+ "content": "tuning, and 20,000- or 500,000-dimensional subspaces for Intrinsic SAID.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 43.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 667,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "The results are shown in fig. 2, with detailed per-dataset results in appendix C. We find that (IA)3",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 678,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 678,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "is the only method that attains higher accuracy than the full-model-fine-tuning baseline. While",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 507,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 507,
+ 702
+ ],
+ "score": 1.0,
+ "content": "other PEFT methods (e.g. Intrinsic SAID and prompt tuning) update or introduce fewer parameters,",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 107,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 699,
+ 132,
+ 711
+ ],
+ "score": 0.62,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "performs considerably better. Our results and setting differ with some past work on the",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "PEFT methods we compare against. Mahabadi et al. [28] report that Compacter and Compacter++",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 72,
+ 505,
+ 193
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "Finally, it must allow for mixed-task batches, since that is a capability of ICL. In order to easily",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "enable mixed-task batches, a PEFT method should ideally not modify the model itself. Otherwise,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 94,
+ 506,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 94,
+ 506,
+ 107
+ ],
+ "score": 1.0,
+ "content": "each example in a batch would effectively need to be processed by a different model or computational",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "graph. A more convenient alternative is provided by methods that directly modify the activations of",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 130
+ ],
+ "score": 1.0,
+ "content": "the model since this can be done independently and cheaply to each example in the batch according",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "to which task the example corresponds to. Prompt tuning and prefix tuning methods [14, 29] work by",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 506,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 506,
+ 150
+ ],
+ "score": 1.0,
+ "content": "concatenating learned vectors to activation or embedding sequences and are therefore examples of",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "activation-modifying PEFT methods that allow for mixed-task batches. However, as we will discuss",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 160,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 160,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "later, we were unable to attain reasonable accuracy with prompt tuning and found that the more",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 505,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 505,
+ 183
+ ],
+ "score": 1.0,
+ "content": "performant PEFT methods did not allow for mixed-task batches. We therefore developed a new PEFT",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 182,
+ 243,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 243,
+ 194
+ ],
+ "score": 1.0,
+ "content": "method that meets our desiderata.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 105,
+ 73,
+ 506,
+ 194
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 198,
+ 505,
+ 308
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "As an alternative, we explored element-wise multiplication (i.e. rescaling) of the model’s activations",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 406,
+ 222
+ ],
+ "score": 1.0,
+ "content": "against a learned vector. Specifically, we consider adaptation of the form",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 209,
+ 429,
+ 219
+ ],
+ "score": 0.86,
+ "content": "l \\odot x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 208,
+ 457,
+ 222
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 458,
+ 208,
+ 487,
+ 219
+ ],
+ "score": 0.91,
+ "content": "l \\in \\mathbb { R } ^ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 208,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "is a",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 104,
+ 217,
+ 504,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 217,
+ 220,
+ 234
+ ],
+ "score": 1.0,
+ "content": "learned task-specific vector,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 220,
+ 221,
+ 229,
+ 230
+ ],
+ "score": 0.81,
+ "content": "\\odot",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 217,
+ 406,
+ 234
+ ],
+ "score": 1.0,
+ "content": "represents element-wise multiplication, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 407,
+ 219,
+ 450,
+ 230
+ ],
+ "score": 0.92,
+ "content": "\\boldsymbol { x } \\in \\mathbb { R } ^ { T \\times d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 217,
+ 495,
+ 234
+ ],
+ "score": 1.0,
+ "content": "is a length-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 221,
+ 504,
+ 230
+ ],
+ "score": 0.8,
+ "content": "T",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 103,
+ 227,
+ 503,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 227,
+ 388,
+ 247
+ ],
+ "score": 1.0,
+ "content": "sequence of activations. We use “broadcasting notation” [46] so that the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 231,
+ 416,
+ 243
+ ],
+ "score": 0.93,
+ "content": "( i , j ) ^ { \\mathrm { t h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 227,
+ 450,
+ 247
+ ],
+ "score": 1.0,
+ "content": "entry of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 231,
+ 470,
+ 242
+ ],
+ "score": 0.89,
+ "content": "l \\odot x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 227,
+ 479,
+ 247
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 231,
+ 503,
+ 243
+ ],
+ "score": 0.91,
+ "content": "l _ { j } x _ { i , j }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 242,
+ 506,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 242,
+ 506,
+ 254
+ ],
+ "score": 1.0,
+ "content": "In preliminary experiments, we found it was not necessary to introduce a learned rescaling vector",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "for each set of activations in the Transformer model. Instead, we found it was sufficient to introduce",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 263,
+ 505,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 263,
+ 505,
+ 276
+ ],
+ "score": 1.0,
+ "content": "rescaling vectors on the keys and values in self-attention and encoder-decoder attention mechanisms",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 273,
+ 506,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 506,
+ 288
+ ],
+ "score": 1.0,
+ "content": "and on the intermediate activation of the position-wise feed-forward networks. Specifically, using",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 284,
+ 506,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 405,
+ 298
+ ],
+ "score": 1.0,
+ "content": "the notation from Vaswani et al. [33], we introduce three learned vectors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 285,
+ 483,
+ 297
+ ],
+ "score": 0.92,
+ "content": "l _ { \\mathbf { k } } \\in \\mathbb { R } ^ { \\hat { d } _ { \\mathbf { k } } } , l _ { \\mathbf { v } } \\in \\mathbb { R } ^ { \\hat { d } _ { \\mathbf { v } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 284,
+ 506,
+ 298
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 107,
+ 295,
+ 372,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 296,
+ 145,
+ 307
+ ],
+ "score": 0.92,
+ "content": "l _ { \\mathrm { f f } } \\in \\mathbb { R } ^ { d _ { \\mathrm { f f } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 295,
+ 372,
+ 309
+ ],
+ "score": 1.0,
+ "content": ", which are introduced into the attention mechanisms as:",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 103,
+ 198,
+ 506,
+ 309
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 234,
+ 312,
+ 377,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 234,
+ 312,
+ 377,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 234,
+ 312,
+ 377,
+ 340
+ ],
+ "score": 0.95,
+ "content": "\\mathrm { s o f t m a x } \\left( \\frac { Q ( l _ { \\mathrm { k } } \\odot K ^ { T } ) } { \\sqrt { d _ { k } } } \\right) ( l _ { \\mathrm { v } } \\odot V )",
+ "type": "interline_equation",
+ "image_path": "ea464786d6df4a5cf212c9add5baf0056ab43c59bad7c04cf8d7096f728c58ee.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 234,
+ 312,
+ 377,
+ 326.0
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 234,
+ 326.0,
+ 377,
+ 340.0
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 345,
+ 505,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 311,
+ 358
+ ],
+ "score": 1.0,
+ "content": "and in the position-wise feed-forward networks as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 345,
+ 387,
+ 357
+ ],
+ "score": 0.93,
+ "content": "( l _ { \\mathrm { f f } } \\odot \\gamma ( W _ { 1 } x ) ) W _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 387,
+ 344,
+ 417,
+ 358
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 418,
+ 347,
+ 425,
+ 357
+ ],
+ "score": 0.83,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 344,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "is the feed-forward",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 316,
+ 369
+ ],
+ "score": 1.0,
+ "content": "network nonlinearity. We introduce a separate set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 357,
+ 337,
+ 367
+ ],
+ "score": 0.9,
+ "content": "l _ { \\mathrm { k } } , l _ { \\mathrm { v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 337,
+ 356,
+ 358,
+ 369
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 357,
+ 368,
+ 367
+ ],
+ "score": 0.87,
+ "content": "l _ { \\mathrm { f f } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 368,
+ 356,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "vectors in each Transformer layer",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 104,
+ 366,
+ 506,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 366,
+ 211,
+ 380
+ ],
+ "score": 1.0,
+ "content": "block. This adds a total of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 367,
+ 279,
+ 379
+ ],
+ "score": 0.92,
+ "content": "L ( d _ { k } + d _ { v } + \\bar { d } _ { \\mathrm { f f } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 279,
+ 366,
+ 364,
+ 380
+ ],
+ "score": 1.0,
+ "content": "new parameters for a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 368,
+ 372,
+ 377
+ ],
+ "score": 0.82,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 366,
+ 506,
+ 380
+ ],
+ "score": 1.0,
+ "content": "-layer-block Transformer encoder",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 378,
+ 505,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 378,
+ 124,
+ 390
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 124,
+ 378,
+ 207,
+ 390
+ ],
+ "score": 0.91,
+ "content": "L ( 2 d _ { k } + 2 d _ { v } + d _ { \\mathrm { f f } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 378,
+ 505,
+ 390
+ ],
+ "score": 1.0,
+ "content": "(with factors of 2 accounting for the presence of both self-attention and",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 388,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 388,
+ 236,
+ 401
+ ],
+ "score": 1.0,
+ "content": "encoder-decoder attention) for a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 389,
+ 245,
+ 399
+ ],
+ "score": 0.78,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 388,
+ 330,
+ 401
+ ],
+ "score": 1.0,
+ "content": "-layer-block decoder.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 389,
+ 352,
+ 400
+ ],
+ "score": 0.91,
+ "content": "l _ { \\mathrm { k } } , l _ { \\mathrm { v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 352,
+ 388,
+ 372,
+ 401
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 372,
+ 389,
+ 382,
+ 400
+ ],
+ "score": 0.86,
+ "content": "l _ { \\mathrm { f f } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 388,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "are all initialized with ones so",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "that the overall function computed by the model does not change when they are added. We call our",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 506,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 140,
+ 423
+ ],
+ "score": 1.0,
+ "content": "method",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 141,
+ 410,
+ 165,
+ 421
+ ],
+ "score": 0.62,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 165,
+ 410,
+ 506,
+ 423
+ ],
+ "score": 1.0,
+ "content": ", which stands for “Infused Adapter by Inhibiting and Amplifying Inner Activations”.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 104,
+ 344,
+ 506,
+ 423
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 426,
+ 505,
+ 504
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 426,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 426,
+ 132,
+ 438
+ ],
+ "score": 0.33,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 426,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "makes mixed-task batches possible because each sequence of activations in the batch can be",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 450
+ ],
+ "score": 1.0,
+ "content": "separately and cheaply multiplied by its associated learned task vector. We also note that, in the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 104,
+ 448,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 448,
+ 462,
+ 462
+ ],
+ "score": 1.0,
+ "content": "event that a model will only be used on a single task, the modifications introduced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 462,
+ 448,
+ 487,
+ 460
+ ],
+ "score": 0.68,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 448,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "can",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 472
+ ],
+ "score": 1.0,
+ "content": "also be applied to weight matrices permanently so that no elementwise multiplication is required and",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 470,
+ 505,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 470,
+ 505,
+ 483
+ ],
+ "score": 1.0,
+ "content": "the model’s architecture remains unchanged. This possible because element-wise multiplications",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 162,
+ 494
+ ],
+ "score": 1.0,
+ "content": "performed in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 481,
+ 187,
+ 492
+ ],
+ "score": 0.32,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 481,
+ 387,
+ 494
+ ],
+ "score": 1.0,
+ "content": "always co-occur with a matrix multiplication, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 481,
+ 473,
+ 493
+ ],
+ "score": 0.91,
+ "content": "l \\odot W x = ( l \\odot W ) x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 474,
+ 481,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": ". In this",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 466,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 466,
+ 505
+ ],
+ "score": 1.0,
+ "content": "case, our method incurs no additional computational cost compared to the original model.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 104,
+ 426,
+ 506,
+ 505
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 508,
+ 506,
+ 662
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 506,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 506,
+ 522
+ ],
+ "score": 1.0,
+ "content": "To validate (IA)3, we compare it to a large variety of existing adaptation methods in our setting of",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 519,
+ 506,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 519,
+ 506,
+ 533
+ ],
+ "score": 1.0,
+ "content": "fine-tuning T0-3B on few-shot datasets from held-out tasks. Specifically, we compare with 9 strong",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 530,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 530,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "PEFT methods: BitFit [47] which updates only the bias parameters; Adapters [23] which introduce",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "task-specific layers after the self-attention and position-wise feed-forward networks; Compacter and",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 552,
+ 506,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 150,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Compacter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 554,
+ 163,
+ 563
+ ],
+ "score": 0.49,
+ "content": "^ { + + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 163,
+ 552,
+ 506,
+ 565
+ ],
+ "score": 1.0,
+ "content": "[28] which improve upon adapters by using low-rank matrices and hypercomplex mul-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "tiplication; prompt tuning [14] which learns task-specific prompt embeddings that are concatenated to",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 507,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 507,
+ 588
+ ],
+ "score": 1.0,
+ "content": "the model’s input; FISH Mask [26] which chooses a subset of parameters to update based on their ap-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 506,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 506,
+ 597
+ ],
+ "score": 1.0,
+ "content": "proximate Fisher information; Intrinsic SAID [27] which performs optimization in a low-dimensional",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "subspace; prefix-tuning [29] which learns task-specific vectors that are concatenated to the model’s",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "activations; and LoRA [13] which assigns low-rank updates to parameter matrices. Additionally, we",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 507,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 507,
+ 631
+ ],
+ "score": 1.0,
+ "content": "include the baselines of full-model fine-tuning and updating only the layer normalization parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 506,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 506,
+ 643
+ ],
+ "score": 1.0,
+ "content": "For certain methods that allow changing the parameter efficiency, we report results for different",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 144,
+ 653
+ ],
+ "score": 1.0,
+ "content": "budgets:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 640,
+ 167,
+ 650
+ ],
+ "score": 0.87,
+ "content": "0 . 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 639,
+ 185,
+ 653
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 640,
+ 213,
+ 650
+ ],
+ "score": 0.88,
+ "content": "0 . 0 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 639,
+ 506,
+ 653
+ ],
+ "score": 1.0,
+ "content": "sparsity for FISH Mask, 10 and 100 learned prompt vectors for prompt",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 405,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 405,
+ 663
+ ],
+ "score": 1.0,
+ "content": "tuning, and 20,000- or 500,000-dimensional subspaces for Intrinsic SAID.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 43.5,
+ "bbox_fs": [
+ 105,
+ 507,
+ 507,
+ 663
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 667,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "The results are shown in fig. 2, with detailed per-dataset results in appendix C. We find that (IA)3",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 678,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 678,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "is the only method that attains higher accuracy than the full-model-fine-tuning baseline. While",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 507,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 507,
+ 702
+ ],
+ "score": 1.0,
+ "content": "other PEFT methods (e.g. Intrinsic SAID and prompt tuning) update or introduce fewer parameters,",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 107,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 699,
+ 132,
+ 711
+ ],
+ "score": 0.62,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "performs considerably better. Our results and setting differ with some past work on the",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "PEFT methods we compare against. Mahabadi et al. [28] report that Compacter and Compacter++",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 106,
+ 73,
+ 506,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 506,
+ 85
+ ],
+ "score": 1.0,
+ "content": "outperform full-model fine-tuning, including in the few-shot setting. Lester et al. [14] found that",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 84,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 84,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "prompt tuning could match full-model fine-tuning, and in subsequent work Wei et al. [48] found that",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 109
+ ],
+ "score": 1.0,
+ "content": "prompt tuning performed well when applied to a multitask fine-tuned model in the few-shot setting.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 106,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "In both cases, we experimented with various hyperparameter choices to try to match past results.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 129
+ ],
+ "score": 1.0,
+ "content": "We hypothesize the disagreement comes from us using a different model and different datasets. For",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "prompt tuning specifically, we noticed that the validation set performance could fluctuate wildly over",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 138,
+ 353,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 138,
+ 353,
+ 151
+ ],
+ "score": 1.0,
+ "content": "the course of training, hinting at possible optimization issues.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 53,
+ "bbox_fs": [
+ 105,
+ 666,
+ 507,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 72,
+ 506,
+ 150
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 506,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 506,
+ 85
+ ],
+ "score": 1.0,
+ "content": "outperform full-model fine-tuning, including in the few-shot setting. Lester et al. [14] found that",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 84,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 84,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "prompt tuning could match full-model fine-tuning, and in subsequent work Wei et al. [48] found that",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 109
+ ],
+ "score": 1.0,
+ "content": "prompt tuning performed well when applied to a multitask fine-tuned model in the few-shot setting.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 106,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "In both cases, we experimented with various hyperparameter choices to try to match past results.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 129
+ ],
+ "score": 1.0,
+ "content": "We hypothesize the disagreement comes from us using a different model and different datasets. For",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "prompt tuning specifically, we noticed that the validation set performance could fluctuate wildly over",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 138,
+ 353,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 138,
+ 353,
+ 151
+ ],
+ "score": 1.0,
+ "content": "the course of training, hinting at possible optimization issues.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 382
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 382
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 382
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "ec5ffa216af6c47257d616d6cc1d5f6ea60664e581ae772fa7d50daf965b6e9f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 176.11111111111111
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 112,
+ 176.11111111111111,
+ 292,
+ 188.22222222222223
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 112,
+ 188.22222222222223,
+ 292,
+ 200.33333333333334
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 112,
+ 200.33333333333334,
+ 292,
+ 212.44444444444446
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 112,
+ 212.44444444444446,
+ 292,
+ 224.55555555555557
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 112,
+ 224.55555555555557,
+ 292,
+ 236.66666666666669
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 112,
+ 236.66666666666669,
+ 292,
+ 248.7777777777778
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 112,
+ 248.7777777777778,
+ 292,
+ 260.8888888888889
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 112,
+ 260.8888888888889,
+ 292,
+ 273.0
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 112,
+ 273.0,
+ 292,
+ 285.1111111111111
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 112,
+ 285.1111111111111,
+ 292,
+ 297.2222222222222
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 112,
+ 297.2222222222222,
+ 292,
+ 309.33333333333326
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 112,
+ 309.33333333333326,
+ 292,
+ 321.44444444444434
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 112,
+ 321.44444444444434,
+ 292,
+ 333.55555555555543
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 112,
+ 333.55555555555543,
+ 292,
+ 345.6666666666665
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 112,
+ 345.6666666666665,
+ 292,
+ 357.7777777777776
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 112,
+ 357.7777777777776,
+ 292,
+ 369.8888888888887
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 112,
+ 369.8888888888887,
+ 292,
+ 381.9999999999998
+ ],
+ "spans": [],
+ "index": 24
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 389,
+ 299,
+ 444
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 298,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 279,
+ 403
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Accuracy of PEFT methods with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 390,
+ 298,
+ 401
+ ],
+ "score": 0.87,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 400,
+ 299,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 123,
+ 412
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 124,
+ 401,
+ 143,
+ 412
+ ],
+ "score": 0.91,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 400,
+ 299,
+ 412
+ ],
+ "score": 1.0,
+ "content": "when applied to T0-3B. Methods that",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 411,
+ 299,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 411,
+ 299,
+ 424
+ ],
+ "score": 1.0,
+ "content": "with variable parameter budgets are represented",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 422,
+ 299,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 422,
+ 299,
+ 434
+ ],
+ "score": 1.0,
+ "content": "with larger and smaller markers for more or less",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 155,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 155,
+ 444
+ ],
+ "score": 1.0,
+ "content": "parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 21.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 356
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 356
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 356
+ ],
+ "score": 0.961,
+ "type": "image",
+ "image_path": "4069406d67c113c81e8fa2fa95996a77df0f6f0e675138606ae99952a7f0b0f2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 37,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 179.6
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 318,
+ 179.6,
+ 500,
+ 192.2
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 318,
+ 192.2,
+ 500,
+ 204.79999999999998
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 318,
+ 204.79999999999998,
+ 500,
+ 217.39999999999998
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 318,
+ 217.39999999999998,
+ 500,
+ 229.99999999999997
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 318,
+ 229.99999999999997,
+ 500,
+ 242.59999999999997
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 318,
+ 242.59999999999997,
+ 500,
+ 255.19999999999996
+ ],
+ "spans": [],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 318,
+ 255.19999999999996,
+ 500,
+ 267.79999999999995
+ ],
+ "spans": [],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 318,
+ 267.79999999999995,
+ 500,
+ 280.4
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 318,
+ 280.4,
+ 500,
+ 293.0
+ ],
+ "spans": [],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 318,
+ 293.0,
+ 500,
+ 305.6
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 318,
+ 305.6,
+ 500,
+ 318.20000000000005
+ ],
+ "spans": [],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 318,
+ 318.20000000000005,
+ 500,
+ 330.80000000000007
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 318,
+ 330.80000000000007,
+ 500,
+ 343.4000000000001
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 318,
+ 343.4000000000001,
+ 500,
+ 356.0000000000001
+ ],
+ "spans": [],
+ "index": 44
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 312,
+ 361,
+ 505,
+ 427
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 311,
+ 360,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 360,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Accuracy of different few-shot learning",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 311,
+ 371,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 371,
+ 402,
+ 384
+ ],
+ "score": 1.0,
+ "content": "methods. T-Few uses",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 402,
+ 372,
+ 427,
+ 383
+ ],
+ "score": 0.27,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 371,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "for PEFT methods",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 311,
+ 383,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 383,
+ 470,
+ 394
+ ],
+ "score": 1.0,
+ "content": "of T0, T0 uses zero-shot learning, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 470,
+ 383,
+ 505,
+ 394
+ ],
+ "score": 0.69,
+ "content": "_ { \\mathrm { T 5 + L M } }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 312,
+ 394,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 394,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "and the GPT-3 variants use few-shot ICL. The",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 312,
+ 405,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 407,
+ 318,
+ 415
+ ],
+ "score": 0.51,
+ "content": "\\mathbf { X }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 319,
+ 405,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "-axis corresponds to inference costs; details are",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 311,
+ 416,
+ 408,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 416,
+ 408,
+ 427
+ ],
+ "score": 1.0,
+ "content": "provided in section 4.2.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 47.5
+ }
+ ],
+ "index": 42.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 459,
+ 210,
+ 471
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 458,
+ 211,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 458,
+ 211,
+ 473
+ ],
+ "score": 1.0,
+ "content": "3.4 Pre-training (IA)3",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 51
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 479,
+ 505,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "In recent work, Gu et al. [18], Vu et al. [19] showed that pre-training the prompt embeddings in",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 491,
+ 506,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 491,
+ 506,
+ 503
+ ],
+ "score": 1.0,
+ "content": "prompt tuning can improve performance when fine-tuning on downstream few-shot tasks. For pre-",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 490,
+ 514
+ ],
+ "score": 1.0,
+ "content": "training, Gu et al. [18] use a suite of self-supervised tasks applied to unlabeled text data, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 491,
+ 502,
+ 505,
+ 512
+ ],
+ "score": 0.45,
+ "content": "\\mathrm { V u }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": "et al. [19] consider using embeddings from a separate task or multitask mixture. We follow Vu et al.",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 352,
+ 536
+ ],
+ "score": 1.0,
+ "content": "[19] and simply pre-train the new parameters introduced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 523,
+ 378,
+ 534
+ ],
+ "score": 0.67,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 378,
+ 523,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "on the same multitask mixture",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 105,
+ 534,
+ 504,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 534,
+ 158,
+ 546
+ ],
+ "score": 1.0,
+ "content": "used to train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 534,
+ 170,
+ 545
+ ],
+ "score": 0.29,
+ "content": "\\mathrm { T 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 534,
+ 479,
+ 546
+ ],
+ "score": 1.0,
+ "content": ". We pre-train for 100,000 steps with a batch size of 16 before fine-tuning the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 479,
+ 534,
+ 504,
+ 545
+ ],
+ "score": 0.43,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 105,
+ 545,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "parameters on each individual downstream dataset. A full comparison of accuracy with and without",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 157,
+ 569
+ ],
+ "score": 1.0,
+ "content": "pre-training",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 157,
+ 556,
+ 182,
+ 567
+ ],
+ "score": 0.54,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 555,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "is detailed in appendix D. We find that pre-training improves fine-tuned accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 314,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 314,
+ 580
+ ],
+ "score": 1.0,
+ "content": "from 64.6 to 65.8 and therefore add it to our recipe.",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ }
+ ],
+ "index": 56
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 592,
+ 244,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 245,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 245,
+ 606
+ ],
+ "score": 1.0,
+ "content": "3.5 Combining the ingredients",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ }
+ ],
+ "index": 61
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 612,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "In summary, the T-Few recipe is defined as follows: We use the T0 model as a backbone. We add",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 107,
+ 622,
+ 506,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 623,
+ 132,
+ 635
+ ],
+ "score": 0.46,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 622,
+ 452,
+ 636
+ ],
+ "score": 1.0,
+ "content": "for downstream task adaptation and use parameters initialized from pre-training",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 452,
+ 623,
+ 477,
+ 635
+ ],
+ "score": 0.48,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 622,
+ 506,
+ 636
+ ],
+ "score": 1.0,
+ "content": "on the",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 218,
+ 648
+ ],
+ "score": 1.0,
+ "content": "same multitask mixture for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 635,
+ 232,
+ 645
+ ],
+ "score": 0.3,
+ "content": "\\mathrm { T 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 633,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": ". As an objective, we use the sum of a standard language modeling",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 105,
+ 645,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 645,
+ 125,
+ 658
+ ],
+ "score": 1.0,
+ "content": "loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 646,
+ 145,
+ 656
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 645,
+ 234,
+ 658
+ ],
+ "score": 1.0,
+ "content": ", an unlikelihood loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 646,
+ 254,
+ 657
+ ],
+ "score": 0.91,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 254,
+ 645,
+ 464,
+ 658
+ ],
+ "score": 1.0,
+ "content": "for incorrect choices, and a length-normalized loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 645,
+ 484,
+ 656
+ ],
+ "score": 0.89,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 645,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": ". We",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "train for 1,000 steps with a batch size of 8 sequences using the Adafactor optimizer [49] with a",
+ "type": "text"
+ }
+ ],
+ "index": 66
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 168,
+ 680
+ ],
+ "score": 1.0,
+ "content": "learning rate of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 667,
+ 190,
+ 677
+ ],
+ "score": 0.91,
+ "content": "3 e ^ { - 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 665,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": "and a linear decay schedule with a 60-step warmup. We apply prompt templates",
+ "type": "text"
+ }
+ ],
+ "index": 67
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "to downstream datasets during training and inference to convert each example into an instructive",
+ "type": "text"
+ }
+ ],
+ "index": 68
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "text-to-text format. Importantly, we apply this recipe to every downstream dataset in exactly the same",
+ "type": "text"
+ }
+ ],
+ "index": 69
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "way without per-dataset hyperparameter tuning or modifications. This makes the recipe a realistic",
+ "type": "text"
+ }
+ ],
+ "index": 70
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 462,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 462,
+ 723
+ ],
+ "score": 1.0,
+ "content": "option for few-shot learning settings where validation sets are tiny by definition [31, 32].",
+ "type": "text"
+ }
+ ],
+ "index": 71
+ }
+ ],
+ "index": 66.5
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "7",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 72,
+ 506,
+ 150
+ ],
+ "lines": [],
+ "index": 3,
+ "bbox_fs": [
+ 105,
+ 73,
+ 506,
+ 151
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 382
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 382
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 382
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "ec5ffa216af6c47257d616d6cc1d5f6ea60664e581ae772fa7d50daf965b6e9f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 112,
+ 164,
+ 292,
+ 176.11111111111111
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 112,
+ 176.11111111111111,
+ 292,
+ 188.22222222222223
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 112,
+ 188.22222222222223,
+ 292,
+ 200.33333333333334
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 112,
+ 200.33333333333334,
+ 292,
+ 212.44444444444446
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 112,
+ 212.44444444444446,
+ 292,
+ 224.55555555555557
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 112,
+ 224.55555555555557,
+ 292,
+ 236.66666666666669
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 112,
+ 236.66666666666669,
+ 292,
+ 248.7777777777778
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 112,
+ 248.7777777777778,
+ 292,
+ 260.8888888888889
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 112,
+ 260.8888888888889,
+ 292,
+ 273.0
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 112,
+ 273.0,
+ 292,
+ 285.1111111111111
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 112,
+ 285.1111111111111,
+ 292,
+ 297.2222222222222
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 112,
+ 297.2222222222222,
+ 292,
+ 309.33333333333326
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 112,
+ 309.33333333333326,
+ 292,
+ 321.44444444444434
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 112,
+ 321.44444444444434,
+ 292,
+ 333.55555555555543
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 112,
+ 333.55555555555543,
+ 292,
+ 345.6666666666665
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 112,
+ 345.6666666666665,
+ 292,
+ 357.7777777777776
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 112,
+ 357.7777777777776,
+ 292,
+ 369.8888888888887
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 112,
+ 369.8888888888887,
+ 292,
+ 381.9999999999998
+ ],
+ "spans": [],
+ "index": 24
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 389,
+ 299,
+ 444
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 298,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 279,
+ 403
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Accuracy of PEFT methods with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 390,
+ 298,
+ 401
+ ],
+ "score": 0.87,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 400,
+ 299,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 123,
+ 412
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 124,
+ 401,
+ 143,
+ 412
+ ],
+ "score": 0.91,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 400,
+ 299,
+ 412
+ ],
+ "score": 1.0,
+ "content": "when applied to T0-3B. Methods that",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 411,
+ 299,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 411,
+ 299,
+ 424
+ ],
+ "score": 1.0,
+ "content": "with variable parameter budgets are represented",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 422,
+ 299,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 422,
+ 299,
+ 434
+ ],
+ "score": 1.0,
+ "content": "with larger and smaller markers for more or less",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 155,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 155,
+ 444
+ ],
+ "score": 1.0,
+ "content": "parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 21.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 356
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 356
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 356
+ ],
+ "score": 0.961,
+ "type": "image",
+ "image_path": "4069406d67c113c81e8fa2fa95996a77df0f6f0e675138606ae99952a7f0b0f2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 37,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 318,
+ 167,
+ 500,
+ 179.6
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 318,
+ 179.6,
+ 500,
+ 192.2
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 318,
+ 192.2,
+ 500,
+ 204.79999999999998
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 318,
+ 204.79999999999998,
+ 500,
+ 217.39999999999998
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 318,
+ 217.39999999999998,
+ 500,
+ 229.99999999999997
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 318,
+ 229.99999999999997,
+ 500,
+ 242.59999999999997
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 318,
+ 242.59999999999997,
+ 500,
+ 255.19999999999996
+ ],
+ "spans": [],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 318,
+ 255.19999999999996,
+ 500,
+ 267.79999999999995
+ ],
+ "spans": [],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 318,
+ 267.79999999999995,
+ 500,
+ 280.4
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 318,
+ 280.4,
+ 500,
+ 293.0
+ ],
+ "spans": [],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 318,
+ 293.0,
+ 500,
+ 305.6
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 318,
+ 305.6,
+ 500,
+ 318.20000000000005
+ ],
+ "spans": [],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 318,
+ 318.20000000000005,
+ 500,
+ 330.80000000000007
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 318,
+ 330.80000000000007,
+ 500,
+ 343.4000000000001
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 318,
+ 343.4000000000001,
+ 500,
+ 356.0000000000001
+ ],
+ "spans": [],
+ "index": 44
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 312,
+ 361,
+ 505,
+ 427
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 311,
+ 360,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 360,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Accuracy of different few-shot learning",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 311,
+ 371,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 371,
+ 402,
+ 384
+ ],
+ "score": 1.0,
+ "content": "methods. T-Few uses",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 402,
+ 372,
+ 427,
+ 383
+ ],
+ "score": 0.27,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 371,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "for PEFT methods",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 311,
+ 383,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 383,
+ 470,
+ 394
+ ],
+ "score": 1.0,
+ "content": "of T0, T0 uses zero-shot learning, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 470,
+ 383,
+ 505,
+ 394
+ ],
+ "score": 0.69,
+ "content": "_ { \\mathrm { T 5 + L M } }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 312,
+ 394,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 394,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "and the GPT-3 variants use few-shot ICL. The",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 312,
+ 405,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 407,
+ 318,
+ 415
+ ],
+ "score": 0.51,
+ "content": "\\mathbf { X }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 319,
+ 405,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "-axis corresponds to inference costs; details are",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 311,
+ 416,
+ 408,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 416,
+ 408,
+ 427
+ ],
+ "score": 1.0,
+ "content": "provided in section 4.2.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 47.5
+ }
+ ],
+ "index": 42.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 459,
+ 210,
+ 471
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 458,
+ 211,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 458,
+ 211,
+ 473
+ ],
+ "score": 1.0,
+ "content": "3.4 Pre-training (IA)3",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 51
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 479,
+ 505,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "In recent work, Gu et al. [18], Vu et al. [19] showed that pre-training the prompt embeddings in",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 491,
+ 506,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 491,
+ 506,
+ 503
+ ],
+ "score": 1.0,
+ "content": "prompt tuning can improve performance when fine-tuning on downstream few-shot tasks. For pre-",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 490,
+ 514
+ ],
+ "score": 1.0,
+ "content": "training, Gu et al. [18] use a suite of self-supervised tasks applied to unlabeled text data, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 491,
+ 502,
+ 505,
+ 512
+ ],
+ "score": 0.45,
+ "content": "\\mathrm { V u }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": "et al. [19] consider using embeddings from a separate task or multitask mixture. We follow Vu et al.",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 352,
+ 536
+ ],
+ "score": 1.0,
+ "content": "[19] and simply pre-train the new parameters introduced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 523,
+ 378,
+ 534
+ ],
+ "score": 0.67,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 378,
+ 523,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "on the same multitask mixture",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 105,
+ 534,
+ 504,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 534,
+ 158,
+ 546
+ ],
+ "score": 1.0,
+ "content": "used to train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 534,
+ 170,
+ 545
+ ],
+ "score": 0.29,
+ "content": "\\mathrm { T 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 534,
+ 479,
+ 546
+ ],
+ "score": 1.0,
+ "content": ". We pre-train for 100,000 steps with a batch size of 16 before fine-tuning the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 479,
+ 534,
+ 504,
+ 545
+ ],
+ "score": 0.43,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 105,
+ 545,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "parameters on each individual downstream dataset. A full comparison of accuracy with and without",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 157,
+ 569
+ ],
+ "score": 1.0,
+ "content": "pre-training",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 157,
+ 556,
+ 182,
+ 567
+ ],
+ "score": 0.54,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 555,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "is detailed in appendix D. We find that pre-training improves fine-tuned accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 314,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 314,
+ 580
+ ],
+ "score": 1.0,
+ "content": "from 64.6 to 65.8 and therefore add it to our recipe.",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ }
+ ],
+ "index": 56,
+ "bbox_fs": [
+ 105,
+ 480,
+ 506,
+ 580
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 592,
+ 244,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 245,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 245,
+ 606
+ ],
+ "score": 1.0,
+ "content": "3.5 Combining the ingredients",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ }
+ ],
+ "index": 61
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 612,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "In summary, the T-Few recipe is defined as follows: We use the T0 model as a backbone. We add",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 107,
+ 622,
+ 506,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 623,
+ 132,
+ 635
+ ],
+ "score": 0.46,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 622,
+ 452,
+ 636
+ ],
+ "score": 1.0,
+ "content": "for downstream task adaptation and use parameters initialized from pre-training",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 452,
+ 623,
+ 477,
+ 635
+ ],
+ "score": 0.48,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 622,
+ 506,
+ 636
+ ],
+ "score": 1.0,
+ "content": "on the",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 218,
+ 648
+ ],
+ "score": 1.0,
+ "content": "same multitask mixture for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 635,
+ 232,
+ 645
+ ],
+ "score": 0.3,
+ "content": "\\mathrm { T 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 633,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": ". As an objective, we use the sum of a standard language modeling",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 105,
+ 645,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 645,
+ 125,
+ 658
+ ],
+ "score": 1.0,
+ "content": "loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 646,
+ 145,
+ 656
+ ],
+ "score": 0.9,
+ "content": "L _ { \\mathrm { L M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 645,
+ 234,
+ 658
+ ],
+ "score": 1.0,
+ "content": ", an unlikelihood loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 646,
+ 254,
+ 657
+ ],
+ "score": 0.91,
+ "content": "L _ { \\mathrm { U L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 254,
+ 645,
+ 464,
+ 658
+ ],
+ "score": 1.0,
+ "content": "for incorrect choices, and a length-normalized loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 645,
+ 484,
+ 656
+ ],
+ "score": 0.89,
+ "content": "L _ { \\mathrm { L N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 645,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": ". We",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "train for 1,000 steps with a batch size of 8 sequences using the Adafactor optimizer [49] with a",
+ "type": "text"
+ }
+ ],
+ "index": 66
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 168,
+ 680
+ ],
+ "score": 1.0,
+ "content": "learning rate of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 667,
+ 190,
+ 677
+ ],
+ "score": 0.91,
+ "content": "3 e ^ { - 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 665,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": "and a linear decay schedule with a 60-step warmup. We apply prompt templates",
+ "type": "text"
+ }
+ ],
+ "index": 67
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "to downstream datasets during training and inference to convert each example into an instructive",
+ "type": "text"
+ }
+ ],
+ "index": 68
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "text-to-text format. Importantly, we apply this recipe to every downstream dataset in exactly the same",
+ "type": "text"
+ }
+ ],
+ "index": 69
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "way without per-dataset hyperparameter tuning or modifications. This makes the recipe a realistic",
+ "type": "text"
+ }
+ ],
+ "index": 70
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 462,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 462,
+ 723
+ ],
+ "score": 1.0,
+ "content": "option for few-shot learning settings where validation sets are tiny by definition [31, 32].",
+ "type": "text"
+ }
+ ],
+ "index": 71
+ }
+ ],
+ "index": 66.5,
+ "bbox_fs": [
+ 105,
+ 613,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 70,
+ 288,
+ 85
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 70,
+ 290,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 70,
+ 290,
+ 87
+ ],
+ "score": 1.0,
+ "content": "4 Outperforming ICL with T-Few",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 95,
+ 505,
+ 129
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 95,
+ 505,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 95,
+ 505,
+ 108
+ ],
+ "score": 1.0,
+ "content": "Having designed and established the T-Few recipe on T0-3B, we now apply it to T0 (with 11 billion",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 106,
+ 506,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 506,
+ 119
+ ],
+ "score": 1.0,
+ "content": "parameters) and compare performance to strong few-shot ICL baselines. From this point onwards,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 118,
+ 382,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 382,
+ 129
+ ],
+ "score": 1.0,
+ "content": "we use exactly the same recipe and hyperparameters across all tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 141,
+ 237,
+ 153
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 237,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 237,
+ 154
+ ],
+ "score": 1.0,
+ "content": "4.1 Performance on T0 tasks",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 161,
+ 505,
+ 271
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 174
+ ],
+ "score": 1.0,
+ "content": "First, we evaluate T-Few on the datasets that were held out from T0’s training mixture. We compare",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 507,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 507,
+ 185
+ ],
+ "score": 1.0,
+ "content": "against zero-shot learning with T0 [1] (since we found few-shot ICL to performed worse than zero-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 182,
+ 506,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 182,
+ 307,
+ 197
+ ],
+ "score": 1.0,
+ "content": "shot for T0, see appendix E); few-shot ICL with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 307,
+ 183,
+ 342,
+ 194
+ ],
+ "score": 0.58,
+ "content": "_ { \\mathrm { T 5 + L M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 182,
+ 506,
+ 197
+ ],
+ "score": 1.0,
+ "content": "[14] (the next-step-prediction language",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 194,
+ 506,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 194,
+ 506,
+ 207
+ ],
+ "score": 1.0,
+ "content": "model upon which T0 is based); and few-shot ICL with the 6.7, 13, and 175 billion parameter variants",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 218
+ ],
+ "score": 1.0,
+ "content": "of GPT-3. See appendix E for more details on these baselines. The accuracy on the held-out T0",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 216,
+ 506,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 216,
+ 506,
+ 228
+ ],
+ "score": 1.0,
+ "content": "datasets (described in section 3.1) is shown in table 1 and fig. 3, with per-dataset results reported",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 104,
+ 226,
+ 507,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 226,
+ 507,
+ 240
+ ],
+ "score": 1.0,
+ "content": "in appendix E. We find that T-Few outperforms all other methods by a substantial margin. Notably,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 504,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 177,
+ 250
+ ],
+ "score": 1.0,
+ "content": "T-Few achieves a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 238,
+ 192,
+ 248
+ ],
+ "score": 0.86,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 193,
+ 237,
+ 484,
+ 250
+ ],
+ "score": 1.0,
+ "content": "higher accuracy than few-shot ICL with GPT-3 175B despite being about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 485,
+ 238,
+ 504,
+ 249
+ ],
+ "score": 0.86,
+ "content": "1 6 \\times",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 248,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "smaller and outperforms the smaller GPT-3 variants by an even larger margin. T-Few also attains",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 259,
+ 502,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 464,
+ 272
+ ],
+ "score": 1.0,
+ "content": "significantly higher accuracy than both zero-shot learning with T0 and few-shot ICL with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 260,
+ 498,
+ 270
+ ],
+ "score": 0.76,
+ "content": "_ { \\mathrm { T 5 + L M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 498,
+ 259,
+ 502,
+ 272
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 355
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 355
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 355
+ ],
+ "score": 0.969,
+ "html": "| Method | Acc. |
| T-Few | 75.8% |
| Human baseline [2] | 73.5% |
| PET[50] | 69.6% |
| SetFit [51] | 66.9% |
| GPT-3 [4] | 62.7% |
",
+ "type": "table",
+ "image_path": "026d7440548c8e41af0452784eff74061f5ca36d54a6d8e830c3887d67b4abe2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 17.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 315.5
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 372,
+ 315.5,
+ 504,
+ 355.0
+ ],
+ "spans": [],
+ "index": 18
+ }
+ ]
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 376
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 376
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 376
+ ],
+ "score": 0.981,
+ "html": "| Method | Inference FLOPs | Training FLOPs | Disk space | Acc. |
| T-Few | 1.1e12 | 2.7e16 | 4.2 MB | 72.4% |
| T0[1] | 1.1e12 | 0 | OB | 66.9% |
| T5+LM[14] | 4.5e13 | 0 | 16 kB | 49.6% |
| GPT-3 6.7B [4] | 5.4e13 | 0 | 16 kB | 57.2% |
| GPT-3 13B [4] | 1.0e14 | 0 | 16 kB | 60.3% |
| GPT-3 175B[4] | 1.4e15 | 0 | 16 kB | 66.6% |
",
+ "type": "table",
+ "image_path": "40b321b5687eae32a98d4d450f37e95533e02b8c81b7852cb682c522236b082d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 309.3333333333333
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 108,
+ 309.3333333333333,
+ 360,
+ 342.66666666666663
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 108,
+ 342.66666666666663,
+ 360,
+ 375.99999999999994
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 371,
+ 360,
+ 505,
+ 426
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 371,
+ 359,
+ 505,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 359,
+ 505,
+ 372
+ ],
+ "score": 1.0,
+ "content": "Table 2: Top-5 best methods on",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 370,
+ 370,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 370,
+ 370,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "RAFT as of writing. T-Few is",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 371,
+ 382,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 382,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "the first method to outperform the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 371,
+ 393,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 393,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "human baseline and achieves over",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 371,
+ 403,
+ 506,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 404,
+ 386,
+ 415
+ ],
+ "score": 0.82,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 387,
+ 403,
+ 506,
+ 416
+ ],
+ "score": 1.0,
+ "content": "higher accuracy than the next-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 371,
+ 415,
+ 425,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 415,
+ 425,
+ 426
+ ],
+ "score": 1.0,
+ "content": "best method.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 24.0
+ }
+ ],
+ "index": 20.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 381,
+ 362,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 363,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 363,
+ 394
+ ],
+ "score": 1.0,
+ "content": "Table 1: Accuracy on held-out T0 tasks and computational costs",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 363,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 363,
+ 405
+ ],
+ "score": 1.0,
+ "content": "for different few-shot learning methods and models. T-Few",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 403,
+ 362,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 403,
+ 243,
+ 416
+ ],
+ "score": 1.0,
+ "content": "attains the highest accuracy with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 243,
+ 404,
+ 275,
+ 414
+ ],
+ "score": 0.87,
+ "content": "1 { , } 0 0 0 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 403,
+ 362,
+ 416
+ ],
+ "score": 1.0,
+ "content": "lower computational",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 414,
+ 363,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 414,
+ 363,
+ 426
+ ],
+ "score": 1.0,
+ "content": "cost than ICL with GPT-3 175B. Fine-tuning with T-Few costs",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 425,
+ 336,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 336,
+ 437
+ ],
+ "score": 1.0,
+ "content": "about as much as ICL on 20 examples with GPT-3 175B.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 450,
+ 266,
+ 461
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 267,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 267,
+ 464
+ ],
+ "score": 1.0,
+ "content": "4.2 Comparing computational costs",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 469,
+ 505,
+ 688
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 483
+ ],
+ "score": 1.0,
+ "content": "Having established that T-Few significantly outperforms ICL-based models, we now compare the",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "relative costs of each few-shot learning approach. For simplicity, we use the FLOPs-per-token",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": "estimates for Transformer-based language models introduced by Kaplan et al. [20]. Specifically, we",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 379,
+ 515
+ ],
+ "score": 1.0,
+ "content": "estimate that a decoder-only Transformer (e.g. the GPT series) with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 503,
+ 390,
+ 513
+ ],
+ "score": 0.75,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 390,
+ 502,
+ 457,
+ 515
+ ],
+ "score": 1.0,
+ "content": "parameters uses",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 457,
+ 502,
+ 473,
+ 513
+ ],
+ "score": 0.68,
+ "content": "2 N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 473,
+ 502,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "FLOPs",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 104,
+ 513,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 513,
+ 217,
+ 526
+ ],
+ "score": 1.0,
+ "content": "per token for inference and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 217,
+ 514,
+ 233,
+ 524
+ ],
+ "score": 0.74,
+ "content": "6 N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 513,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "FLOPs per token for training. Encoder-decoder models like T0 and",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "T5 (where the encoder and decoder have the same number of layers and layer sizes) only process",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 535,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "each token with either the encoder or decoder (each having roughly half the parameters of the full",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 505,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 323,
+ 559
+ ],
+ "score": 1.0,
+ "content": "model), so the FLOPs per token estimates are halved to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 547,
+ 334,
+ 556
+ ],
+ "score": 0.81,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 546,
+ 351,
+ 559
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 351,
+ 546,
+ 367,
+ 556
+ ],
+ "score": 0.74,
+ "content": "3 N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 546,
+ 505,
+ 559
+ ],
+ "score": 1.0,
+ "content": "FLOPs per token for inference and",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 557,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "training. We note that FLOPs are not a direct measurement of real-world computational cost because",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 580
+ ],
+ "score": 1.0,
+ "content": "latency, power usage, and other costs can vary significantly depending on hardware and other factors",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "[52]. However, we focus on FLOPs because it is a hardware-independent metric that closely with",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 590,
+ 505,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 505,
+ 602
+ ],
+ "score": 1.0,
+ "content": "real-world costs the hardware setup used for running the different methods we consider would likely",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "vary significantly across methods. We summarize the costs in table 1 and discuss them below. For all",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "estimates, we use the median number of shots (41) across the datasets we consider. Rank evaluation",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "and our unlikelihood loss both require processing every possible output choice to attain a prediction",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "for an unlabeled example. The median combined tokenized sequence length for the input and all",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "possible targets is 103 for the datasets we consider. For in-context examples processed for few-shot",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "ICL, only the correct target is required, producing a median sequence length of 98. Assuming that",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 104,
+ 664,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 664,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": "key and value vectors are cached, processing a single example with ICL therefore involves processing",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 676,
+ 423,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 167,
+ 687
+ ],
+ "score": 0.9,
+ "content": "4 1 \\times 9 8 + 1 0 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 676,
+ 423,
+ 689
+ ],
+ "score": 1.0,
+ "content": "tokens. A summary of our cost estimates is provided in table 1.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 41.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 700,
+ 502,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Inference cost. Beyond improved accuracy, the primary advantage of avoiding few-shot ICL is",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 504,
+ 723
+ ],
+ "score": 1.0,
+ "content": "dramatically lower inference costs. Processing a single input and all target choices with T-Few",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 52.5
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "8",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 70,
+ 288,
+ 85
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 70,
+ 290,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 70,
+ 290,
+ 87
+ ],
+ "score": 1.0,
+ "content": "4 Outperforming ICL with T-Few",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 95,
+ 505,
+ 129
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 95,
+ 505,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 95,
+ 505,
+ 108
+ ],
+ "score": 1.0,
+ "content": "Having designed and established the T-Few recipe on T0-3B, we now apply it to T0 (with 11 billion",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 106,
+ 506,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 506,
+ 119
+ ],
+ "score": 1.0,
+ "content": "parameters) and compare performance to strong few-shot ICL baselines. From this point onwards,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 118,
+ 382,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 382,
+ 129
+ ],
+ "score": 1.0,
+ "content": "we use exactly the same recipe and hyperparameters across all tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 105,
+ 95,
+ 506,
+ 129
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 141,
+ 237,
+ 153
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 237,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 237,
+ 154
+ ],
+ "score": 1.0,
+ "content": "4.1 Performance on T0 tasks",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 161,
+ 505,
+ 271
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 174
+ ],
+ "score": 1.0,
+ "content": "First, we evaluate T-Few on the datasets that were held out from T0’s training mixture. We compare",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 507,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 507,
+ 185
+ ],
+ "score": 1.0,
+ "content": "against zero-shot learning with T0 [1] (since we found few-shot ICL to performed worse than zero-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 182,
+ 506,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 182,
+ 307,
+ 197
+ ],
+ "score": 1.0,
+ "content": "shot for T0, see appendix E); few-shot ICL with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 307,
+ 183,
+ 342,
+ 194
+ ],
+ "score": 0.58,
+ "content": "_ { \\mathrm { T 5 + L M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 182,
+ 506,
+ 197
+ ],
+ "score": 1.0,
+ "content": "[14] (the next-step-prediction language",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 194,
+ 506,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 194,
+ 506,
+ 207
+ ],
+ "score": 1.0,
+ "content": "model upon which T0 is based); and few-shot ICL with the 6.7, 13, and 175 billion parameter variants",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 218
+ ],
+ "score": 1.0,
+ "content": "of GPT-3. See appendix E for more details on these baselines. The accuracy on the held-out T0",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 216,
+ 506,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 216,
+ 506,
+ 228
+ ],
+ "score": 1.0,
+ "content": "datasets (described in section 3.1) is shown in table 1 and fig. 3, with per-dataset results reported",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 104,
+ 226,
+ 507,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 226,
+ 507,
+ 240
+ ],
+ "score": 1.0,
+ "content": "in appendix E. We find that T-Few outperforms all other methods by a substantial margin. Notably,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 504,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 177,
+ 250
+ ],
+ "score": 1.0,
+ "content": "T-Few achieves a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 238,
+ 192,
+ 248
+ ],
+ "score": 0.86,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 193,
+ 237,
+ 484,
+ 250
+ ],
+ "score": 1.0,
+ "content": "higher accuracy than few-shot ICL with GPT-3 175B despite being about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 485,
+ 238,
+ 504,
+ 249
+ ],
+ "score": 0.86,
+ "content": "1 6 \\times",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 248,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "smaller and outperforms the smaller GPT-3 variants by an even larger margin. T-Few also attains",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 259,
+ 502,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 464,
+ 272
+ ],
+ "score": 1.0,
+ "content": "significantly higher accuracy than both zero-shot learning with T0 and few-shot ICL with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 260,
+ 498,
+ 270
+ ],
+ "score": 0.76,
+ "content": "_ { \\mathrm { T 5 + L M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 498,
+ 259,
+ 502,
+ 272
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 104,
+ 160,
+ 507,
+ 272
+ ]
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 355
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 355
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 355
+ ],
+ "score": 0.969,
+ "html": "| Method | Acc. |
| T-Few | 75.8% |
| Human baseline [2] | 73.5% |
| PET[50] | 69.6% |
| SetFit [51] | 66.9% |
| GPT-3 [4] | 62.7% |
",
+ "type": "table",
+ "image_path": "026d7440548c8e41af0452784eff74061f5ca36d54a6d8e830c3887d67b4abe2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 17.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 372,
+ 276,
+ 504,
+ 315.5
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 372,
+ 315.5,
+ 504,
+ 355.0
+ ],
+ "spans": [],
+ "index": 18
+ }
+ ]
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 376
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 376
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 376
+ ],
+ "score": 0.981,
+ "html": "| Method | Inference FLOPs | Training FLOPs | Disk space | Acc. |
| T-Few | 1.1e12 | 2.7e16 | 4.2 MB | 72.4% |
| T0[1] | 1.1e12 | 0 | OB | 66.9% |
| T5+LM[14] | 4.5e13 | 0 | 16 kB | 49.6% |
| GPT-3 6.7B [4] | 5.4e13 | 0 | 16 kB | 57.2% |
| GPT-3 13B [4] | 1.0e14 | 0 | 16 kB | 60.3% |
| GPT-3 175B[4] | 1.4e15 | 0 | 16 kB | 66.6% |
",
+ "type": "table",
+ "image_path": "40b321b5687eae32a98d4d450f37e95533e02b8c81b7852cb682c522236b082d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 276,
+ 360,
+ 309.3333333333333
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 108,
+ 309.3333333333333,
+ 360,
+ 342.66666666666663
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 108,
+ 342.66666666666663,
+ 360,
+ 375.99999999999994
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 371,
+ 360,
+ 505,
+ 426
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 371,
+ 359,
+ 505,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 359,
+ 505,
+ 372
+ ],
+ "score": 1.0,
+ "content": "Table 2: Top-5 best methods on",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 370,
+ 370,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 370,
+ 370,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "RAFT as of writing. T-Few is",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 371,
+ 382,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 382,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "the first method to outperform the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 371,
+ 393,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 393,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "human baseline and achieves over",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 371,
+ 403,
+ 506,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 404,
+ 386,
+ 415
+ ],
+ "score": 0.82,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 387,
+ 403,
+ 506,
+ 416
+ ],
+ "score": 1.0,
+ "content": "higher accuracy than the next-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 371,
+ 415,
+ 425,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 415,
+ 425,
+ 426
+ ],
+ "score": 1.0,
+ "content": "best method.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 24.0
+ }
+ ],
+ "index": 20.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 381,
+ 362,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 363,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 363,
+ 394
+ ],
+ "score": 1.0,
+ "content": "Table 1: Accuracy on held-out T0 tasks and computational costs",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 363,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 363,
+ 405
+ ],
+ "score": 1.0,
+ "content": "for different few-shot learning methods and models. T-Few",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 403,
+ 362,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 403,
+ 243,
+ 416
+ ],
+ "score": 1.0,
+ "content": "attains the highest accuracy with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 243,
+ 404,
+ 275,
+ 414
+ ],
+ "score": 0.87,
+ "content": "1 { , } 0 0 0 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 403,
+ 362,
+ 416
+ ],
+ "score": 1.0,
+ "content": "lower computational",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 414,
+ 363,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 414,
+ 363,
+ 426
+ ],
+ "score": 1.0,
+ "content": "cost than ICL with GPT-3 175B. Fine-tuning with T-Few costs",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 425,
+ 336,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 336,
+ 437
+ ],
+ "score": 1.0,
+ "content": "about as much as ICL on 20 examples with GPT-3 175B.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 105,
+ 381,
+ 363,
+ 437
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 450,
+ 266,
+ 461
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 267,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 267,
+ 464
+ ],
+ "score": 1.0,
+ "content": "4.2 Comparing computational costs",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 469,
+ 505,
+ 688
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 483
+ ],
+ "score": 1.0,
+ "content": "Having established that T-Few significantly outperforms ICL-based models, we now compare the",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "relative costs of each few-shot learning approach. For simplicity, we use the FLOPs-per-token",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": "estimates for Transformer-based language models introduced by Kaplan et al. [20]. Specifically, we",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 379,
+ 515
+ ],
+ "score": 1.0,
+ "content": "estimate that a decoder-only Transformer (e.g. the GPT series) with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 503,
+ 390,
+ 513
+ ],
+ "score": 0.75,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 390,
+ 502,
+ 457,
+ 515
+ ],
+ "score": 1.0,
+ "content": "parameters uses",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 457,
+ 502,
+ 473,
+ 513
+ ],
+ "score": 0.68,
+ "content": "2 N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 473,
+ 502,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "FLOPs",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 104,
+ 513,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 513,
+ 217,
+ 526
+ ],
+ "score": 1.0,
+ "content": "per token for inference and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 217,
+ 514,
+ 233,
+ 524
+ ],
+ "score": 0.74,
+ "content": "6 N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 513,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "FLOPs per token for training. Encoder-decoder models like T0 and",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "T5 (where the encoder and decoder have the same number of layers and layer sizes) only process",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 535,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "each token with either the encoder or decoder (each having roughly half the parameters of the full",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 505,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 323,
+ 559
+ ],
+ "score": 1.0,
+ "content": "model), so the FLOPs per token estimates are halved to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 547,
+ 334,
+ 556
+ ],
+ "score": 0.81,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 546,
+ 351,
+ 559
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 351,
+ 546,
+ 367,
+ 556
+ ],
+ "score": 0.74,
+ "content": "3 N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 546,
+ 505,
+ 559
+ ],
+ "score": 1.0,
+ "content": "FLOPs per token for inference and",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 557,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "training. We note that FLOPs are not a direct measurement of real-world computational cost because",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 580
+ ],
+ "score": 1.0,
+ "content": "latency, power usage, and other costs can vary significantly depending on hardware and other factors",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "[52]. However, we focus on FLOPs because it is a hardware-independent metric that closely with",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 590,
+ 505,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 505,
+ 602
+ ],
+ "score": 1.0,
+ "content": "real-world costs the hardware setup used for running the different methods we consider would likely",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "vary significantly across methods. We summarize the costs in table 1 and discuss them below. For all",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "estimates, we use the median number of shots (41) across the datasets we consider. Rank evaluation",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "and our unlikelihood loss both require processing every possible output choice to attain a prediction",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "for an unlabeled example. The median combined tokenized sequence length for the input and all",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "possible targets is 103 for the datasets we consider. For in-context examples processed for few-shot",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "ICL, only the correct target is required, producing a median sequence length of 98. Assuming that",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 104,
+ 664,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 664,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": "key and value vectors are cached, processing a single example with ICL therefore involves processing",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 676,
+ 423,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 167,
+ 687
+ ],
+ "score": 0.9,
+ "content": "4 1 \\times 9 8 + 1 0 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 676,
+ 423,
+ 689
+ ],
+ "score": 1.0,
+ "content": "tokens. A summary of our cost estimates is provided in table 1.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 41.5,
+ "bbox_fs": [
+ 104,
+ 469,
+ 506,
+ 689
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 700,
+ 502,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Inference cost. Beyond improved accuracy, the primary advantage of avoiding few-shot ICL is",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 504,
+ 723
+ ],
+ "score": 1.0,
+ "content": "dramatically lower inference costs. Processing a single input and all target choices with T-Few",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 140,
+ 85
+ ],
+ "score": 1.0,
+ "content": "requires",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 141,
+ 73,
+ 227,
+ 83
+ ],
+ "score": 0.86,
+ "content": "1 1 \\mathrm { { e 9 \\times 1 0 3 = 1 . 1 \\mathrm { { e 1 2 } } } }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 227,
+ 73,
+ 453,
+ 85
+ ],
+ "score": 1.0,
+ "content": "FLOPs, whereas few-shot ICL with GPT-3 175B requires",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 454,
+ 73,
+ 505,
+ 83
+ ],
+ "score": 0.88,
+ "content": "2 \\times 1 7 5 \\mathrm { e 9 \\times }",
+ "type": "inline_equation",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 213,
+ 95
+ ],
+ "score": 0.66,
+ "content": "( 4 1 \\times 9 8 + 1 0 3 ) = 1 . 4 \\mathrm { { e } 1 5 }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 213,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "FLOPs – more than 3 orders of magnitude more. Inference costs with ICL",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 95,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 95,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "using the smaller GPT-3 variants are also dramatically higher than the inference cost of T-Few. As",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "discussed in section 2.1, caching the key and value vectors of the in-context examples can reduce",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 116,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 116,
+ 443,
+ 129
+ ],
+ "score": 1.0,
+ "content": "the computational cost of ICL. However, this would only result in an approximately",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 443,
+ 117,
+ 462,
+ 127
+ ],
+ "score": 0.86,
+ "content": "4 1 \\times",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 463,
+ 116,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "reduction,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 433,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 433,
+ 139
+ ],
+ "score": 1.0,
+ "content": "which is not nearly enough to make any of the GPT-3 ICL costs as low as T-Few.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 52.5,
+ "bbox_fs": [
+ 105,
+ 699,
+ 505,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 138
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 73,
+ 140,
+ 85
+ ],
+ "score": 1.0,
+ "content": "requires",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 141,
+ 73,
+ 227,
+ 83
+ ],
+ "score": 0.86,
+ "content": "1 1 \\mathrm { { e 9 \\times 1 0 3 = 1 . 1 \\mathrm { { e 1 2 } } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 73,
+ 453,
+ 85
+ ],
+ "score": 1.0,
+ "content": "FLOPs, whereas few-shot ICL with GPT-3 175B requires",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 454,
+ 73,
+ 505,
+ 83
+ ],
+ "score": 0.88,
+ "content": "2 \\times 1 7 5 \\mathrm { e 9 \\times }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 213,
+ 95
+ ],
+ "score": 0.66,
+ "content": "( 4 1 \\times 9 8 + 1 0 3 ) = 1 . 4 \\mathrm { { e } 1 5 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "FLOPs – more than 3 orders of magnitude more. Inference costs with ICL",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 95,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 95,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "using the smaller GPT-3 variants are also dramatically higher than the inference cost of T-Few. As",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "discussed in section 2.1, caching the key and value vectors of the in-context examples can reduce",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 116,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 116,
+ 443,
+ 129
+ ],
+ "score": 1.0,
+ "content": "the computational cost of ICL. However, this would only result in an approximately",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 117,
+ 462,
+ 127
+ ],
+ "score": 0.86,
+ "content": "4 1 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 463,
+ 116,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "reduction,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 433,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 433,
+ 139
+ ],
+ "score": 1.0,
+ "content": "which is not nearly enough to make any of the GPT-3 ICL costs as low as T-Few.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 151,
+ 505,
+ 238
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 150,
+ 505,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 150,
+ 505,
+ 163
+ ],
+ "score": 1.0,
+ "content": "Training cost. Since T-Few is the only method that involves updating parameters, it is the only",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 162,
+ 506,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 162,
+ 506,
+ 174
+ ],
+ "score": 1.0,
+ "content": "method that incurs a training cost. Training an eleven billion parameter encoder-decoder model for",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 419,
+ 186
+ ],
+ "score": 1.0,
+ "content": "1,000 steps with a batch size of 8 length-103 sequences requires approximately",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 173,
+ 506,
+ 184
+ ],
+ "score": 0.63,
+ "content": "3 \\times 1 1 { \\mathrm { e 9 } } \\times 1 , 0 0 0 \\times",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 181,
+ 194
+ ],
+ "score": 0.87,
+ "content": "8 \\times 1 0 3 = 2 . 7 { \\mathrm { e 1 6 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 184,
+ 505,
+ 196
+ ],
+ "score": 1.0,
+ "content": "FLOPs. While not insignificant, this is only about 20 times larger than the FLOPs",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 193,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 193,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "required to process a single example with few-shot ICL using GPT-3 175B. In other words, training",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 205,
+ 505,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 505,
+ 218
+ ],
+ "score": 1.0,
+ "content": "T-Few costs as much as using GPT-3 175B to process 20 examples with few-shot ICL. We also",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 229
+ ],
+ "score": 1.0,
+ "content": "found that fine-tuning T0 with T-Few on a single dataset only takes about a half an hour on a single",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 227,
+ 473,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 339,
+ 239
+ ],
+ "score": 1.0,
+ "content": "NVIDIA A100 GPU. As of writing, this would cost about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 339,
+ 227,
+ 351,
+ 238
+ ],
+ "score": 0.85,
+ "content": "\\$ 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 227,
+ 473,
+ 239
+ ],
+ "score": 1.0,
+ "content": "USD using Microsoft Azure.2",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 251,
+ 505,
+ 317
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 250,
+ 505,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 250,
+ 505,
+ 263
+ ],
+ "score": 1.0,
+ "content": "Storage cost. T-Few also incurs the largest storage cost. When stored as single-precision floats, the",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 261,
+ 505,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 261,
+ 192,
+ 276
+ ],
+ "score": 1.0,
+ "content": "parameters added by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 261,
+ 218,
+ 273
+ ],
+ "score": 0.7,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 261,
+ 250,
+ 276
+ ],
+ "score": 1.0,
+ "content": "take up",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 262,
+ 282,
+ 272
+ ],
+ "score": 0.36,
+ "content": "4 . 2 \\mathrm { M B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 282,
+ 261,
+ 505,
+ 276
+ ],
+ "score": 1.0,
+ "content": "of space on disk. In contrast, ICL methods only require",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 273,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 273,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "storing the tokenized in-context examples (typically stored as 32-bit integers), resulting in a smaller",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 283,
+ 505,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 284,
+ 162,
+ 294
+ ],
+ "score": 0.89,
+ "content": "4 1 \\times 9 8 \\times 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 283,
+ 179,
+ 296
+ ],
+ "score": 1.0,
+ "content": "bits",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 179,
+ 284,
+ 216,
+ 294
+ ],
+ "score": 0.45,
+ "content": "= 1 6 \\mathrm { k B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 283,
+ 402,
+ 296
+ ],
+ "score": 1.0,
+ "content": "disk space requirement. However, we note that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 284,
+ 434,
+ 294
+ ],
+ "score": 0.47,
+ "content": "4 . 2 \\mathrm { M B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 283,
+ 505,
+ 296
+ ],
+ "score": 1.0,
+ "content": "is dwarfed by the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 294,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 294,
+ 360,
+ 307
+ ],
+ "score": 1.0,
+ "content": "on-disk size of the model checkpoints themselves – storing the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 360,
+ 294,
+ 385,
+ 305
+ ],
+ "score": 0.62,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 294,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "adaptation vectors for 10,000",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 306,
+ 390,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 306,
+ 390,
+ 317
+ ],
+ "score": 1.0,
+ "content": "tasks would take about as much space as the T0 checkpoint (41.5 GB).",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 329,
+ 505,
+ 384
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 506,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 506,
+ 342
+ ],
+ "score": 1.0,
+ "content": "Memory usage. During inference, the primary memory cost is incurred by the model’s parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 339,
+ 506,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 506,
+ 353
+ ],
+ "score": 1.0,
+ "content": "The only model smaller than T0 (used by T-Few) is GPT-3 6.7B; otherwise, T-Few will incur a lower",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 364
+ ],
+ "score": 1.0,
+ "content": "memory cost during inference. Additional memory costs are incurred when training T-Few due to the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 362,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "need to cache intermediate activations for backpropagation and for the gradient accumulator variables",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 479,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 479,
+ 386
+ ],
+ "score": 1.0,
+ "content": "in Adafactor. However, as mentioned above, a single 80GB A100 GPU is enough for T-Few.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 397,
+ 352,
+ 409
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 353,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 353,
+ 411
+ ],
+ "score": 1.0,
+ "content": "4.3 Performance on Real-world Few-shot Tasks (RAFT)",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 418,
+ 506,
+ 538
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 431
+ ],
+ "score": 1.0,
+ "content": "So far, we have run evaluatation on a collection of datasets not explicitly designed for benchmarking",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "few-shot learning. To better evaluate T-Few in the real world, we took our approach to the RAFT",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "benchmark [2]. RAFT comprise 11 “economically valuable” tasks that mirror real-world applications.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "Importantly, each RAFT datasets has only 50 training examples with no validation set and a (larger)",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 460,
+ 506,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 460,
+ 506,
+ 476
+ ],
+ "score": 1.0,
+ "content": "test set with no public labels, so it is impossible to “cheat” by tuning on an unrealistically-large",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 473,
+ 506,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 506,
+ 486
+ ],
+ "score": 1.0,
+ "content": "validation set or by peeking at the test set [32, 31]. We apply T-Few to RAFT by using the standard",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "score": 1.0,
+ "content": "prompts released alongside the dataset. The accuracy of the current top-5 methods is shown in table 2,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 411,
+ 507
+ ],
+ "score": 1.0,
+ "content": "with further details in appendix G. T-Few attains a state-of-the-art accuracy of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 495,
+ 438,
+ 505
+ ],
+ "score": 0.87,
+ "content": "7 5 . 8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 495,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "and outperforms",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 186,
+ 518
+ ],
+ "score": 1.0,
+ "content": "the human baseline",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 187,
+ 506,
+ 213,
+ 516
+ ],
+ "score": 0.81,
+ "content": "7 3 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 505,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "accuracy) for the first time. The next-best model (from Schick and Schütze",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 165,
+ 528
+ ],
+ "score": 1.0,
+ "content": "[50]) achieves",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 516,
+ 181,
+ 527
+ ],
+ "score": 0.86,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 516,
+ 365,
+ 528
+ ],
+ "score": 1.0,
+ "content": "lower accuracy and GPT-3 175B attains only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 516,
+ 393,
+ 527
+ ],
+ "score": 0.87,
+ "content": "6 2 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 393,
+ 516,
+ 505,
+ 528
+ ],
+ "score": 1.0,
+ "content": ". These results validate that",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 444,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 444,
+ 541
+ ],
+ "score": 1.0,
+ "content": "T-Few can be readily applied as-is to novel real-world tasks for strong performance.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 552,
+ 222,
+ 564
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 223,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 223,
+ 566
+ ],
+ "score": 1.0,
+ "content": "4.4 Ablation experiments",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 639
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "Given that our T-Few design experiments were on T0-3B, we perform an ablation of some of the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "ingredients of T-Few on T0. Results are shown in appendix F. While the gains from adding each",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "ingredient does not always significant increase the accuracy on each individual dataset, each ingredient",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "consistently improves the average performance across datasets: Removing pre-training decreases",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 104,
+ 615,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 615,
+ 157,
+ 630
+ ],
+ "score": 1.0,
+ "content": "accuracy by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 617,
+ 179,
+ 627
+ ],
+ "score": 0.85,
+ "content": "1 . 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 180,
+ 615,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": ", removing unlikelihood training and length normalization decreases accuracy by",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 486,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 128,
+ 638
+ ],
+ "score": 0.85,
+ "content": "4 . 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 128,
+ 627,
+ 459,
+ 639
+ ],
+ "score": 1.0,
+ "content": ", and removing both pre-training and our additional loss terms reduces accuracy by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 460,
+ 627,
+ 482,
+ 638
+ ],
+ "score": 0.85,
+ "content": "2 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 627,
+ 486,
+ 639
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 654,
+ 197,
+ 668
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 198,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 198,
+ 669
+ ],
+ "score": 1.0,
+ "content": "5 Related Work",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 680,
+ 504,
+ 702
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 680,
+ 504,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 680,
+ 504,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Currently, prompt tuning is one of the most parameter-efficient methods for large language models",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 690,
+ 506,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 690,
+ 506,
+ 703
+ ],
+ "score": 1.0,
+ "content": "[29, 14, 53]. Liu et al. [54] introduce several tricks to improve prompt tuning, An et al. [55] tune",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45.5
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 115,
+ 712,
+ 475,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 709,
+ 476,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 709,
+ 476,
+ 723
+ ],
+ "score": 1.0,
+ "content": "2https://docs.microsoft.com/en-us/azure/virtual-machines/ndm-a100-v4-series",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 138
+ ],
+ "lines": [],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 73,
+ 506,
+ 139
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 151,
+ 505,
+ 238
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 150,
+ 505,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 150,
+ 505,
+ 163
+ ],
+ "score": 1.0,
+ "content": "Training cost. Since T-Few is the only method that involves updating parameters, it is the only",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 162,
+ 506,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 162,
+ 506,
+ 174
+ ],
+ "score": 1.0,
+ "content": "method that incurs a training cost. Training an eleven billion parameter encoder-decoder model for",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 419,
+ 186
+ ],
+ "score": 1.0,
+ "content": "1,000 steps with a batch size of 8 length-103 sequences requires approximately",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 173,
+ 506,
+ 184
+ ],
+ "score": 0.63,
+ "content": "3 \\times 1 1 { \\mathrm { e 9 } } \\times 1 , 0 0 0 \\times",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 181,
+ 194
+ ],
+ "score": 0.87,
+ "content": "8 \\times 1 0 3 = 2 . 7 { \\mathrm { e 1 6 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 184,
+ 505,
+ 196
+ ],
+ "score": 1.0,
+ "content": "FLOPs. While not insignificant, this is only about 20 times larger than the FLOPs",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 193,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 193,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "required to process a single example with few-shot ICL using GPT-3 175B. In other words, training",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 205,
+ 505,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 505,
+ 218
+ ],
+ "score": 1.0,
+ "content": "T-Few costs as much as using GPT-3 175B to process 20 examples with few-shot ICL. We also",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 229
+ ],
+ "score": 1.0,
+ "content": "found that fine-tuning T0 with T-Few on a single dataset only takes about a half an hour on a single",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 227,
+ 473,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 339,
+ 239
+ ],
+ "score": 1.0,
+ "content": "NVIDIA A100 GPU. As of writing, this would cost about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 339,
+ 227,
+ 351,
+ 238
+ ],
+ "score": 0.85,
+ "content": "\\$ 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 227,
+ 473,
+ 239
+ ],
+ "score": 1.0,
+ "content": "USD using Microsoft Azure.2",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 104,
+ 150,
+ 506,
+ 239
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 251,
+ 505,
+ 317
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 250,
+ 505,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 250,
+ 505,
+ 263
+ ],
+ "score": 1.0,
+ "content": "Storage cost. T-Few also incurs the largest storage cost. When stored as single-precision floats, the",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 261,
+ 505,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 261,
+ 192,
+ 276
+ ],
+ "score": 1.0,
+ "content": "parameters added by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 261,
+ 218,
+ 273
+ ],
+ "score": 0.7,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 261,
+ 250,
+ 276
+ ],
+ "score": 1.0,
+ "content": "take up",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 262,
+ 282,
+ 272
+ ],
+ "score": 0.36,
+ "content": "4 . 2 \\mathrm { M B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 282,
+ 261,
+ 505,
+ 276
+ ],
+ "score": 1.0,
+ "content": "of space on disk. In contrast, ICL methods only require",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 273,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 273,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "storing the tokenized in-context examples (typically stored as 32-bit integers), resulting in a smaller",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 283,
+ 505,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 284,
+ 162,
+ 294
+ ],
+ "score": 0.89,
+ "content": "4 1 \\times 9 8 \\times 3 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 283,
+ 179,
+ 296
+ ],
+ "score": 1.0,
+ "content": "bits",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 179,
+ 284,
+ 216,
+ 294
+ ],
+ "score": 0.45,
+ "content": "= 1 6 \\mathrm { k B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 283,
+ 402,
+ 296
+ ],
+ "score": 1.0,
+ "content": "disk space requirement. However, we note that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 284,
+ 434,
+ 294
+ ],
+ "score": 0.47,
+ "content": "4 . 2 \\mathrm { M B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 283,
+ 505,
+ 296
+ ],
+ "score": 1.0,
+ "content": "is dwarfed by the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 294,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 294,
+ 360,
+ 307
+ ],
+ "score": 1.0,
+ "content": "on-disk size of the model checkpoints themselves – storing the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 360,
+ 294,
+ 385,
+ 305
+ ],
+ "score": 0.62,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 294,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "adaptation vectors for 10,000",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 306,
+ 390,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 306,
+ 390,
+ 317
+ ],
+ "score": 1.0,
+ "content": "tasks would take about as much space as the T0 checkpoint (41.5 GB).",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 104,
+ 250,
+ 505,
+ 317
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 329,
+ 505,
+ 384
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 506,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 506,
+ 342
+ ],
+ "score": 1.0,
+ "content": "Memory usage. During inference, the primary memory cost is incurred by the model’s parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 339,
+ 506,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 506,
+ 353
+ ],
+ "score": 1.0,
+ "content": "The only model smaller than T0 (used by T-Few) is GPT-3 6.7B; otherwise, T-Few will incur a lower",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 364
+ ],
+ "score": 1.0,
+ "content": "memory cost during inference. Additional memory costs are incurred when training T-Few due to the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 362,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "need to cache intermediate activations for backpropagation and for the gradient accumulator variables",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 479,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 479,
+ 386
+ ],
+ "score": 1.0,
+ "content": "in Adafactor. However, as mentioned above, a single 80GB A100 GPU is enough for T-Few.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 330,
+ 506,
+ 386
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 397,
+ 352,
+ 409
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 353,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 353,
+ 411
+ ],
+ "score": 1.0,
+ "content": "4.3 Performance on Real-world Few-shot Tasks (RAFT)",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 418,
+ 506,
+ 538
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 431
+ ],
+ "score": 1.0,
+ "content": "So far, we have run evaluatation on a collection of datasets not explicitly designed for benchmarking",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "few-shot learning. To better evaluate T-Few in the real world, we took our approach to the RAFT",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "benchmark [2]. RAFT comprise 11 “economically valuable” tasks that mirror real-world applications.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "Importantly, each RAFT datasets has only 50 training examples with no validation set and a (larger)",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 460,
+ 506,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 460,
+ 506,
+ 476
+ ],
+ "score": 1.0,
+ "content": "test set with no public labels, so it is impossible to “cheat” by tuning on an unrealistically-large",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 473,
+ 506,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 506,
+ 486
+ ],
+ "score": 1.0,
+ "content": "validation set or by peeking at the test set [32, 31]. We apply T-Few to RAFT by using the standard",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "score": 1.0,
+ "content": "prompts released alongside the dataset. The accuracy of the current top-5 methods is shown in table 2,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 411,
+ 507
+ ],
+ "score": 1.0,
+ "content": "with further details in appendix G. T-Few attains a state-of-the-art accuracy of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 495,
+ 438,
+ 505
+ ],
+ "score": 0.87,
+ "content": "7 5 . 8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 495,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "and outperforms",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 186,
+ 518
+ ],
+ "score": 1.0,
+ "content": "the human baseline",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 187,
+ 506,
+ 213,
+ 516
+ ],
+ "score": 0.81,
+ "content": "7 3 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 505,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "accuracy) for the first time. The next-best model (from Schick and Schütze",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 165,
+ 528
+ ],
+ "score": 1.0,
+ "content": "[50]) achieves",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 516,
+ 181,
+ 527
+ ],
+ "score": 0.86,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 516,
+ 365,
+ 528
+ ],
+ "score": 1.0,
+ "content": "lower accuracy and GPT-3 175B attains only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 516,
+ 393,
+ 527
+ ],
+ "score": 0.87,
+ "content": "6 2 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 393,
+ 516,
+ 505,
+ 528
+ ],
+ "score": 1.0,
+ "content": ". These results validate that",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 444,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 444,
+ 541
+ ],
+ "score": 1.0,
+ "content": "T-Few can be readily applied as-is to novel real-world tasks for strong performance.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 104,
+ 417,
+ 506,
+ 541
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 552,
+ 222,
+ 564
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 223,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 223,
+ 566
+ ],
+ "score": 1.0,
+ "content": "4.4 Ablation experiments",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 639
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "Given that our T-Few design experiments were on T0-3B, we perform an ablation of some of the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "ingredients of T-Few on T0. Results are shown in appendix F. While the gains from adding each",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "ingredient does not always significant increase the accuracy on each individual dataset, each ingredient",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "consistently improves the average performance across datasets: Removing pre-training decreases",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 104,
+ 615,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 615,
+ 157,
+ 630
+ ],
+ "score": 1.0,
+ "content": "accuracy by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 617,
+ 179,
+ 627
+ ],
+ "score": 0.85,
+ "content": "1 . 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 180,
+ 615,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": ", removing unlikelihood training and length normalization decreases accuracy by",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 486,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 128,
+ 638
+ ],
+ "score": 0.85,
+ "content": "4 . 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 128,
+ 627,
+ 459,
+ 639
+ ],
+ "score": 1.0,
+ "content": ", and removing both pre-training and our additional loss terms reduces accuracy by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 460,
+ 627,
+ 482,
+ 638
+ ],
+ "score": 0.85,
+ "content": "2 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 627,
+ 486,
+ 639
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 104,
+ 573,
+ 506,
+ 639
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 654,
+ 197,
+ 668
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 198,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 198,
+ 669
+ ],
+ "score": 1.0,
+ "content": "5 Related Work",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 680,
+ 504,
+ 702
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 680,
+ 504,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 680,
+ 504,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Currently, prompt tuning is one of the most parameter-efficient methods for large language models",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 690,
+ 506,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 690,
+ 506,
+ 703
+ ],
+ "score": 1.0,
+ "content": "[29, 14, 53]. Liu et al. [54] introduce several tricks to improve prompt tuning, An et al. [55] tune",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 72,
+ 505,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 505,
+ 86
+ ],
+ "score": 1.0,
+ "content": "prompts along with input embeddings for boost in performance, and Chen et al. [56] improve prompt",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 97
+ ],
+ "score": 1.0,
+ "content": "embeddings through continued pre-training. Given optimization difficulties when training prompt",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 94,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 94,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "embeddings, Diao et al. [57] recently used black-box optimization to train prompt embeddings",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "without requiring gradients. Several works have analyzed prompt tuning from the perspective of",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 116,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 116,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "interpretability Khashabi et al. [58] and its similarity to other PEFT methods He et al. [30]. Prompt",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "tuning has been applied to various applications for NLP including continual learning [59], model",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 152
+ ],
+ "score": 1.0,
+ "content": "robustness [60, 61], summarization [62], machine translation [63], co-training [64], probing language",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "models [65, 65], inverse prompting [66] and transfer learning [67]. He et al. [68] recently proposed",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "the use of a hypernetwork to predict prompts for new tasks (rather than training the prompt parameters",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "with gradient descent). Prompt tuning and other PEFT methods have also been explored outside of",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 182,
+ 471,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 471,
+ 194
+ ],
+ "score": 1.0,
+ "content": "the context of language models (e.g. vision [22, 69] and vision-and-language models [26]).",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 45.5,
+ "bbox_fs": [
+ 106,
+ 680,
+ 506,
+ 703
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 72,
+ 505,
+ 193
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 505,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 505,
+ 86
+ ],
+ "score": 1.0,
+ "content": "prompts along with input embeddings for boost in performance, and Chen et al. [56] improve prompt",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 506,
+ 97
+ ],
+ "score": 1.0,
+ "content": "embeddings through continued pre-training. Given optimization difficulties when training prompt",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 94,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 94,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "embeddings, Diao et al. [57] recently used black-box optimization to train prompt embeddings",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "without requiring gradients. Several works have analyzed prompt tuning from the perspective of",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 116,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 116,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "interpretability Khashabi et al. [58] and its similarity to other PEFT methods He et al. [30]. Prompt",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "tuning has been applied to various applications for NLP including continual learning [59], model",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 152
+ ],
+ "score": 1.0,
+ "content": "robustness [60, 61], summarization [62], machine translation [63], co-training [64], probing language",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "models [65, 65], inverse prompting [66] and transfer learning [67]. He et al. [68] recently proposed",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "the use of a hypernetwork to predict prompts for new tasks (rather than training the prompt parameters",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "with gradient descent). Prompt tuning and other PEFT methods have also been explored outside of",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 182,
+ 471,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 471,
+ 194
+ ],
+ "score": 1.0,
+ "content": "the context of language models (e.g. vision [22, 69] and vision-and-language models [26]).",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 198,
+ 504,
+ 264
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 197,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 197,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "Separately, various studies have considered few-shot full-model fine-tuning with discrete prompts",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "[70]. Recent work has analyzed training with discrete prompts, demonstrating a boost in performance",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "with prompting when training on various numbers of examples [71], finding that models perform",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 244
+ ],
+ "score": 1.0,
+ "content": "similarly when trained on good and bad prompts [11], and exploring which prompts work well for",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 254
+ ],
+ "score": 1.0,
+ "content": "few-shot and full-shot setting [72]. There have also been efforts to develop methods that find discrete",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 443,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 443,
+ 266
+ ],
+ "score": 1.0,
+ "content": "prompts [73, 74] and training prompts using methods similar to prompt tuning [75].",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 269,
+ 504,
+ 324
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "There has also been a great deal of work on improving ICL. Chen et al. [5], Min et al. [6] use ICL for",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "meta-learning to perform few-shot learning on new tasks. Lampinen et al. [7] show ICL can improve",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "when explanations are provided and [8] use ICL with text retrieved from the web for open-domain",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "score": 1.0,
+ "content": "question-answering. Meanwhile, Min et al. [9] analyze how ICL works and show that ICL can still",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 414,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 414,
+ 325
+ ],
+ "score": 1.0,
+ "content": "perform well when incorrect labels are provided for the in-context examples.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 329,
+ 505,
+ 449
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 328,
+ 506,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 328,
+ 506,
+ 342
+ ],
+ "score": 1.0,
+ "content": "With the advent of billion-parameter language models, there has been a great deal of recent interest",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 352
+ ],
+ "score": 1.0,
+ "content": "in PEFT methods and their compatibility in the few-shot setting. Mahabadi et al. [28] found that",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "PEFT outperforms standard fine-tuning in the low-resource setting. In concurrent work, Mahabadi",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 104,
+ 361,
+ 506,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 361,
+ 506,
+ 376
+ ],
+ "score": 1.0,
+ "content": "et al. [76] find that PEFT compares favorably in few-shot fine-tuning against discrete prompts (e.g.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 104,
+ 371,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 371,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "PET [70]). Also concurrently, Moosavi et al. [77] propose a framework for introducing adapters",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "whose architecture and design vary from task to task and demonstrate improved results in few-shot",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "settings. Gu et al. [18] and Vu et al. [19] both explored how pre-training prompt tuning parameters",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": "can improve when limited labeled data is available. For few-shot learning, Triantafillou et al. [78]",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "explore learning universal and dataset dependent parameters that can be blended for generalization.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 440
+ ],
+ "score": 1.0,
+ "content": "Requeima et al. [79] use conditional neural adaptive processes and Li et al. [80] leverage distillation",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 449,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 449,
+ 451
+ ],
+ "score": 1.0,
+ "content": "from multiple feature extractors to learn new classes or domains in few-shot learning.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 466,
+ 182,
+ 480
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 185,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 185,
+ 483
+ ],
+ "score": 1.0,
+ "content": "6 Conclusion",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 492,
+ 505,
+ 635
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "We introduced T-Few, a parameter-efficient few-shot learning recipe that attains higher accuracy than",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 503,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 503,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "few-shot ICL at a lower computational cost. T-Few uses (IA)3, a new PEFT method that rescales",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 285,
+ 527
+ ],
+ "score": 1.0,
+ "content": "inner activations with learned vectors. Using",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 514,
+ 310,
+ 525
+ ],
+ "score": 0.34,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 311,
+ 513,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "produces better performance than fine-tuning the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "full model while introducing minimal additional parameters. T-Few also uses two additional loss",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 537,
+ 505,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 537,
+ 505,
+ 548
+ ],
+ "score": 1.0,
+ "content": "terms that encourage the model to output lower probabilities for incorrect choices and account for the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "length of different answer choices. When applying T-Few as-is (with no task-specific hyperparameter",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 571
+ ],
+ "score": 1.0,
+ "content": "tuning or other changes) to the RAFT benchmark, we attained super-human performance for the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "first time and outperformed prior submissions by a large margin. Through detailed characterization",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 328,
+ 592
+ ],
+ "score": 1.0,
+ "content": "of computational costs, we found that T-Few uses over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 580,
+ 360,
+ 591
+ ],
+ "score": 0.87,
+ "content": "1 { , } 0 0 0 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 360,
+ 580,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "fewer FLOPs during inference than",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "few-shot ICL with GPT-3 and only requires 30 minutes to train on a single NVIDIA A100 GPU.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 602,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 602,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Since all of our experiments were on classification tasks, we are interested in applying T-Few to",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 626
+ ],
+ "score": 1.0,
+ "content": "generative tasks like as summarization and question answering in future work. We hope our results",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 495,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 495,
+ 637
+ ],
+ "score": 1.0,
+ "content": "provide a new perspective on how best to perform few-shot learning with large language models.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 651,
+ 339,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 341,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 341,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Acknowledgments and Disclosure of Funding",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 47
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 678,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 691
+ ],
+ "score": 1.0,
+ "content": "We thank Brian Lester and Noah Constant for helpful discussion on debugging prompt tuning and",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Rabeeh Karimi Mahabadi for help with Compacter and Intrinsic SAID. We also thank Stella Biderman",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "and the Google TPU Research Cloud who provided valuable computational resources to support this",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 412,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 412,
+ 723
+ ],
+ "score": 1.0,
+ "content": "work. This work was supported by NSF-AI Engage Institute DRL-2112635.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49.5
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 301,
+ 742,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 14
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 72,
+ 505,
+ 193
+ ],
+ "lines": [],
+ "index": 5,
+ "bbox_fs": [
+ 105,
+ 72,
+ 506,
+ 194
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 198,
+ 504,
+ 264
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 197,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 197,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "Separately, various studies have considered few-shot full-model fine-tuning with discrete prompts",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "[70]. Recent work has analyzed training with discrete prompts, demonstrating a boost in performance",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "with prompting when training on various numbers of examples [71], finding that models perform",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 244
+ ],
+ "score": 1.0,
+ "content": "similarly when trained on good and bad prompts [11], and exploring which prompts work well for",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 254
+ ],
+ "score": 1.0,
+ "content": "few-shot and full-shot setting [72]. There have also been efforts to develop methods that find discrete",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 443,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 443,
+ 266
+ ],
+ "score": 1.0,
+ "content": "prompts [73, 74] and training prompts using methods similar to prompt tuning [75].",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 105,
+ 197,
+ 506,
+ 266
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 269,
+ 504,
+ 324
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "There has also been a great deal of work on improving ICL. Chen et al. [5], Min et al. [6] use ICL for",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "meta-learning to perform few-shot learning on new tasks. Lampinen et al. [7] show ICL can improve",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "when explanations are provided and [8] use ICL with text retrieved from the web for open-domain",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "score": 1.0,
+ "content": "question-answering. Meanwhile, Min et al. [9] analyze how ICL works and show that ICL can still",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 414,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 414,
+ 325
+ ],
+ "score": 1.0,
+ "content": "perform well when incorrect labels are provided for the in-context examples.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 269,
+ 506,
+ 325
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 329,
+ 505,
+ 449
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 328,
+ 506,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 328,
+ 506,
+ 342
+ ],
+ "score": 1.0,
+ "content": "With the advent of billion-parameter language models, there has been a great deal of recent interest",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 352
+ ],
+ "score": 1.0,
+ "content": "in PEFT methods and their compatibility in the few-shot setting. Mahabadi et al. [28] found that",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "PEFT outperforms standard fine-tuning in the low-resource setting. In concurrent work, Mahabadi",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 104,
+ 361,
+ 506,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 361,
+ 506,
+ 376
+ ],
+ "score": 1.0,
+ "content": "et al. [76] find that PEFT compares favorably in few-shot fine-tuning against discrete prompts (e.g.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 104,
+ 371,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 371,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "PET [70]). Also concurrently, Moosavi et al. [77] propose a framework for introducing adapters",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "whose architecture and design vary from task to task and demonstrate improved results in few-shot",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "settings. Gu et al. [18] and Vu et al. [19] both explored how pre-training prompt tuning parameters",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": "can improve when limited labeled data is available. For few-shot learning, Triantafillou et al. [78]",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "explore learning universal and dataset dependent parameters that can be blended for generalization.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 440
+ ],
+ "score": 1.0,
+ "content": "Requeima et al. [79] use conditional neural adaptive processes and Li et al. [80] leverage distillation",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 449,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 449,
+ 451
+ ],
+ "score": 1.0,
+ "content": "from multiple feature extractors to learn new classes or domains in few-shot learning.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 104,
+ 328,
+ 506,
+ 451
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 466,
+ 182,
+ 480
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 185,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 185,
+ 483
+ ],
+ "score": 1.0,
+ "content": "6 Conclusion",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 492,
+ 505,
+ 635
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "We introduced T-Few, a parameter-efficient few-shot learning recipe that attains higher accuracy than",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 503,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 503,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "few-shot ICL at a lower computational cost. T-Few uses (IA)3, a new PEFT method that rescales",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 285,
+ 527
+ ],
+ "score": 1.0,
+ "content": "inner activations with learned vectors. Using",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 514,
+ 310,
+ 525
+ ],
+ "score": 0.34,
+ "content": "( \\mathtt { I A } ) ^ { 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 311,
+ 513,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "produces better performance than fine-tuning the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "full model while introducing minimal additional parameters. T-Few also uses two additional loss",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 537,
+ 505,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 537,
+ 505,
+ 548
+ ],
+ "score": 1.0,
+ "content": "terms that encourage the model to output lower probabilities for incorrect choices and account for the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "length of different answer choices. When applying T-Few as-is (with no task-specific hyperparameter",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 571
+ ],
+ "score": 1.0,
+ "content": "tuning or other changes) to the RAFT benchmark, we attained super-human performance for the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "first time and outperformed prior submissions by a large margin. Through detailed characterization",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 328,
+ 592
+ ],
+ "score": 1.0,
+ "content": "of computational costs, we found that T-Few uses over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 580,
+ 360,
+ 591
+ ],
+ "score": 0.87,
+ "content": "1 { , } 0 0 0 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 360,
+ 580,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "fewer FLOPs during inference than",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "few-shot ICL with GPT-3 and only requires 30 minutes to train on a single NVIDIA A100 GPU.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 602,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 602,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Since all of our experiments were on classification tasks, we are interested in applying T-Few to",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 626
+ ],
+ "score": 1.0,
+ "content": "generative tasks like as summarization and question answering in future work. We hope our results",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 495,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 495,
+ 637
+ ],
+ "score": 1.0,
+ "content": "provide a new perspective on how best to perform few-shot learning with large language models.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 105,
+ 492,
+ 506,
+ 637
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 651,
+ 339,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 341,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 341,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Acknowledgments and Disclosure of Funding",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 47
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 678,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 691
+ ],
+ "score": 1.0,
+ "content": "We thank Brian Lester and Noah Constant for helpful discussion on debugging prompt tuning and",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Rabeeh Karimi Mahabadi for help with Compacter and Intrinsic SAID. We also thank Stella Biderman",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "and the Google TPU Research Cloud who provided valuable computational resources to support this",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 412,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 412,
+ 723
+ ],
+ "score": 1.0,
+ "content": "work. This work was supported by NSF-AI Engage Institute DRL-2112635.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49.5,
+ "bbox_fs": [
+ 105,
+ 677,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 60,
+ 507,
+ 726
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "score": 1.0,
+ "content": "References",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 109,
+ 90,
+ 506,
+ 103
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 90,
+ 506,
+ 103
+ ],
+ "score": 1.0,
+ "content": "[1] Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 125,
+ 99,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 99,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": "Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 126,
+ 112,
+ 442,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 112,
+ 442,
+ 125
+ ],
+ "score": 1.0,
+ "content": "enables zero-shot task generalization. arXiv preprint arXiv:2110.08207, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 111,
+ 130,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 130,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "[2] Neel Alex, Eli Lifland, Lewis Tunstall, Abhishek Thakur, Pegah Maham, C Jess Riedel, Emmie",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 127,
+ 141,
+ 506,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 141,
+ 506,
+ 155
+ ],
+ "score": 1.0,
+ "content": "Hine, Carolyn Ashurst, Paul Sedille, Alexis Carlier, et al. RAFT: A real-world few-shot text",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 126,
+ 152,
+ 396,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 152,
+ 396,
+ 165
+ ],
+ "score": 1.0,
+ "content": "classification benchmark. arXiv preprint arXiv:2109.14076, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 110,
+ 170,
+ 507,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 170,
+ 507,
+ 184
+ ],
+ "score": 1.0,
+ "content": "[3] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 126,
+ 182,
+ 429,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 182,
+ 429,
+ 194
+ ],
+ "score": 1.0,
+ "content": "Language models are unsupervised multitask learners. OpenAI blog, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 109,
+ 199,
+ 507,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 199,
+ 507,
+ 214
+ ],
+ "score": 1.0,
+ "content": "[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 127,
+ 211,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 211,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 127,
+ 222,
+ 366,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 222,
+ 366,
+ 236
+ ],
+ "score": 1.0,
+ "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 109,
+ 239,
+ 506,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 239,
+ 506,
+ 256
+ ],
+ "score": 1.0,
+ "content": "[5] Yanda Chen, Ruiqi Zhong, Sheng Zha, George Karypis, and He He. Meta-learning via language",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 126,
+ 252,
+ 393,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 252,
+ 393,
+ 264
+ ],
+ "score": 1.0,
+ "content": "model in-context tuning. arXiv preprint arXiv:2110.07814, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 110,
+ 270,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 270,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "[6] Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. Metaicl: Learning to",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 127,
+ 281,
+ 360,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 281,
+ 360,
+ 295
+ ],
+ "score": 1.0,
+ "content": "learn in context. arXiv preprint arXiv:2110.15943, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 110,
+ 299,
+ 506,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 299,
+ 506,
+ 313
+ ],
+ "score": 1.0,
+ "content": "[7] Andrew Kyle Lampinen, Ishita Dasgupta, Stephanie C. Y. Chan, Kory Matthewson,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 127,
+ 310,
+ 506,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 310,
+ 506,
+ 324
+ ],
+ "score": 1.0,
+ "content": "Michael Henry Tessler, Antonia Creswell, James L. McClelland, Jane X. Wang, and Felix",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 127,
+ 321,
+ 505,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 321,
+ 505,
+ 335
+ ],
+ "score": 1.0,
+ "content": "Hill. Can language models learn from explanations in context? ArXiv, abs/2204.02329, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 111,
+ 340,
+ 506,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 340,
+ 506,
+ 354
+ ],
+ "score": 1.0,
+ "content": "[8] Angeliki Lazaridou, Elena Gribovskaya, Wojciech Stokowiec, and Nikolai Grigorev. Internet-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 124,
+ 349,
+ 508,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 349,
+ 508,
+ 366
+ ],
+ "score": 1.0,
+ "content": "augmented language models through few-shot prompting for open-domain question answering.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 127,
+ 362,
+ 291,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 362,
+ 291,
+ 376
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2203.05115, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 110,
+ 380,
+ 506,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 380,
+ 506,
+ 394
+ ],
+ "score": 1.0,
+ "content": "[9] Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 125,
+ 389,
+ 506,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 389,
+ 506,
+ 407
+ ],
+ "score": 1.0,
+ "content": "Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 127,
+ 402,
+ 318,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 402,
+ 318,
+ 416
+ ],
+ "score": 1.0,
+ "content": "work? arXiv preprint arXiv:2202.12837, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 434
+ ],
+ "score": 1.0,
+ "content": "[10] Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 126,
+ 431,
+ 507,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 431,
+ 507,
+ 445
+ ],
+ "score": 1.0,
+ "content": "Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 127,
+ 442,
+ 506,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 442,
+ 378,
+ 456
+ ],
+ "score": 1.0,
+ "content": "Benchmarking generalization via in-context instructions on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 378,
+ 443,
+ 408,
+ 453
+ ],
+ "score": 0.82,
+ "content": "1 { , } 6 0 0 { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 409,
+ 442,
+ 506,
+ 456
+ ],
+ "score": 1.0,
+ "content": "language tasks. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 124,
+ 453,
+ 267,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 453,
+ 267,
+ 468
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2204.07705, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 471,
+ 506,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 471,
+ 506,
+ 486
+ ],
+ "score": 1.0,
+ "content": "[11] Albert Webson and Ellie Pavlick. Do prompt-based models really understand the meaning of",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 127,
+ 483,
+ 353,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 483,
+ 353,
+ 496
+ ],
+ "score": 1.0,
+ "content": "their prompts? arXiv preprint arXiv:2109.01247, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 506,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 506,
+ 515
+ ],
+ "score": 1.0,
+ "content": "[12] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use:",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 127,
+ 513,
+ 507,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 513,
+ 507,
+ 527
+ ],
+ "score": 1.0,
+ "content": "Improving few-shot performance of language models. arXiv preprint arXiv:2102.09690, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 529,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 529,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 127,
+ 542,
+ 506,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 542,
+ 506,
+ 554
+ ],
+ "score": 1.0,
+ "content": "Weizhu Chen. LoRA: Low-rank adaptation of large language models. ArXiv, abs/2106.09685,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 126,
+ 552,
+ 155,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 552,
+ 155,
+ 567
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "[14] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 126,
+ 582,
+ 353,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 582,
+ 353,
+ 595
+ ],
+ "score": 1.0,
+ "content": "prompt tuning. arXiv preprint arXiv:2104.08691, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 614
+ ],
+ "score": 1.0,
+ "content": "[15] Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena,",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 128,
+ 612,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 612,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 126,
+ 622,
+ 349,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 622,
+ 349,
+ 635
+ ],
+ "score": 1.0,
+ "content": "text-to-text transformer. ArXiv, abs/1910.10683, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "[16] Derek Tam, Rakesh R Menon, Mohit Bansal, Shashank Srivastava, and Colin Raffel. Improving",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 127,
+ 651,
+ 466,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 651,
+ 466,
+ 666
+ ],
+ "score": 1.0,
+ "content": "and simplifying pattern exploiting training. arXiv preprint arXiv:2103.11955, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 104,
+ 669,
+ 508,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 669,
+ 508,
+ 684
+ ],
+ "score": 1.0,
+ "content": "[17] Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 127,
+ 681,
+ 492,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 681,
+ 492,
+ 695
+ ],
+ "score": 1.0,
+ "content": "Neural text generation with unlikelihood training. arXiv preprint arXiv:1908.04319, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 714
+ ],
+ "score": 1.0,
+ "content": "[18] Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. PPT: Pre-trained prompt tuning for",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 126,
+ 711,
+ 367,
+ 725
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 711,
+ 367,
+ 725
+ ],
+ "score": 1.0,
+ "content": "few-shot learning. arXiv preprint arXiv:2109.04332, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "page_idx": 10,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 755
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 755
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 15,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 60,
+ 507,
+ 726
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "score": 1.0,
+ "content": "References",
+ "type": "text"
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 109,
+ 90,
+ 506,
+ 103
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 90,
+ 506,
+ 103
+ ],
+ "score": 1.0,
+ "content": "[1] Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai,",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 125,
+ 99,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 99,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": "Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 126,
+ 112,
+ 442,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 112,
+ 442,
+ 125
+ ],
+ "score": 1.0,
+ "content": "enables zero-shot task generalization. arXiv preprint arXiv:2110.08207, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 130,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 130,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "[2] Neel Alex, Eli Lifland, Lewis Tunstall, Abhishek Thakur, Pegah Maham, C Jess Riedel, Emmie",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 141,
+ 506,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 141,
+ 506,
+ 155
+ ],
+ "score": 1.0,
+ "content": "Hine, Carolyn Ashurst, Paul Sedille, Alexis Carlier, et al. RAFT: A real-world few-shot text",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 126,
+ 152,
+ 396,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 152,
+ 396,
+ 165
+ ],
+ "score": 1.0,
+ "content": "classification benchmark. arXiv preprint arXiv:2109.14076, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 6,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 170,
+ 507,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 170,
+ 507,
+ 184
+ ],
+ "score": 1.0,
+ "content": "[3] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever.",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 182,
+ 429,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 182,
+ 429,
+ 194
+ ],
+ "score": 1.0,
+ "content": "Language models are unsupervised multitask learners. OpenAI blog, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 109,
+ 199,
+ 507,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 199,
+ 507,
+ 214
+ ],
+ "score": 1.0,
+ "content": "[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 211,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 211,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 127,
+ 222,
+ 366,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 222,
+ 366,
+ 236
+ ],
+ "score": 1.0,
+ "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 109,
+ 239,
+ 506,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 239,
+ 506,
+ 256
+ ],
+ "score": 1.0,
+ "content": "[5] Yanda Chen, Ruiqi Zhong, Sheng Zha, George Karypis, and He He. Meta-learning via language",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 252,
+ 393,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 252,
+ 393,
+ 264
+ ],
+ "score": 1.0,
+ "content": "model in-context tuning. arXiv preprint arXiv:2110.07814, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 270,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 270,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "[6] Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. Metaicl: Learning to",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 281,
+ 360,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 281,
+ 360,
+ 295
+ ],
+ "score": 1.0,
+ "content": "learn in context. arXiv preprint arXiv:2110.15943, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 15,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 299,
+ 506,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 299,
+ 506,
+ 313
+ ],
+ "score": 1.0,
+ "content": "[7] Andrew Kyle Lampinen, Ishita Dasgupta, Stephanie C. Y. Chan, Kory Matthewson,",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 310,
+ 506,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 310,
+ 506,
+ 324
+ ],
+ "score": 1.0,
+ "content": "Michael Henry Tessler, Antonia Creswell, James L. McClelland, Jane X. Wang, and Felix",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 127,
+ 321,
+ 505,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 321,
+ 505,
+ 335
+ ],
+ "score": 1.0,
+ "content": "Hill. Can language models learn from explanations in context? ArXiv, abs/2204.02329, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 111,
+ 340,
+ 506,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 340,
+ 506,
+ 354
+ ],
+ "score": 1.0,
+ "content": "[8] Angeliki Lazaridou, Elena Gribovskaya, Wojciech Stokowiec, and Nikolai Grigorev. Internet-",
+ "type": "text"
+ }
+ ],
+ "index": 19,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 124,
+ 349,
+ 508,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 349,
+ 508,
+ 366
+ ],
+ "score": 1.0,
+ "content": "augmented language models through few-shot prompting for open-domain question answering.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 127,
+ 362,
+ 291,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 362,
+ 291,
+ 376
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2203.05115, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 21,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 380,
+ 506,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 380,
+ 506,
+ 394
+ ],
+ "score": 1.0,
+ "content": "[9] Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and",
+ "type": "text"
+ }
+ ],
+ "index": 22,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 125,
+ 389,
+ 506,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 389,
+ 506,
+ 407
+ ],
+ "score": 1.0,
+ "content": "Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 127,
+ 402,
+ 318,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 402,
+ 318,
+ 416
+ ],
+ "score": 1.0,
+ "content": "work? arXiv preprint arXiv:2202.12837, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 24,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 434
+ ],
+ "score": 1.0,
+ "content": "[10] Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei,",
+ "type": "text"
+ }
+ ],
+ "index": 25,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 431,
+ 507,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 431,
+ 507,
+ 445
+ ],
+ "score": 1.0,
+ "content": "Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 127,
+ 442,
+ 506,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 442,
+ 378,
+ 456
+ ],
+ "score": 1.0,
+ "content": "Benchmarking generalization via in-context instructions on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 378,
+ 443,
+ 408,
+ 453
+ ],
+ "score": 0.82,
+ "content": "1 { , } 6 0 0 { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 409,
+ 442,
+ 506,
+ 456
+ ],
+ "score": 1.0,
+ "content": "language tasks. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 124,
+ 453,
+ 267,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 453,
+ 267,
+ 468
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2204.07705, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 28,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 471,
+ 506,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 471,
+ 506,
+ 486
+ ],
+ "score": 1.0,
+ "content": "[11] Albert Webson and Ellie Pavlick. Do prompt-based models really understand the meaning of",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 483,
+ 353,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 483,
+ 353,
+ 496
+ ],
+ "score": 1.0,
+ "content": "their prompts? arXiv preprint arXiv:2109.01247, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 30,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 506,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 506,
+ 515
+ ],
+ "score": 1.0,
+ "content": "[12] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use:",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 513,
+ 507,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 513,
+ 507,
+ 527
+ ],
+ "score": 1.0,
+ "content": "Improving few-shot performance of language models. arXiv preprint arXiv:2102.09690, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 529,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 529,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 542,
+ 506,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 542,
+ 506,
+ 554
+ ],
+ "score": 1.0,
+ "content": "Weizhu Chen. LoRA: Low-rank adaptation of large language models. ArXiv, abs/2106.09685,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 126,
+ 552,
+ 155,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 552,
+ 155,
+ 567
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 35,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "[14] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient",
+ "type": "text"
+ }
+ ],
+ "index": 36,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 582,
+ 353,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 582,
+ 353,
+ 595
+ ],
+ "score": 1.0,
+ "content": "prompt tuning. arXiv preprint arXiv:2104.08691, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 37,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 614
+ ],
+ "score": 1.0,
+ "content": "[15] Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena,",
+ "type": "text"
+ }
+ ],
+ "index": 38,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 612,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 612,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 126,
+ 622,
+ 349,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 622,
+ 349,
+ 635
+ ],
+ "score": 1.0,
+ "content": "text-to-text transformer. ArXiv, abs/1910.10683, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 40,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "[16] Derek Tam, Rakesh R Menon, Mohit Bansal, Shashank Srivastava, and Colin Raffel. Improving",
+ "type": "text"
+ }
+ ],
+ "index": 41,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 651,
+ 466,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 651,
+ 466,
+ 666
+ ],
+ "score": 1.0,
+ "content": "and simplifying pattern exploiting training. arXiv preprint arXiv:2103.11955, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 42,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 669,
+ 508,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 669,
+ 508,
+ 684
+ ],
+ "score": 1.0,
+ "content": "[17] Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston.",
+ "type": "text"
+ }
+ ],
+ "index": 43,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 681,
+ 492,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 681,
+ 492,
+ 695
+ ],
+ "score": 1.0,
+ "content": "Neural text generation with unlikelihood training. arXiv preprint arXiv:1908.04319, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 714
+ ],
+ "score": 1.0,
+ "content": "[18] Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. PPT: Pre-trained prompt tuning for",
+ "type": "text"
+ }
+ ],
+ "index": 45,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 711,
+ 367,
+ 725
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 711,
+ 367,
+ 725
+ ],
+ "score": 1.0,
+ "content": "few-shot learning. arXiv preprint arXiv:2109.04332, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 46,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 104,
+ 70,
+ 508,
+ 725
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 72,
+ 505,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "[19] Tu Vu, Brian Lester, Noah Constant, Rami Al-Rfou, and Daniel Cer. SPoT: Better frozen model",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 127,
+ 83,
+ 452,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 83,
+ 452,
+ 96
+ ],
+ "score": 1.0,
+ "content": "adaptation through soft prompt transfer. arXiv preprint arXiv:2110.07904, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 102,
+ 504,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 115
+ ],
+ "score": 1.0,
+ "content": "[20] Jared Kaplan, Sam McCandlish, T. J. Henighan, Tom B. Brown, Benjamin Chess, Rewon Child,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 127,
+ 114,
+ 504,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 114,
+ 504,
+ 127
+ ],
+ "score": 1.0,
+ "content": "Scott Gray, Alec Radford, Jeff Wu, and Dario Amodei. Scaling laws for neural language models.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 127,
+ 125,
+ 290,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 125,
+ 290,
+ 137
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2001.08361, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 144,
+ 504,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 142,
+ 506,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 142,
+ 506,
+ 159
+ ],
+ "score": 1.0,
+ "content": "[21] Sewon Min, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Noisy channel language",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 127,
+ 155,
+ 488,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 155,
+ 488,
+ 168
+ ],
+ "score": 1.0,
+ "content": "model prompting for few-shot text classification. arXiv preprint arXiv:2108.04106, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 174,
+ 504,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 175,
+ 505,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 175,
+ 505,
+ 188
+ ],
+ "score": 1.0,
+ "content": "[22] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 127,
+ 185,
+ 470,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 185,
+ 470,
+ 198
+ ],
+ "score": 1.0,
+ "content": "with residual adapters. Advances in neural information processing systems, 30, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 205,
+ 505,
+ 239
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 218
+ ],
+ "score": 1.0,
+ "content": "[23] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 215,
+ 505,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 215,
+ 505,
+ 230
+ ],
+ "score": 1.0,
+ "content": "Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 128,
+ 227,
+ 329,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 227,
+ 329,
+ 240
+ ],
+ "score": 1.0,
+ "content": "for NLP. arXiv preprint arXiv:1902.00751, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 246,
+ 504,
+ 270
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "[24] Ankur Bapna, Naveen Arivazhagan, and Orhan Firat. Simple, scalable adaptation for neural",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 127,
+ 257,
+ 375,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 257,
+ 375,
+ 270
+ ],
+ "score": 1.0,
+ "content": "machine translation. arXiv preprint arXiv:1909.08478, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 277,
+ 504,
+ 300
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "[25] Demi Guo, Alexander M. Rush, and Yoon Kim. Parameter-efficient transfer learning with diff",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 127,
+ 289,
+ 327,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 289,
+ 327,
+ 300
+ ],
+ "score": 1.0,
+ "content": "pruning. arXiv preprint arXiv:2012.07463, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 307,
+ 504,
+ 330
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 321
+ ],
+ "score": 1.0,
+ "content": "[26] Yi-Lin Sung, Varun Nair, and Colin Raffel. Training neural networks with fixed sparse masks.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 127,
+ 319,
+ 290,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 319,
+ 290,
+ 330
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2111.09839, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 337,
+ 504,
+ 362
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 337,
+ 505,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 337,
+ 505,
+ 351
+ ],
+ "score": 1.0,
+ "content": "[27] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 127,
+ 348,
+ 470,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 348,
+ 470,
+ 362
+ ],
+ "score": 1.0,
+ "content": "effectiveness of language model fine-tuning. arXiv preprint arXiv:2012.13255, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 368,
+ 504,
+ 392
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 368,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 368,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "[28] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 127,
+ 379,
+ 451,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 379,
+ 451,
+ 391
+ ],
+ "score": 1.0,
+ "content": "low-rank hypercomplex adapter layers. arXiv preprint arXiv:2106.04647, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "[29] Xiang Lisa Li and Percy Liang. Prefix-Tuning: Optimizing continuous prompts for generation.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 127,
+ 410,
+ 290,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 410,
+ 290,
+ 421
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2101.00190, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 429,
+ 505,
+ 463
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "[30] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. To-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 127,
+ 441,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 441,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "wards a unified view of parameter-efficient transfer learning. arXiv preprint arXiv:2110.04366,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 126,
+ 450,
+ 154,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 450,
+ 154,
+ 464
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 470,
+ 504,
+ 494
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "[31] Ethan Perez, Douwe Kiela, and Kyunghyun Cho. True few-shot learning with language models.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 128,
+ 482,
+ 291,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 482,
+ 291,
+ 494
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2105.11447, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 501,
+ 504,
+ 535
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "[32] Avital Oliver, Augustus Odena, Colin Raffel, Ekin Dogus Cubuk, and Ian Goodfellow. Realistic",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 127,
+ 513,
+ 505,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 513,
+ 505,
+ 525
+ ],
+ "score": 1.0,
+ "content": "evaluation of deep semi-supervised learning algorithms. Advances in Neural Information",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 127,
+ 523,
+ 236,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 523,
+ 236,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Processing Systems, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 576
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 507,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 507,
+ 556
+ ],
+ "score": 1.0,
+ "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 126,
+ 553,
+ 505,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 553,
+ 505,
+ 566
+ ],
+ "score": 1.0,
+ "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 127,
+ 564,
+ 236,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 564,
+ 236,
+ 577
+ ],
+ "score": 1.0,
+ "content": "Processing Systems, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 617
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 584,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 584,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "[34] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 127,
+ 595,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 595,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 126,
+ 604,
+ 154,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 604,
+ 154,
+ 618
+ ],
+ "score": 1.0,
+ "content": "2018.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 678,
+ 506,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "[36] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 127,
+ 689,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 689,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. Transformers: State-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 126,
+ 711,
+ 425,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 711,
+ 425,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Methods in Natural Language Processing: System Demonstrations, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 39.5
+ }
+ ],
+ "page_idx": 11,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "12",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 72,
+ 505,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "[19] Tu Vu, Brian Lester, Noah Constant, Rami Al-Rfou, and Daniel Cer. SPoT: Better frozen model",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 127,
+ 83,
+ 452,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 83,
+ 452,
+ 96
+ ],
+ "score": 1.0,
+ "content": "adaptation through soft prompt transfer. arXiv preprint arXiv:2110.07904, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 106,
+ 72,
+ 505,
+ 96
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 102,
+ 504,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 115
+ ],
+ "score": 1.0,
+ "content": "[20] Jared Kaplan, Sam McCandlish, T. J. Henighan, Tom B. Brown, Benjamin Chess, Rewon Child,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 127,
+ 114,
+ 504,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 114,
+ 504,
+ 127
+ ],
+ "score": 1.0,
+ "content": "Scott Gray, Alec Radford, Jeff Wu, and Dario Amodei. Scaling laws for neural language models.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 127,
+ 125,
+ 290,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 125,
+ 290,
+ 137
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2001.08361, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 105,
+ 102,
+ 505,
+ 137
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 144,
+ 504,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 142,
+ 506,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 142,
+ 506,
+ 159
+ ],
+ "score": 1.0,
+ "content": "[21] Sewon Min, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Noisy channel language",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 127,
+ 155,
+ 488,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 155,
+ 488,
+ 168
+ ],
+ "score": 1.0,
+ "content": "model prompting for few-shot text classification. arXiv preprint arXiv:2108.04106, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 142,
+ 506,
+ 168
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 174,
+ 504,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 175,
+ 505,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 175,
+ 505,
+ 188
+ ],
+ "score": 1.0,
+ "content": "[22] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 127,
+ 185,
+ 470,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 185,
+ 470,
+ 198
+ ],
+ "score": 1.0,
+ "content": "with residual adapters. Advances in neural information processing systems, 30, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 106,
+ 175,
+ 505,
+ 198
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 205,
+ 505,
+ 239
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 218
+ ],
+ "score": 1.0,
+ "content": "[23] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 215,
+ 505,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 215,
+ 505,
+ 230
+ ],
+ "score": 1.0,
+ "content": "Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 128,
+ 227,
+ 329,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 227,
+ 329,
+ 240
+ ],
+ "score": 1.0,
+ "content": "for NLP. arXiv preprint arXiv:1902.00751, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 105,
+ 205,
+ 506,
+ 240
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 246,
+ 504,
+ 270
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "[24] Ankur Bapna, Naveen Arivazhagan, and Orhan Firat. Simple, scalable adaptation for neural",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 127,
+ 257,
+ 375,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 257,
+ 375,
+ 270
+ ],
+ "score": 1.0,
+ "content": "machine translation. arXiv preprint arXiv:1909.08478, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 106,
+ 246,
+ 505,
+ 270
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 277,
+ 504,
+ 300
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "[25] Demi Guo, Alexander M. Rush, and Yoon Kim. Parameter-efficient transfer learning with diff",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 127,
+ 289,
+ 327,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 289,
+ 327,
+ 300
+ ],
+ "score": 1.0,
+ "content": "pruning. arXiv preprint arXiv:2012.07463, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 106,
+ 277,
+ 505,
+ 300
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 307,
+ 504,
+ 330
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 321
+ ],
+ "score": 1.0,
+ "content": "[26] Yi-Lin Sung, Varun Nair, and Colin Raffel. Training neural networks with fixed sparse masks.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 127,
+ 319,
+ 290,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 319,
+ 290,
+ 330
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2111.09839, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 106,
+ 308,
+ 506,
+ 330
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 337,
+ 504,
+ 362
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 337,
+ 505,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 337,
+ 505,
+ 351
+ ],
+ "score": 1.0,
+ "content": "[27] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 127,
+ 348,
+ 470,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 348,
+ 470,
+ 362
+ ],
+ "score": 1.0,
+ "content": "effectiveness of language model fine-tuning. arXiv preprint arXiv:2012.13255, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 106,
+ 337,
+ 505,
+ 362
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 368,
+ 504,
+ 392
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 368,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 368,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "[28] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 127,
+ 379,
+ 451,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 379,
+ 451,
+ 391
+ ],
+ "score": 1.0,
+ "content": "low-rank hypercomplex adapter layers. arXiv preprint arXiv:2106.04647, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 106,
+ 368,
+ 505,
+ 391
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "[29] Xiang Lisa Li and Percy Liang. Prefix-Tuning: Optimizing continuous prompts for generation.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 127,
+ 410,
+ 290,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 410,
+ 290,
+ 421
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2101.00190, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 106,
+ 398,
+ 506,
+ 421
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 429,
+ 505,
+ 463
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "[30] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. To-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 127,
+ 441,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 441,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "wards a unified view of parameter-efficient transfer learning. arXiv preprint arXiv:2110.04366,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 126,
+ 450,
+ 154,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 450,
+ 154,
+ 464
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 106,
+ 429,
+ 506,
+ 464
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 470,
+ 504,
+ 494
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "[31] Ethan Perez, Douwe Kiela, and Kyunghyun Cho. True few-shot learning with language models.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 128,
+ 482,
+ 291,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 482,
+ 291,
+ 494
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2105.11447, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 106,
+ 470,
+ 506,
+ 494
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 501,
+ 504,
+ 535
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "[32] Avital Oliver, Augustus Odena, Colin Raffel, Ekin Dogus Cubuk, and Ian Goodfellow. Realistic",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 127,
+ 513,
+ 505,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 513,
+ 505,
+ 525
+ ],
+ "score": 1.0,
+ "content": "evaluation of deep semi-supervised learning algorithms. Advances in Neural Information",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 127,
+ 523,
+ 236,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 523,
+ 236,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Processing Systems, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 501,
+ 506,
+ 536
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 576
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 507,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 507,
+ 556
+ ],
+ "score": 1.0,
+ "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 126,
+ 553,
+ 505,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 553,
+ 505,
+ 566
+ ],
+ "score": 1.0,
+ "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 127,
+ 564,
+ 236,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 564,
+ 236,
+ 577
+ ],
+ "score": 1.0,
+ "content": "Processing Systems, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 105,
+ 542,
+ 507,
+ 577
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 617
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 584,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 584,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "[34] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 127,
+ 595,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 595,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 126,
+ 604,
+ 154,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 604,
+ 154,
+ 618
+ ],
+ "score": 1.0,
+ "content": "2018.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 106,
+ 584,
+ 506,
+ 618
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 678,
+ 506,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "[36] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 127,
+ 689,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 689,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. Transformers: State-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 126,
+ 711,
+ 425,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 711,
+ 425,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Methods in Natural Language Processing: System Demonstrations, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 39.5,
+ "bbox_fs": [
+ 105,
+ 677,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 24,
+ 507,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 87
+ ],
+ "score": 1.0,
+ "content": "[37] Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S. Gordon. Choice of plausible",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 125,
+ 81,
+ 507,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 81,
+ 507,
+ 99
+ ],
+ "score": 1.0,
+ "content": "alternatives: An evaluation of commonsense causal reasoning. 2011 AAAI Spring Symposium",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 126,
+ 94,
+ 183,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 94,
+ 183,
+ 108
+ ],
+ "score": 1.0,
+ "content": "Series, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 115,
+ 506,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 115,
+ 506,
+ 127
+ ],
+ "score": 1.0,
+ "content": "[38] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag: Can a",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 126,
+ 125,
+ 439,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 125,
+ 439,
+ 139
+ ],
+ "score": 1.0,
+ "content": "machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 107,
+ 146,
+ 506,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 146,
+ 506,
+ 159
+ ],
+ "score": 1.0,
+ "content": "[39] Rishi Sharma, James Allen, Omid Bakhshandeh, and Nasrin Mostafazadeh. Tackling the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 127,
+ 156,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 156,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "story ending biases in the story cloze test. In Proceedings of the 56th Annual Meeting of the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 125,
+ 167,
+ 496,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 167,
+ 496,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Association for Computational Linguistics (Volume 2: Short Papers), pages 752–757, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 506,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 506,
+ 202
+ ],
+ "score": 1.0,
+ "content": "[40] Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 127,
+ 198,
+ 506,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 198,
+ 506,
+ 212
+ ],
+ "score": 1.0,
+ "content": "Adversarial NLI: A new benchmark for natural language understanding. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 209,
+ 232,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 209,
+ 232,
+ 223
+ ],
+ "score": 1.0,
+ "content": "arXiv:1910.14599, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 229,
+ 507,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 229,
+ 507,
+ 243
+ ],
+ "score": 1.0,
+ "content": "[41] Marie-Catherine de Marneffe, Mandy Simons, and Judith Tonhauser. The CommitmentBank:",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 126,
+ 240,
+ 506,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 240,
+ 506,
+ 257
+ ],
+ "score": 1.0,
+ "content": "Investigating projection in naturally occurring discourse. Proceedings of Sinn und Bedeutung",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 127,
+ 252,
+ 168,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 252,
+ 168,
+ 264
+ ],
+ "score": 1.0,
+ "content": "23, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 286
+ ],
+ "score": 1.0,
+ "content": "[42] Ido Dagan, Oren Glickman, and Bernardo Magnini. The pascal recognising textual entailment",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 127,
+ 282,
+ 480,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 282,
+ 480,
+ 297
+ ],
+ "score": 1.0,
+ "content": "challenge. In Machine Learning Challenges Workshop, pages 177–190. Springer, 2005.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 303,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "[43] Hector Levesque, Ernest Davis, and Leora Morgenstern. The winograd schema challenge. Thir-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 126,
+ 312,
+ 506,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 312,
+ 506,
+ 329
+ ],
+ "score": 1.0,
+ "content": "teenth International Conference on the Principles of Knowledge Representation and Reasoning,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 127,
+ 324,
+ 154,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 324,
+ 154,
+ 338
+ ],
+ "score": 1.0,
+ "content": "2012.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 344,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 344,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "[44] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 127,
+ 356,
+ 506,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 356,
+ 506,
+ 370
+ ],
+ "score": 1.0,
+ "content": "adversarial winograd schema challenge at scale. In Proceedings of the AAAI Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 125,
+ 366,
+ 244,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 366,
+ 244,
+ 380
+ ],
+ "score": 1.0,
+ "content": "Artificial Intelligence, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 388,
+ 505,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 388,
+ 505,
+ 400
+ ],
+ "score": 1.0,
+ "content": "[45] Mohammad Taher Pilehvar and Jose Camacho-Collados. WiC: the word-in-context dataset for",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 127,
+ 398,
+ 506,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 398,
+ 506,
+ 412
+ ],
+ "score": 1.0,
+ "content": "evaluating context-sensitive meaning representations. arXiv preprint arXiv:1808.09121, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 417,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "[46] Stefan Van Der Walt, S. Chris Colbert, and Gael Varoquaux. The numpy array: a structure for",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 127,
+ 430,
+ 464,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 430,
+ 464,
+ 442
+ ],
+ "score": 1.0,
+ "content": "efficient numerical computation. Computing in science & engineering, 13(2), 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 449,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "[47] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. BitFit: Simple parameter-efficient",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 126,
+ 459,
+ 507,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 459,
+ 507,
+ 476
+ ],
+ "score": 1.0,
+ "content": "fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 126,
+ 470,
+ 155,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 470,
+ 155,
+ 485
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 492,
+ 506,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 492,
+ 506,
+ 504
+ ],
+ "score": 1.0,
+ "content": "[48] Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 127,
+ 501,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 501,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": "Du, Andrew M Dai, and Quoc V. Le. Finetuned language models are zero-shot learners. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 125,
+ 513,
+ 268,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 513,
+ 268,
+ 526
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2109.01652, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 548
+ ],
+ "score": 1.0,
+ "content": "[49] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 126,
+ 544,
+ 412,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 544,
+ 412,
+ 558
+ ],
+ "score": 1.0,
+ "content": "cost. In International Conference on Machine Learning. PMLR, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 506,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 506,
+ 578
+ ],
+ "score": 1.0,
+ "content": "[50] Timo Schick and Hinrich Schütze. True few-shot learning with prompts–a real-world perspective.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 127,
+ 576,
+ 291,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 576,
+ 291,
+ 590
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2111.13440, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 596,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 596,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "[51] Moshe Wasserblat. Sentence transformer fine-tuning (SetFit): Outperforming GPT-3 on few-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 127,
+ 607,
+ 374,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 607,
+ 374,
+ 619
+ ],
+ "score": 1.0,
+ "content": "shot text-classification while being 1600 times smaller, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 506,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 506,
+ 641
+ ],
+ "score": 1.0,
+ "content": "[52] Mostafa Dehghani, Anurag Arnab, Lucas Beyer, Ashish Vaswani, and Yi Tay. The efficiency",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 127,
+ 638,
+ 335,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 638,
+ 335,
+ 650
+ ],
+ "score": 1.0,
+ "content": "misnomer. arXiv preprint arXiv:2110.12894, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 657,
+ 506,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 506,
+ 672
+ ],
+ "score": 1.0,
+ "content": "[53] Guanghui Qin and Jason Eisner. Learning how to ask: Querying LMs with mixtures of soft",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 126,
+ 669,
+ 330,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 669,
+ 330,
+ 682
+ ],
+ "score": 1.0,
+ "content": "prompts. arXiv preprint arXiv:2104.06599, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 507,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 507,
+ 703
+ ],
+ "score": 1.0,
+ "content": "[54] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-Tuning v2:",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 714
+ ],
+ "score": 1.0,
+ "content": "Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 125,
+ 711,
+ 266,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 711,
+ 266,
+ 724
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2110.07602, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "page_idx": 12,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "13",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 24,
+ 507,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 87
+ ],
+ "score": 1.0,
+ "content": "[37] Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S. Gordon. Choice of plausible",
+ "type": "text"
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 125,
+ 81,
+ 507,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 81,
+ 507,
+ 99
+ ],
+ "score": 1.0,
+ "content": "alternatives: An evaluation of commonsense causal reasoning. 2011 AAAI Spring Symposium",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 126,
+ 94,
+ 183,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 94,
+ 183,
+ 108
+ ],
+ "score": 1.0,
+ "content": "Series, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 107,
+ 115,
+ 506,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 115,
+ 506,
+ 127
+ ],
+ "score": 1.0,
+ "content": "[38] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag: Can a",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 125,
+ 439,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 125,
+ 439,
+ 139
+ ],
+ "score": 1.0,
+ "content": "machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 107,
+ 146,
+ 506,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 146,
+ 506,
+ 159
+ ],
+ "score": 1.0,
+ "content": "[39] Rishi Sharma, James Allen, Omid Bakhshandeh, and Nasrin Mostafazadeh. Tackling the",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 156,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 156,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "story ending biases in the story cloze test. In Proceedings of the 56th Annual Meeting of the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 125,
+ 167,
+ 496,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 167,
+ 496,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Association for Computational Linguistics (Volume 2: Short Papers), pages 752–757, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 506,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 506,
+ 202
+ ],
+ "score": 1.0,
+ "content": "[40] Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela.",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 198,
+ 506,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 198,
+ 506,
+ 212
+ ],
+ "score": 1.0,
+ "content": "Adversarial NLI: A new benchmark for natural language understanding. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 209,
+ 232,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 209,
+ 232,
+ 223
+ ],
+ "score": 1.0,
+ "content": "arXiv:1910.14599, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 229,
+ 507,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 229,
+ 507,
+ 243
+ ],
+ "score": 1.0,
+ "content": "[41] Marie-Catherine de Marneffe, Mandy Simons, and Judith Tonhauser. The CommitmentBank:",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 240,
+ 506,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 240,
+ 506,
+ 257
+ ],
+ "score": 1.0,
+ "content": "Investigating projection in naturally occurring discourse. Proceedings of Sinn und Bedeutung",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 127,
+ 252,
+ 168,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 252,
+ 168,
+ 264
+ ],
+ "score": 1.0,
+ "content": "23, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 286
+ ],
+ "score": 1.0,
+ "content": "[42] Ido Dagan, Oren Glickman, and Bernardo Magnini. The pascal recognising textual entailment",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 282,
+ 480,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 282,
+ 480,
+ 297
+ ],
+ "score": 1.0,
+ "content": "challenge. In Machine Learning Challenges Workshop, pages 177–190. Springer, 2005.",
+ "type": "text"
+ }
+ ],
+ "index": 15,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 303,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "[43] Hector Levesque, Ernest Davis, and Leora Morgenstern. The winograd schema challenge. Thir-",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 312,
+ 506,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 312,
+ 506,
+ 329
+ ],
+ "score": 1.0,
+ "content": "teenth International Conference on the Principles of Knowledge Representation and Reasoning,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 127,
+ 324,
+ 154,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 324,
+ 154,
+ 338
+ ],
+ "score": 1.0,
+ "content": "2012.",
+ "type": "text"
+ }
+ ],
+ "index": 18,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 344,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 344,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "[44] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An",
+ "type": "text"
+ }
+ ],
+ "index": 19,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 356,
+ 506,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 356,
+ 506,
+ 370
+ ],
+ "score": 1.0,
+ "content": "adversarial winograd schema challenge at scale. In Proceedings of the AAAI Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 125,
+ 366,
+ 244,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 366,
+ 244,
+ 380
+ ],
+ "score": 1.0,
+ "content": "Artificial Intelligence, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 21,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 388,
+ 505,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 388,
+ 505,
+ 400
+ ],
+ "score": 1.0,
+ "content": "[45] Mohammad Taher Pilehvar and Jose Camacho-Collados. WiC: the word-in-context dataset for",
+ "type": "text"
+ }
+ ],
+ "index": 22,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 398,
+ 506,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 398,
+ 506,
+ 412
+ ],
+ "score": 1.0,
+ "content": "evaluating context-sensitive meaning representations. arXiv preprint arXiv:1808.09121, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 417,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "[46] Stefan Van Der Walt, S. Chris Colbert, and Gael Varoquaux. The numpy array: a structure for",
+ "type": "text"
+ }
+ ],
+ "index": 24,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 430,
+ 464,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 430,
+ 464,
+ 442
+ ],
+ "score": 1.0,
+ "content": "efficient numerical computation. Computing in science & engineering, 13(2), 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 25,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 449,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "[47] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. BitFit: Simple parameter-efficient",
+ "type": "text"
+ }
+ ],
+ "index": 26,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 459,
+ 507,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 459,
+ 507,
+ 476
+ ],
+ "score": 1.0,
+ "content": "fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 126,
+ 470,
+ 155,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 470,
+ 155,
+ 485
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 28,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 492,
+ 506,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 492,
+ 506,
+ 504
+ ],
+ "score": 1.0,
+ "content": "[48] Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 501,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 501,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": "Du, Andrew M Dai, and Quoc V. Le. Finetuned language models are zero-shot learners. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 125,
+ 513,
+ 268,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 513,
+ 268,
+ 526
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2109.01652, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 548
+ ],
+ "score": 1.0,
+ "content": "[49] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory",
+ "type": "text"
+ }
+ ],
+ "index": 32,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 544,
+ 412,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 544,
+ 412,
+ 558
+ ],
+ "score": 1.0,
+ "content": "cost. In International Conference on Machine Learning. PMLR, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 506,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 506,
+ 578
+ ],
+ "score": 1.0,
+ "content": "[50] Timo Schick and Hinrich Schütze. True few-shot learning with prompts–a real-world perspective.",
+ "type": "text"
+ }
+ ],
+ "index": 34,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 576,
+ 291,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 576,
+ 291,
+ 590
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2111.13440, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 35,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 596,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 596,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "[51] Moshe Wasserblat. Sentence transformer fine-tuning (SetFit): Outperforming GPT-3 on few-",
+ "type": "text"
+ }
+ ],
+ "index": 36,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 607,
+ 374,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 607,
+ 374,
+ 619
+ ],
+ "score": 1.0,
+ "content": "shot text-classification while being 1600 times smaller, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 37,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 506,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 506,
+ 641
+ ],
+ "score": 1.0,
+ "content": "[52] Mostafa Dehghani, Anurag Arnab, Lucas Beyer, Ashish Vaswani, and Yi Tay. The efficiency",
+ "type": "text"
+ }
+ ],
+ "index": 38,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 638,
+ 335,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 638,
+ 335,
+ 650
+ ],
+ "score": 1.0,
+ "content": "misnomer. arXiv preprint arXiv:2110.12894, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 39,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 657,
+ 506,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 506,
+ 672
+ ],
+ "score": 1.0,
+ "content": "[53] Guanghui Qin and Jason Eisner. Learning how to ask: Querying LMs with mixtures of soft",
+ "type": "text"
+ }
+ ],
+ "index": 40,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 669,
+ 330,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 669,
+ 330,
+ 682
+ ],
+ "score": 1.0,
+ "content": "prompts. arXiv preprint arXiv:2104.06599, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 41,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 507,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 507,
+ 703
+ ],
+ "score": 1.0,
+ "content": "[54] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-Tuning v2:",
+ "type": "text"
+ }
+ ],
+ "index": 42,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 699,
+ 506,
+ 714
+ ],
+ "score": 1.0,
+ "content": "Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 125,
+ 711,
+ 266,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 711,
+ 266,
+ 724
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2110.07602, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 44,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 71,
+ 507,
+ 88
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 71,
+ 507,
+ 88
+ ],
+ "score": 1.0,
+ "content": "[55] Shengnan An, Yifei Li, Zeqi Lin, Qian Liu, Bei Chen, Qiang Fu, Weizhu Chen, Nanning Zheng,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 83,
+ 507,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 83,
+ 507,
+ 97
+ ],
+ "score": 1.0,
+ "content": "and Jian-Guang Lou. Input-Tuning: Adapting unfamiliar inputs to frozen pretrained models.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 95,
+ 290,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 95,
+ 290,
+ 109
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2203.03131, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "[56] Yulong Chen, Yang Liu, Li Dong, Shuohang Wang, Chenguang Zhu, Michael Zeng, and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 124,
+ 507,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 124,
+ 507,
+ 141
+ ],
+ "score": 1.0,
+ "content": "Yue Zhang. AdaPrompt: Adaptive model training for prompt-based NLP. arXiv preprint",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 127,
+ 136,
+ 231,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 136,
+ 231,
+ 150
+ ],
+ "score": 1.0,
+ "content": "arXiv:2202.04824, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 154,
+ 506,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 154,
+ 506,
+ 173
+ ],
+ "score": 1.0,
+ "content": "[57] Shizhe Diao, Xuechun Li, Yong Lin, Zhichao Huang, and Tong Zhang. Black-box prompt",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 166,
+ 459,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 166,
+ 459,
+ 182
+ ],
+ "score": 1.0,
+ "content": "learning for pre-trained language models. arXiv preprint arXiv:2201.08531, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 188,
+ 506,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 188,
+ 506,
+ 202
+ ],
+ "score": 1.0,
+ "content": "[58] Daniel Khashabi, Shane Lyu, Sewon Min, Lianhui Qin, Kyle Richardson, Sameer Singh,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 199,
+ 507,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 199,
+ 507,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Sean Welleck, Hannaneh Hajishirzi, Tushar Khot, Ashish Sabharwal, et al. Prompt wayward-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 208,
+ 506,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 208,
+ 506,
+ 224
+ ],
+ "score": 1.0,
+ "content": "ness: The curious case of discretized interpretation of continuous prompts. arXiv preprint",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 126,
+ 220,
+ 232,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 220,
+ 232,
+ 234
+ ],
+ "score": 1.0,
+ "content": "arXiv:2112.08348, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "[59] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 250,
+ 506,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 250,
+ 506,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. arXiv",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 124,
+ 262,
+ 267,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 262,
+ 267,
+ 277
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2112.08654, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 506,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 506,
+ 297
+ ],
+ "score": 1.0,
+ "content": "[60] Zonghan Yang and Yang Liu. On robust prefix-tuning for text classification. arXiv preprint",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 15,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 294,
+ 231,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 294,
+ 231,
+ 308
+ ],
+ "score": 1.0,
+ "content": "arXiv:2203.10378, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 16,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 312,
+ 507,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 312,
+ 507,
+ 330
+ ],
+ "score": 1.0,
+ "content": "[61] Yuting Yang, Pei Huang, Juan Cao, Jintao Li, Yun Lin, Jin Song Dong, Feifei Ma, and Jian Zhang.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 17,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 325,
+ 507,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 325,
+ 507,
+ 339
+ ],
+ "score": 1.0,
+ "content": "A prompting-based approach for adversarial example generation and robustness enhancement.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 127,
+ 336,
+ 290,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 336,
+ 290,
+ 350
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2203.10714, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 19,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 356,
+ 506,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 506,
+ 370
+ ],
+ "score": 1.0,
+ "content": "[62] Xiaochen Liu, Yu Bai, Jiawei Li, Yinan Hu, and Yang Gao. PSP: Pre-trained soft prompts for",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 20,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 367,
+ 437,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 367,
+ 437,
+ 381
+ ],
+ "score": 1.0,
+ "content": "few-shot abstractive summarization. arXiv preprint arXiv:2204.04413, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 21,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 385,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 385,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "[63] Xavier Garcia and Orhan Firat. Using natural language prompts for machine translation. arXiv",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 22,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 397,
+ 265,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 397,
+ 265,
+ 412
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2202.11822, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 23,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 417,
+ 507,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 417,
+ 507,
+ 433
+ ],
+ "score": 1.0,
+ "content": "[64] Hunter Lang, Monica Agrawal, Yoon Kim, and David Sontag. Co-training improves prompt-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 24,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 429,
+ 460,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 429,
+ 460,
+ 443
+ ],
+ "score": 1.0,
+ "content": "based learning for large language models. arXiv preprint arXiv:2202.00828, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 25,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 450,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "[65] Boshi Wang, Xiang Deng, and Huan Sun. Shepherd pre-trained language models to develop a",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 26,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 460,
+ 495,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 460,
+ 495,
+ 475
+ ],
+ "score": 1.0,
+ "content": "train of thought: An iterative prompting approach. arXiv preprint arXiv:2203.08383, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 495
+ ],
+ "score": 1.0,
+ "content": "[66] Xu Zou, Da Yin, Qingyang Zhong, Hongxia Yang, Zhilin Yang, and Jie Tang. Controllable gener-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 28,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 491,
+ 507,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 491,
+ 507,
+ 507
+ ],
+ "score": 1.0,
+ "content": "ation from pre-trained language models via inverse prompting. arXiv preprint arXiv:2103.10685,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 127,
+ 503,
+ 154,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 503,
+ 154,
+ 515
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 30,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 521,
+ 507,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 507,
+ 537
+ ],
+ "score": 1.0,
+ "content": "[67] Yusheng Su, Xiaozhi Wang, Yujia Qin, Chi-Min Chan, Yankai Lin, Zhiyuan Liu, Peng Li,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 31,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 125,
+ 532,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 532,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "Juanzi Li, Lei Hou, Maosong Sun, et al. On transferability of prompt tuning for natural language",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 127,
+ 544,
+ 353,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 544,
+ 353,
+ 558
+ ],
+ "score": 1.0,
+ "content": "understanding. arXiv preprint arXiv:2111.06719, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 33,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 563,
+ 506,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 506,
+ 579
+ ],
+ "score": 1.0,
+ "content": "[68] Yun He, Huaixiu Steven Zheng, Yi Tay, Jai Gupta, Yu Du, Vamsi Aribandi, Zhe Zhao, YaGuang",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 34,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 574,
+ 506,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 574,
+ 506,
+ 591
+ ],
+ "score": 1.0,
+ "content": "Li, Zhao Chen, Donald Metzler, et al. HyperPrompt: Prompt-based task-conditioning of",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 125,
+ 585,
+ 347,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 585,
+ 347,
+ 600
+ ],
+ "score": 1.0,
+ "content": "transformers. arXiv preprint arXiv:2203.00759, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 36,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "[69] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 37,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 617,
+ 460,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 617,
+ 460,
+ 633
+ ],
+ "score": 1.0,
+ "content": "and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 38,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 637,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "[70] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 39,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 648,
+ 418,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 648,
+ 418,
+ 662
+ ],
+ "score": 1.0,
+ "content": "and natural language inference. arXiv preprint arXiv:2001.07676, 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 40,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 668,
+ 505,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 668,
+ 505,
+ 682
+ ],
+ "score": 1.0,
+ "content": "[71] Teven Le Scao and Alexander M. Rush. How many data points is a prompt worth? arXiv",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 41,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 124,
+ 679,
+ 267,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 679,
+ 267,
+ 695
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2103.08493, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 42,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 699,
+ 506,
+ 716
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 699,
+ 506,
+ 716
+ ],
+ "score": 1.0,
+ "content": "[72] Sen Yang, Yunchen Zhang, Leyang Cui, and Yue Zhang. Do prompts solve NLP tasks using",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 43,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 125,
+ 709,
+ 366,
+ 725
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 709,
+ 366,
+ 725
+ ],
+ "score": 1.0,
+ "content": "natural language? arXiv preprint arXiv:2203.00902, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 44,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 72,
+ 506,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 506,
+ 87
+ ],
+ "score": 1.0,
+ "content": "[73] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Auto-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 84,
+ 507,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 84,
+ 507,
+ 98
+ ],
+ "score": 1.0,
+ "content": "Prompt: Eliciting knowledge from language models with automatically generated prompts.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 95,
+ 290,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 95,
+ 290,
+ 108
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2010.15980, 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "[74] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pre-trained language models better few-shot",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 126,
+ 328,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 126,
+ 328,
+ 138
+ ],
+ "score": 1.0,
+ "content": "learners. arXiv preprint arXiv:2012.15723, 2020.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 143,
+ 507,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 143,
+ 507,
+ 159
+ ],
+ "score": 1.0,
+ "content": "[75] Ningyu Zhang, Luoqiu Li, Xiang Chen, Shumin Deng, Zhen Bi, Chuanqi Tan, Fei Huang,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 156,
+ 505,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 156,
+ 505,
+ 169
+ ],
+ "score": 1.0,
+ "content": "and Huajun Chen. Differentiable prompt makes pre-trained language models better few-shot",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 127,
+ 167,
+ 328,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 167,
+ 328,
+ 179
+ ],
+ "score": 1.0,
+ "content": "learners. arXiv preprint arXiv:2108.13161, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 185,
+ 506,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 185,
+ 506,
+ 199
+ ],
+ "score": 1.0,
+ "content": "[76] Rabeeh Karimi Mahabadi, Luke Zettlemoyer, James Henderson, Marzieh Saeidi, Lambert",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 196,
+ 506,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 196,
+ 506,
+ 209
+ ],
+ "score": 1.0,
+ "content": "Mathias, Veselin Stoyanov, and Majid Yazdani. PERFECT: Prompt-free and efficient few-shot",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 207,
+ 419,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 207,
+ 419,
+ 221
+ ],
+ "score": 1.0,
+ "content": "learning with language models. arXiv preprint arXiv:2204.01172, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 227,
+ 505,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 227,
+ 505,
+ 241
+ ],
+ "score": 1.0,
+ "content": "[77] Nafise Sadat Moosavi, Quentin Delfosse, Kristian Kersting, and Iryna Gurevych. Adaptable",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 239,
+ 330,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 239,
+ 330,
+ 250
+ ],
+ "score": 1.0,
+ "content": "adapters. arXiv preprint arXiv:2205.01549, 2022.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 256,
+ 507,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 507,
+ 271
+ ],
+ "score": 1.0,
+ "content": "[78] Eleni Triantafillou, Hugo Larochelle, Richard Zemel, and Vincent Dumoulin. Learning a",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 268,
+ 507,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 268,
+ 507,
+ 282
+ ],
+ "score": 1.0,
+ "content": "universal template for few-shot dataset generalization. arXiv preprint arXiv:/2105.07029, 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 302
+ ],
+ "score": 1.0,
+ "content": "[79] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E. Turner.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 15,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 298,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 298,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "Fast and flexible multi-task classification using conditional neural adaptive processes. arXiv",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 126,
+ 310,
+ 265,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 310,
+ 265,
+ 322
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1906.07697, 2019.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 17,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 329,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "[80] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 18,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 126,
+ 338,
+ 506,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 338,
+ 506,
+ 354
+ ],
+ "score": 1.0,
+ "content": "domains for few-shot classification. Proceedings of the IEEE/CVF International Conference on",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 127,
+ 351,
+ 227,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 351,
+ 227,
+ 363
+ ],
+ "score": 1.0,
+ "content": "Computer Vision., 2021.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 20,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 384
+ ],
+ "score": 1.0,
+ "content": "[81] Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 21,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 127,
+ 380,
+ 506,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 380,
+ 506,
+ 395
+ ],
+ "score": 1.0,
+ "content": "Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 127,
+ 391,
+ 506,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 391,
+ 506,
+ 406
+ ],
+ "score": 1.0,
+ "content": "Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 127,
+ 403,
+ 498,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 403,
+ 498,
+ 415
+ ],
+ "score": 1.0,
+ "content": "model evaluation, September 2021. URL https://doi.org/10.5281/zenodo.5371628.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 72,
+ 507,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 54,
+ 507,
+ 727
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 71,
+ 507,
+ 88
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 71,
+ 507,
+ 88
+ ],
+ "score": 1.0,
+ "content": "[55] Shengnan An, Yifei Li, Zeqi Lin, Qian Liu, Bei Chen, Qiang Fu, Weizhu Chen, Nanning Zheng,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 127,
+ 83,
+ 507,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 83,
+ 507,
+ 97
+ ],
+ "score": 1.0,
+ "content": "and Jian-Guang Lou. Input-Tuning: Adapting unfamiliar inputs to frozen pretrained models.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 95,
+ 290,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 95,
+ 290,
+ 109
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2203.03131, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "[56] Yulong Chen, Yang Liu, Li Dong, Shuohang Wang, Chenguang Zhu, Michael Zeng, and",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 126,
+ 124,
+ 507,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 124,
+ 507,
+ 141
+ ],
+ "score": 1.0,
+ "content": "Yue Zhang. AdaPrompt: Adaptive model training for prompt-based NLP. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 127,
+ 136,
+ 231,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 136,
+ 231,
+ 150
+ ],
+ "score": 1.0,
+ "content": "arXiv:2202.04824, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 154,
+ 506,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 154,
+ 506,
+ 173
+ ],
+ "score": 1.0,
+ "content": "[57] Shizhe Diao, Xuechun Li, Yong Lin, Zhichao Huang, and Tong Zhang. Black-box prompt",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 126,
+ 166,
+ 459,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 166,
+ 459,
+ 182
+ ],
+ "score": 1.0,
+ "content": "learning for pre-trained language models. arXiv preprint arXiv:2201.08531, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 188,
+ 506,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 188,
+ 506,
+ 202
+ ],
+ "score": 1.0,
+ "content": "[58] Daniel Khashabi, Shane Lyu, Sewon Min, Lianhui Qin, Kyle Richardson, Sameer Singh,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 127,
+ 199,
+ 507,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 199,
+ 507,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Sean Welleck, Hannaneh Hajishirzi, Tushar Khot, Ashish Sabharwal, et al. Prompt wayward-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 208,
+ 506,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 208,
+ 506,
+ 224
+ ],
+ "score": 1.0,
+ "content": "ness: The curious case of discretized interpretation of continuous prompts. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 126,
+ 220,
+ 232,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 220,
+ 232,
+ 234
+ ],
+ "score": 1.0,
+ "content": "arXiv:2112.08348, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "[59] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 126,
+ 250,
+ 506,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 250,
+ 506,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 124,
+ 262,
+ 267,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 262,
+ 267,
+ 277
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2112.08654, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 506,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 506,
+ 297
+ ],
+ "score": 1.0,
+ "content": "[60] Zonghan Yang and Yang Liu. On robust prefix-tuning for text classification. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 127,
+ 294,
+ 231,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 294,
+ 231,
+ 308
+ ],
+ "score": 1.0,
+ "content": "arXiv:2203.10378, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 312,
+ 507,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 312,
+ 507,
+ 330
+ ],
+ "score": 1.0,
+ "content": "[61] Yuting Yang, Pei Huang, Juan Cao, Jintao Li, Yun Lin, Jin Song Dong, Feifei Ma, and Jian Zhang.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 127,
+ 325,
+ 507,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 325,
+ 507,
+ 339
+ ],
+ "score": 1.0,
+ "content": "A prompting-based approach for adversarial example generation and robustness enhancement.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 127,
+ 336,
+ 290,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 336,
+ 290,
+ 350
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2203.10714, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 356,
+ 506,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 506,
+ 370
+ ],
+ "score": 1.0,
+ "content": "[62] Xiaochen Liu, Yu Bai, Jiawei Li, Yinan Hu, and Yang Gao. PSP: Pre-trained soft prompts for",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 127,
+ 367,
+ 437,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 367,
+ 437,
+ 381
+ ],
+ "score": 1.0,
+ "content": "few-shot abstractive summarization. arXiv preprint arXiv:2204.04413, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 385,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 385,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "[63] Xavier Garcia and Orhan Firat. Using natural language prompts for machine translation. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 126,
+ 397,
+ 265,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 397,
+ 265,
+ 412
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2202.11822, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 417,
+ 507,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 417,
+ 507,
+ 433
+ ],
+ "score": 1.0,
+ "content": "[64] Hunter Lang, Monica Agrawal, Yoon Kim, and David Sontag. Co-training improves prompt-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 127,
+ 429,
+ 460,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 429,
+ 460,
+ 443
+ ],
+ "score": 1.0,
+ "content": "based learning for large language models. arXiv preprint arXiv:2202.00828, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 450,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "[65] Boshi Wang, Xiang Deng, and Huan Sun. Shepherd pre-trained language models to develop a",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 127,
+ 460,
+ 495,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 460,
+ 495,
+ 475
+ ],
+ "score": 1.0,
+ "content": "train of thought: An iterative prompting approach. arXiv preprint arXiv:2203.08383, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 495
+ ],
+ "score": 1.0,
+ "content": "[66] Xu Zou, Da Yin, Qingyang Zhong, Hongxia Yang, Zhilin Yang, and Jie Tang. Controllable gener-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 126,
+ 491,
+ 507,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 491,
+ 507,
+ 507
+ ],
+ "score": 1.0,
+ "content": "ation from pre-trained language models via inverse prompting. arXiv preprint arXiv:2103.10685,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 127,
+ 503,
+ 154,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 503,
+ 154,
+ 515
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 521,
+ 507,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 507,
+ 537
+ ],
+ "score": 1.0,
+ "content": "[67] Yusheng Su, Xiaozhi Wang, Yujia Qin, Chi-Min Chan, Yankai Lin, Zhiyuan Liu, Peng Li,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 125,
+ 532,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 532,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "Juanzi Li, Lei Hou, Maosong Sun, et al. On transferability of prompt tuning for natural language",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 127,
+ 544,
+ 353,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 544,
+ 353,
+ 558
+ ],
+ "score": 1.0,
+ "content": "understanding. arXiv preprint arXiv:2111.06719, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 563,
+ 506,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 506,
+ 579
+ ],
+ "score": 1.0,
+ "content": "[68] Yun He, Huaixiu Steven Zheng, Yi Tay, Jai Gupta, Yu Du, Vamsi Aribandi, Zhe Zhao, YaGuang",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 126,
+ 574,
+ 506,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 574,
+ 506,
+ 591
+ ],
+ "score": 1.0,
+ "content": "Li, Zhao Chen, Donald Metzler, et al. HyperPrompt: Prompt-based task-conditioning of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 125,
+ 585,
+ 347,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 585,
+ 347,
+ 600
+ ],
+ "score": 1.0,
+ "content": "transformers. arXiv preprint arXiv:2203.00759, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "[69] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 126,
+ 617,
+ 460,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 617,
+ 460,
+ 633
+ ],
+ "score": 1.0,
+ "content": "and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 637,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "[70] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 127,
+ 648,
+ 418,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 648,
+ 418,
+ 662
+ ],
+ "score": 1.0,
+ "content": "and natural language inference. arXiv preprint arXiv:2001.07676, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 668,
+ 505,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 668,
+ 505,
+ 682
+ ],
+ "score": 1.0,
+ "content": "[71] Teven Le Scao and Alexander M. Rush. How many data points is a prompt worth? arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 124,
+ 679,
+ 267,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 679,
+ 267,
+ 695
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:2103.08493, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 104,
+ 699,
+ 506,
+ 716
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 699,
+ 506,
+ 716
+ ],
+ "score": 1.0,
+ "content": "[72] Sen Yang, Yunchen Zhang, Leyang Cui, and Yue Zhang. Do prompts solve NLP tasks using",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 125,
+ 709,
+ 366,
+ 725
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 709,
+ 366,
+ 725
+ ],
+ "score": 1.0,
+ "content": "natural language? arXiv preprint arXiv:2203.00902, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "page_idx": 13,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "14",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 104,
+ 54,
+ 507,
+ 727
+ ],
+ "lines": [],
+ "index": 22,
+ "bbox_fs": [
+ 104,
+ 71,
+ 507,
+ 725
+ ],
+ "lines_deleted": true
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 73,
+ 506,
+ 416
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 506,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 506,
+ 87
+ ],
+ "score": 1.0,
+ "content": "[73] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Auto-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 127,
+ 84,
+ 507,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 84,
+ 507,
+ 98
+ ],
+ "score": 1.0,
+ "content": "Prompt: Eliciting knowledge from language models with automatically generated prompts.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 95,
+ 290,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 95,
+ 290,
+ 108
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2010.15980, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "[74] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pre-trained language models better few-shot",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 127,
+ 126,
+ 328,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 126,
+ 328,
+ 138
+ ],
+ "score": 1.0,
+ "content": "learners. arXiv preprint arXiv:2012.15723, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 143,
+ 507,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 143,
+ 507,
+ 159
+ ],
+ "score": 1.0,
+ "content": "[75] Ningyu Zhang, Luoqiu Li, Xiang Chen, Shumin Deng, Zhen Bi, Chuanqi Tan, Fei Huang,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 127,
+ 156,
+ 505,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 156,
+ 505,
+ 169
+ ],
+ "score": 1.0,
+ "content": "and Huajun Chen. Differentiable prompt makes pre-trained language models better few-shot",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 127,
+ 167,
+ 328,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 167,
+ 328,
+ 179
+ ],
+ "score": 1.0,
+ "content": "learners. arXiv preprint arXiv:2108.13161, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 185,
+ 506,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 185,
+ 506,
+ 199
+ ],
+ "score": 1.0,
+ "content": "[76] Rabeeh Karimi Mahabadi, Luke Zettlemoyer, James Henderson, Marzieh Saeidi, Lambert",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 127,
+ 196,
+ 506,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 196,
+ 506,
+ 209
+ ],
+ "score": 1.0,
+ "content": "Mathias, Veselin Stoyanov, and Majid Yazdani. PERFECT: Prompt-free and efficient few-shot",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 126,
+ 207,
+ 419,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 207,
+ 419,
+ 221
+ ],
+ "score": 1.0,
+ "content": "learning with language models. arXiv preprint arXiv:2204.01172, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 227,
+ 505,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 227,
+ 505,
+ 241
+ ],
+ "score": 1.0,
+ "content": "[77] Nafise Sadat Moosavi, Quentin Delfosse, Kristian Kersting, and Iryna Gurevych. Adaptable",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 127,
+ 239,
+ 330,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 239,
+ 330,
+ 250
+ ],
+ "score": 1.0,
+ "content": "adapters. arXiv preprint arXiv:2205.01549, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 256,
+ 507,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 507,
+ 271
+ ],
+ "score": 1.0,
+ "content": "[78] Eleni Triantafillou, Hugo Larochelle, Richard Zemel, and Vincent Dumoulin. Learning a",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 127,
+ 268,
+ 507,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 268,
+ 507,
+ 282
+ ],
+ "score": 1.0,
+ "content": "universal template for few-shot dataset generalization. arXiv preprint arXiv:/2105.07029, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 302
+ ],
+ "score": 1.0,
+ "content": "[79] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E. Turner.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 127,
+ 298,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 298,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "Fast and flexible multi-task classification using conditional neural adaptive processes. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 126,
+ 310,
+ 265,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 310,
+ 265,
+ 322
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1906.07697, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 329,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "[80] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 126,
+ 338,
+ 506,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 338,
+ 506,
+ 354
+ ],
+ "score": 1.0,
+ "content": "domains for few-shot classification. Proceedings of the IEEE/CVF International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 127,
+ 351,
+ 227,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 351,
+ 227,
+ 363
+ ],
+ "score": 1.0,
+ "content": "Computer Vision., 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 384
+ ],
+ "score": 1.0,
+ "content": "[81] Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 127,
+ 380,
+ 506,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 380,
+ 506,
+ 395
+ ],
+ "score": 1.0,
+ "content": "Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 127,
+ 391,
+ 506,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 391,
+ 506,
+ 406
+ ],
+ "score": 1.0,
+ "content": "Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 127,
+ 403,
+ 498,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 403,
+ 498,
+ 415
+ ],
+ "score": 1.0,
+ "content": "model evaluation, September 2021. URL https://doi.org/10.5281/zenodo.5371628.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 435,
+ 156,
+ 448
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 158,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 158,
+ 450
+ ],
+ "score": 1.0,
+ "content": "Checklist",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 457,
+ 208,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 455,
+ 210,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 455,
+ 210,
+ 469
+ ],
+ "score": 1.0,
+ "content": "1. For all authors...",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 472,
+ 505,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 471,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 471,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 161,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 482,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "contributions and scope? [Yes] We compare different PEFT methods in Figure 2 and",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 162,
+ 493,
+ 357,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 493,
+ 357,
+ 505
+ ],
+ "score": 1.0,
+ "content": "demonstate PEFT outperforms ICL in section 4.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 145,
+ 506,
+ 416,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 506,
+ 416,
+ 519
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe the limitations of your work? [Yes] Section 5",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 146,
+ 519,
+ 465,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 519,
+ 465,
+ 532
+ ],
+ "score": 1.0,
+ "content": "(c) Did you discuss any potential negative societal impacts of your work? [No]",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 145,
+ 532,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 532,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 161,
+ 543,
+ 214,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 543,
+ 214,
+ 556
+ ],
+ "score": 1.0,
+ "content": "them? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 559,
+ 302,
+ 570
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 558,
+ 304,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 558,
+ 304,
+ 572
+ ],
+ "score": 1.0,
+ "content": "2. If you are including theoretical results...",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 573,
+ 452,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 573,
+ 453,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 573,
+ 453,
+ 586
+ ],
+ "score": 1.0,
+ "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 145,
+ 586,
+ 424,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 586,
+ 424,
+ 599
+ ],
+ "score": 1.0,
+ "content": "(b) Did you include complete proofs of all theoretical results? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 130,
+ 603,
+ 241,
+ 613
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 600,
+ 243,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 600,
+ 243,
+ 617
+ ],
+ "score": 1.0,
+ "content": "3. If you ran experiments...",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 617,
+ 505,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 617,
+ 507,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 617,
+ 507,
+ 630
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 161,
+ 628,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 628,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "mental results (either in the supplemental material or as a URL)? [Yes] Included code",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 162,
+ 639,
+ 227,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 639,
+ 227,
+ 650
+ ],
+ "score": 1.0,
+ "content": "with README",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 145,
+ 651,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 651,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 161,
+ 663,
+ 410,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 663,
+ 410,
+ 675
+ ],
+ "score": 1.0,
+ "content": "were chosen)? [Yes] Section 3.1, Section 3.5, Appendix D, H",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 147,
+ 675,
+ 507,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 675,
+ 507,
+ 689
+ ],
+ "score": 1.0,
+ "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 162,
+ 686,
+ 376,
+ 699
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 686,
+ 376,
+ 699
+ ],
+ "score": 1.0,
+ "content": "ments multiple times)? [Yes] Appendix C, D, E, F, G",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 146,
+ 698,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 698,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 162,
+ 711,
+ 421,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 711,
+ 421,
+ 723
+ ],
+ "score": 1.0,
+ "content": "of GPUs, internal cluster, or cloud provider)? [Yes] Appendix A",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "page_idx": 14,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "score": 1.0,
+ "content": "15",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 73,
+ 506,
+ 416
+ ],
+ "lines": [],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 72,
+ 507,
+ 415
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 435,
+ 156,
+ 448
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 158,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 158,
+ 450
+ ],
+ "score": 1.0,
+ "content": "Checklist",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 457,
+ 208,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 455,
+ 210,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 455,
+ 210,
+ 469
+ ],
+ "score": 1.0,
+ "content": "1. For all authors...",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 129,
+ 455,
+ 210,
+ 469
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 472,
+ 505,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 471,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 471,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s",
+ "type": "text"
+ }
+ ],
+ "index": 27,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 482,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "contributions and scope? [Yes] We compare different PEFT methods in Figure 2 and",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 162,
+ 493,
+ 357,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 493,
+ 357,
+ 505
+ ],
+ "score": 1.0,
+ "content": "demonstate PEFT outperforms ICL in section 4.",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 506,
+ 416,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 506,
+ 416,
+ 519
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe the limitations of your work? [Yes] Section 5",
+ "type": "text"
+ }
+ ],
+ "index": 30,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 519,
+ 465,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 519,
+ 465,
+ 532
+ ],
+ "score": 1.0,
+ "content": "(c) Did you discuss any potential negative societal impacts of your work? [No]",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 532,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 532,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to",
+ "type": "text"
+ }
+ ],
+ "index": 32,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 543,
+ 214,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 543,
+ 214,
+ 556
+ ],
+ "score": 1.0,
+ "content": "them? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 145,
+ 471,
+ 506,
+ 556
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 559,
+ 302,
+ 570
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 558,
+ 304,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 558,
+ 304,
+ 572
+ ],
+ "score": 1.0,
+ "content": "2. If you are including theoretical results...",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 129,
+ 558,
+ 304,
+ 572
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 573,
+ 452,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 573,
+ 453,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 573,
+ 453,
+ 586
+ ],
+ "score": 1.0,
+ "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 145,
+ 586,
+ 424,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 586,
+ 424,
+ 599
+ ],
+ "score": 1.0,
+ "content": "(b) Did you include complete proofs of all theoretical results? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 145,
+ 573,
+ 453,
+ 599
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 130,
+ 603,
+ 241,
+ 613
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 600,
+ 243,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 600,
+ 243,
+ 617
+ ],
+ "score": 1.0,
+ "content": "3. If you ran experiments...",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 128,
+ 600,
+ 243,
+ 617
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 617,
+ 505,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 617,
+ 507,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 617,
+ 507,
+ 630
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-",
+ "type": "text"
+ }
+ ],
+ "index": 38,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 628,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 628,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "mental results (either in the supplemental material or as a URL)? [Yes] Included code",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 162,
+ 639,
+ 227,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 639,
+ 227,
+ 650
+ ],
+ "score": 1.0,
+ "content": "with README",
+ "type": "text"
+ }
+ ],
+ "index": 40,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 651,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 651,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they",
+ "type": "text"
+ }
+ ],
+ "index": 41,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 663,
+ 410,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 663,
+ 410,
+ 675
+ ],
+ "score": 1.0,
+ "content": "were chosen)? [Yes] Section 3.1, Section 3.5, Appendix D, H",
+ "type": "text"
+ }
+ ],
+ "index": 42,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 147,
+ 675,
+ 507,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 675,
+ 507,
+ 689
+ ],
+ "score": 1.0,
+ "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-",
+ "type": "text"
+ }
+ ],
+ "index": 43,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 686,
+ 376,
+ 699
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 686,
+ 376,
+ 699
+ ],
+ "score": 1.0,
+ "content": "ments multiple times)? [Yes] Appendix C, D, E, F, G",
+ "type": "text"
+ }
+ ],
+ "index": 44,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 698,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 698,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type",
+ "type": "text"
+ }
+ ],
+ "index": 45,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 711,
+ 421,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 711,
+ 421,
+ 723
+ ],
+ "score": 1.0,
+ "content": "of GPUs, internal cluster, or cloud provider)? [Yes] Appendix A",
+ "type": "text"
+ }
+ ],
+ "index": 46,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 145,
+ 617,
+ 507,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 72,
+ 505,
+ 85
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 71,
+ 507,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 71,
+ 507,
+ 87
+ ],
+ "score": 1.0,
+ "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 87,
+ 505,
+ 195
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 87,
+ 505,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 87,
+ 505,
+ 100
+ ],
+ "score": 1.0,
+ "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] Section 3.1, Section",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 160,
+ 97,
+ 179,
+ 110
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 97,
+ 179,
+ 110
+ ],
+ "score": 1.0,
+ "content": "4.3",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 145,
+ 110,
+ 405,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 110,
+ 405,
+ 124
+ ],
+ "score": 1.0,
+ "content": "(b) Did you mention the license of the assets? [Yes] Section 3.1",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 145,
+ 123,
+ 506,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 123,
+ 506,
+ 137
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 144,
+ 147,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 147,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 162,
+ 159,
+ 249,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 159,
+ 249,
+ 172
+ ],
+ "score": 1.0,
+ "content": "using/curating? [No]",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 147,
+ 172,
+ 505,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 172,
+ 505,
+ 184
+ ],
+ "score": 1.0,
+ "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 162,
+ 183,
+ 320,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 183,
+ 320,
+ 196
+ ],
+ "score": 1.0,
+ "content": "information or offensive content? [No]",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 198,
+ 432,
+ 209
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 196,
+ 433,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 196,
+ 433,
+ 212
+ ],
+ "score": 1.0,
+ "content": "5. If you used crowdsourcing or conducted research with human subjects...",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 212,
+ 505,
+ 284
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 212,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 212,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the full text of instructions given to participants and screenshots, if",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 162,
+ 223,
+ 238,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 223,
+ 238,
+ 237
+ ],
+ "score": 1.0,
+ "content": "applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 145,
+ 236,
+ 505,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 236,
+ 505,
+ 249
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe any potential participant risks, with links to Institutional Review",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 161,
+ 247,
+ 342,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 247,
+ 342,
+ 260
+ ],
+ "score": 1.0,
+ "content": "Board (IRB) approvals, if applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 147,
+ 260,
+ 505,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 260,
+ 505,
+ 273
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 162,
+ 272,
+ 333,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 272,
+ 333,
+ 284
+ ],
+ "score": 1.0,
+ "content": "spent on participant compensation? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12.5
+ }
+ ],
+ "page_idx": 15,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "16",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 72,
+ 505,
+ 85
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 71,
+ 507,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 71,
+ 507,
+ 87
+ ],
+ "score": 1.0,
+ "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0,
+ "bbox_fs": [
+ 128,
+ 71,
+ 507,
+ 87
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 87,
+ 505,
+ 195
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 87,
+ 505,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 87,
+ 505,
+ 100
+ ],
+ "score": 1.0,
+ "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] Section 3.1, Section",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 160,
+ 97,
+ 179,
+ 110
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 97,
+ 179,
+ 110
+ ],
+ "score": 1.0,
+ "content": "4.3",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 110,
+ 405,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 110,
+ 405,
+ 124
+ ],
+ "score": 1.0,
+ "content": "(b) Did you mention the license of the assets? [Yes] Section 3.1",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 123,
+ 506,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 123,
+ 506,
+ 137
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 144,
+ 147,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 147,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 159,
+ 249,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 159,
+ 249,
+ 172
+ ],
+ "score": 1.0,
+ "content": "using/curating? [No]",
+ "type": "text"
+ }
+ ],
+ "index": 6,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 147,
+ 172,
+ 505,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 172,
+ 505,
+ 184
+ ],
+ "score": 1.0,
+ "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 183,
+ 320,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 183,
+ 320,
+ 196
+ ],
+ "score": 1.0,
+ "content": "information or offensive content? [No]",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 144,
+ 87,
+ 506,
+ 196
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 198,
+ 432,
+ 209
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 196,
+ 433,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 196,
+ 433,
+ 212
+ ],
+ "score": 1.0,
+ "content": "5. If you used crowdsourcing or conducted research with human subjects...",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 128,
+ 196,
+ 433,
+ 212
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 212,
+ 505,
+ 284
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 212,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 212,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the full text of instructions given to participants and screenshots, if",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 223,
+ 238,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 223,
+ 238,
+ 237
+ ],
+ "score": 1.0,
+ "content": "applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 236,
+ 505,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 236,
+ 505,
+ 249
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe any potential participant risks, with links to Institutional Review",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 247,
+ 342,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 247,
+ 342,
+ 260
+ ],
+ "score": 1.0,
+ "content": "Board (IRB) approvals, if applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 147,
+ 260,
+ 505,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 260,
+ 505,
+ 273
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 272,
+ 333,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 272,
+ 333,
+ 284
+ ],
+ "score": 1.0,
+ "content": "spent on participant compensation? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 15,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 145,
+ 212,
+ 506,
+ 284
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file