diff --git "a/parse/dev/UpN2wfrLec/UpN2wfrLec_middle.json" "b/parse/dev/UpN2wfrLec/UpN2wfrLec_middle.json" new file mode 100644--- /dev/null +++ "b/parse/dev/UpN2wfrLec/UpN2wfrLec_middle.json" @@ -0,0 +1,32364 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 115, + 97, + 497, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 96, + 497, + 120 + ], + "spans": [ + { + "bbox": [ + 113, + 96, + 497, + 120 + ], + "score": 1.0, + "content": "Language Is Not All You Need: Aligning Perception", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 219, + 118, + 391, + 138 + ], + "spans": [ + { + "bbox": [ + 219, + 118, + 391, + 138 + ], + "score": 1.0, + "content": "with Language Models", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 122, + 172, + 490, + 228 + ], + "lines": [ + { + "bbox": [ + 129, + 172, + 486, + 185 + ], + "spans": [ + { + "bbox": [ + 129, + 172, + 486, + 185 + ], + "score": 1.0, + "content": "Shaohan Huang∗, Li Dong∗, Wenhui Wang∗, Yaru Hao∗, Saksham Singhal∗, Shuming Ma∗", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 183, + 491, + 196 + ], + "spans": [ + { + "bbox": [ + 120, + 183, + 491, + 196 + ], + "score": 1.0, + "content": "Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 194, + 478, + 207 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 478, + 207 + ], + "score": 1.0, + "content": "Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, Furu Wei†", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 284, + 206, + 327, + 217 + ], + "spans": [ + { + "bbox": [ + 284, + 206, + 327, + 217 + ], + "score": 1.0, + "content": "Microsoft", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 213, + 216, + 395, + 229 + ], + "spans": [ + { + "bbox": [ + 213, + 216, + 395, + 229 + ], + "score": 1.0, + "content": "https://github.com/microsoft/unilm", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 283, + 245, + 328, + 258 + ], + "lines": [ + { + "bbox": [ + 281, + 244, + 331, + 260 + ], + "spans": [ + { + "bbox": [ + 281, + 244, + 331, + 260 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 270, + 469, + 456 + ], + "lines": [ + { + "bbox": [ + 141, + 270, + 471, + 283 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 471, + 283 + ], + "score": 1.0, + "content": "A big convergence of language, multimodal perception, action, and world model-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 281, + 470, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 470, + 293 + ], + "score": 1.0, + "content": "ing is a key step toward artificial general intelligence. In this work, we introduce", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 292, + 469, + 304 + ], + "spans": [ + { + "bbox": [ + 142, + 292, + 469, + 304 + ], + "score": 1.0, + "content": "KOSMOS-1, a Multimodal Large Language Model (MLLM) that can perceive", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 302, + 471, + 315 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 471, + 315 + ], + "score": 1.0, + "content": "general modalities, learn in context (i.e., few-shot), and follow instructions (i.e.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 314, + 470, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 314, + 470, + 325 + ], + "score": 1.0, + "content": "zero-shot). Specifically, we train KOSMOS-1 from scratch on web-scale multi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 324, + 469, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 324, + 469, + 337 + ], + "score": 1.0, + "content": "modal corpora, including arbitrarily interleaved text and images, image-caption", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 336, + 470, + 348 + ], + "spans": [ + { + "bbox": [ + 141, + 336, + 470, + 348 + ], + "score": 1.0, + "content": "pairs, and text data. We evaluate various settings, including zero-shot, few-shot,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 347, + 470, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 347, + 470, + 358 + ], + "score": 1.0, + "content": "and multimodal chain-of-thought prompting, on a wide range of tasks without", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 357, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 357, + 470, + 369 + ], + "score": 1.0, + "content": "any gradient updates or finetuning. Experimental results show that KOSMOS-1", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 368, + 470, + 381 + ], + "spans": [ + { + "bbox": [ + 141, + 368, + 470, + 381 + ], + "score": 1.0, + "content": "achieves impressive performance on (i) language understanding, generation, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "score": 1.0, + "content": "even OCR-free NLP (directly fed with document images), (ii) perception-language", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 389, + 470, + 403 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 403 + ], + "score": 1.0, + "content": "tasks, including multimodal dialogue, image captioning, visual question answering,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 401, + 469, + 414 + ], + "spans": [ + { + "bbox": [ + 141, + 401, + 469, + 414 + ], + "score": 1.0, + "content": "and (iii) vision tasks, such as image recognition with descriptions (specifying", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 412, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 142, + 412, + 469, + 423 + ], + "score": 1.0, + "content": "classification via text instructions). We also show that MLLMs can benefit from", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 423, + 469, + 434 + ], + "spans": [ + { + "bbox": [ + 142, + 423, + 469, + 434 + ], + "score": 1.0, + "content": "cross-modal transfer, i.e., transfer knowledge from language to multimodal, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 434, + 470, + 446 + ], + "spans": [ + { + "bbox": [ + 142, + 434, + 470, + 446 + ], + "score": 1.0, + "content": "from multimodal to language. In addition, we introduce a dataset of Raven IQ test,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 445, + 399, + 457 + ], + "spans": [ + { + "bbox": [ + 142, + 445, + 399, + 457 + ], + "score": 1.0, + "content": "which diagnoses the nonverbal reasoning capability of MLLMs.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 474, + 321, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 322, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 322, + 490 + ], + "score": 1.0, + "content": "1 Introduction: From LLMs to MLLMs", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 499, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "Large language models (LLMs) have successfully served as a general-purpose interface across various", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "natural language tasks [1]. The LLM-based interface can be adapted to a task as long as we are able", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "to transform the input and output into texts. For example, the input of the summarization task is a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "document and the output is its summary. So we can feed the input document into the language model", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 543, + 275, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 275, + 557 + ], + "score": 1.0, + "content": "and then produce the generated summary.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "Despite the successful applications in natural language processing, it is still struggling to natively use", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "LLMs for multimodal data, such as image, and audio. Being a basic part of intelligence, multimodal", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "perception is a necessity to achieve artificial general intelligence, in terms of knowledge acquisition", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "and grounding to the real world. More importantly, unlocking multimodal input [2, 3, 4, 5, 6, 7]", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 603, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 616 + ], + "score": 1.0, + "content": "greatly widens the applications of language models to more high-value areas, such as multimodal", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 615, + 327, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 327, + 626 + ], + "score": 1.0, + "content": "machine learning, document intelligence, and robotics.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 631, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "In this work, we introduce KOSMOS-1, a Multimodal Large Language Model (MLLM) that can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 641, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 104, + 641, + 507, + 654 + ], + "score": 1.0, + "content": "perceive general modalities, follow instructions (i.e., zero-shot learning), and learn in context (i.e.,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 652, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 506, + 665 + ], + "score": 1.0, + "content": "few-shot learning). The goal is to align perception with LLMs, so that the models are able to see and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 664, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 675 + ], + "score": 1.0, + "content": "talk. To be specific, we follow METALM [3] to train the KOSMOS-1 model from scratch. As shown", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 675, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 505, + 686 + ], + "score": 1.0, + "content": "in Figure 1, a Transformer-based language model is regarded as the general-purpose interface, and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 685, + 506, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 506, + 697 + ], + "score": 1.0, + "content": "perception modules are docked with the language model. We train the model on web-scale multimodal", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 704, + 284, + 714 + ], + "lines": [ + { + "bbox": [ + 118, + 701, + 286, + 716 + ], + "spans": [ + { + "bbox": [ + 118, + 701, + 286, + 716 + ], + "score": 1.0, + "content": "∗ Equal contribution. † Corresponding author.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "score": 1.0, + "content": "37th Conference on Neural Information Processing Systems (NeurIPS 2023).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 115, + 97, + 497, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 96, + 497, + 120 + ], + "spans": [ + { + "bbox": [ + 113, + 96, + 497, + 120 + ], + "score": 1.0, + "content": "Language Is Not All You Need: Aligning Perception", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 219, + 118, + 391, + 138 + ], + "spans": [ + { + "bbox": [ + 219, + 118, + 391, + 138 + ], + "score": 1.0, + "content": "with Language Models", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 122, + 172, + 490, + 228 + ], + "lines": [ + { + "bbox": [ + 129, + 172, + 486, + 185 + ], + "spans": [ + { + "bbox": [ + 129, + 172, + 486, + 185 + ], + "score": 1.0, + "content": "Shaohan Huang∗, Li Dong∗, Wenhui Wang∗, Yaru Hao∗, Saksham Singhal∗, Shuming Ma∗", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 183, + 491, + 196 + ], + "spans": [ + { + "bbox": [ + 120, + 183, + 491, + 196 + ], + "score": 1.0, + "content": "Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 194, + 478, + 207 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 478, + 207 + ], + "score": 1.0, + "content": "Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, Furu Wei†", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 284, + 206, + 327, + 217 + ], + "spans": [ + { + "bbox": [ + 284, + 206, + 327, + 217 + ], + "score": 1.0, + "content": "Microsoft", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 213, + 216, + 395, + 229 + ], + "spans": [ + { + "bbox": [ + 213, + 216, + 395, + 229 + ], + "score": 1.0, + "content": "https://github.com/microsoft/unilm", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 120, + 172, + 491, + 229 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 245, + 328, + 258 + ], + "lines": [ + { + "bbox": [ + 281, + 244, + 331, + 260 + ], + "spans": [ + { + "bbox": [ + 281, + 244, + 331, + 260 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 270, + 469, + 456 + ], + "lines": [ + { + "bbox": [ + 141, + 270, + 471, + 283 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 471, + 283 + ], + "score": 1.0, + "content": "A big convergence of language, multimodal perception, action, and world model-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 281, + 470, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 470, + 293 + ], + "score": 1.0, + "content": "ing is a key step toward artificial general intelligence. In this work, we introduce", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 292, + 469, + 304 + ], + "spans": [ + { + "bbox": [ + 142, + 292, + 469, + 304 + ], + "score": 1.0, + "content": "KOSMOS-1, a Multimodal Large Language Model (MLLM) that can perceive", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 302, + 471, + 315 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 471, + 315 + ], + "score": 1.0, + "content": "general modalities, learn in context (i.e., few-shot), and follow instructions (i.e.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 314, + 470, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 314, + 470, + 325 + ], + "score": 1.0, + "content": "zero-shot). Specifically, we train KOSMOS-1 from scratch on web-scale multi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 324, + 469, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 324, + 469, + 337 + ], + "score": 1.0, + "content": "modal corpora, including arbitrarily interleaved text and images, image-caption", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 336, + 470, + 348 + ], + "spans": [ + { + "bbox": [ + 141, + 336, + 470, + 348 + ], + "score": 1.0, + "content": "pairs, and text data. We evaluate various settings, including zero-shot, few-shot,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 347, + 470, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 347, + 470, + 358 + ], + "score": 1.0, + "content": "and multimodal chain-of-thought prompting, on a wide range of tasks without", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 357, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 357, + 470, + 369 + ], + "score": 1.0, + "content": "any gradient updates or finetuning. Experimental results show that KOSMOS-1", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 368, + 470, + 381 + ], + "spans": [ + { + "bbox": [ + 141, + 368, + 470, + 381 + ], + "score": 1.0, + "content": "achieves impressive performance on (i) language understanding, generation, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "score": 1.0, + "content": "even OCR-free NLP (directly fed with document images), (ii) perception-language", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 389, + 470, + 403 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 403 + ], + "score": 1.0, + "content": "tasks, including multimodal dialogue, image captioning, visual question answering,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 401, + 469, + 414 + ], + "spans": [ + { + "bbox": [ + 141, + 401, + 469, + 414 + ], + "score": 1.0, + "content": "and (iii) vision tasks, such as image recognition with descriptions (specifying", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 412, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 142, + 412, + 469, + 423 + ], + "score": 1.0, + "content": "classification via text instructions). We also show that MLLMs can benefit from", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 423, + 469, + 434 + ], + "spans": [ + { + "bbox": [ + 142, + 423, + 469, + 434 + ], + "score": 1.0, + "content": "cross-modal transfer, i.e., transfer knowledge from language to multimodal, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 434, + 470, + 446 + ], + "spans": [ + { + "bbox": [ + 142, + 434, + 470, + 446 + ], + "score": 1.0, + "content": "from multimodal to language. In addition, we introduce a dataset of Raven IQ test,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 445, + 399, + 457 + ], + "spans": [ + { + "bbox": [ + 142, + 445, + 399, + 457 + ], + "score": 1.0, + "content": "which diagnoses the nonverbal reasoning capability of MLLMs.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16, + "bbox_fs": [ + 141, + 270, + 471, + 457 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 474, + 321, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 322, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 322, + 490 + ], + "score": 1.0, + "content": "1 Introduction: From LLMs to MLLMs", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 499, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "Large language models (LLMs) have successfully served as a general-purpose interface across various", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "natural language tasks [1]. The LLM-based interface can be adapted to a task as long as we are able", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "to transform the input and output into texts. For example, the input of the summarization task is a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "document and the output is its summary. So we can feed the input document into the language model", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 543, + 275, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 275, + 557 + ], + "score": 1.0, + "content": "and then produce the generated summary.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 500, + 505, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "Despite the successful applications in natural language processing, it is still struggling to natively use", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "LLMs for multimodal data, such as image, and audio. Being a basic part of intelligence, multimodal", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "perception is a necessity to achieve artificial general intelligence, in terms of knowledge acquisition", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "and grounding to the real world. More importantly, unlocking multimodal input [2, 3, 4, 5, 6, 7]", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 603, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 616 + ], + "score": 1.0, + "content": "greatly widens the applications of language models to more high-value areas, such as multimodal", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 615, + 327, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 327, + 626 + ], + "score": 1.0, + "content": "machine learning, document intelligence, and robotics.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 559, + 506, + 626 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 631, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "In this work, we introduce KOSMOS-1, a Multimodal Large Language Model (MLLM) that can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 641, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 104, + 641, + 507, + 654 + ], + "score": 1.0, + "content": "perceive general modalities, follow instructions (i.e., zero-shot learning), and learn in context (i.e.,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 652, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 506, + 665 + ], + "score": 1.0, + "content": "few-shot learning). The goal is to align perception with LLMs, so that the models are able to see and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 664, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 675 + ], + "score": 1.0, + "content": "talk. To be specific, we follow METALM [3] to train the KOSMOS-1 model from scratch. As shown", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 675, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 505, + 686 + ], + "score": 1.0, + "content": "in Figure 1, a Transformer-based language model is regarded as the general-purpose interface, and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 685, + 506, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 506, + 697 + ], + "score": 1.0, + "content": "perception modules are docked with the language model. We train the model on web-scale multimodal", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 313, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 328 + ], + "score": 1.0, + "content": "corpora, i.e., text data, arbitrarily interleaved images and texts, and image-caption pairs. In addition,", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 326, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 338 + ], + "score": 1.0, + "content": "we calibrate the instruction-following capability across modalities by transferring language-only data.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 630, + 507, + 697 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 122, + 67, + 489, + 238 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 67, + 489, + 238 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 67, + 489, + 238 + ], + "spans": [ + { + "bbox": [ + 122, + 67, + 489, + 238 + ], + "score": 0.426, + "type": "image", + "image_path": "8e3232a1c4f23f7b25cb3ce3da7195272b497bbf8c8cfd492e0f378fc617ec3b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 122, + 67, + 489, + 124.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 122, + 124.0, + 489, + 181.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 181.0, + 489, + 238.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 250, + 506, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "Figure 1: KOSMOS-1 is a multimodal large language model (MLLM) that is capable of perceiving", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 260, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 506, + 274 + ], + "score": 1.0, + "content": "multimodal input, following instructions, and performing in-context learning for not only language", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 271, + 507, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 507, + 285 + ], + "score": 1.0, + "content": "tasks but also multimodal tasks. In this work, we align vision with large language models (LLMs),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 283, + 321, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 321, + 295 + ], + "score": 1.0, + "content": "advancing the trend of going from LLMs to MLLMs.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 314, + 504, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 328 + ], + "score": 1.0, + "content": "corpora, i.e., text data, arbitrarily interleaved images and texts, and image-caption pairs. In addition,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 326, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 338 + ], + "score": 1.0, + "content": "we calibrate the instruction-following capability across modalities by transferring language-only data.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 341, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 355 + ], + "score": 1.0, + "content": "The KOSMOS-1 model natively supports language, perception-language, and vision tasks. In addition", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "score": 1.0, + "content": "to various natural language tasks, the KOSMOS-1 models natively handle a wide range of perception-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "intensive tasks, spanning visual dialogue, visual explanation, visual question answering, image", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "captioning, simple math equation, OCR, and zero-shot image classification with descriptions. We", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "also build an IQ test benchmark following Raven’s Progressive Matrices [8, 9], which evaluates", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "the capability of nonverbal reasoning for MLLMs. The examples show that the native support of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "multimodal perception enables new opportunities to apply LLMs to new tasks. Moreover, we show", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "score": 1.0, + "content": "that MLLMs achieve better commonsense reasoning performance compared with LLMs, which", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 428, + 346, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 346, + 442 + ], + "score": 1.0, + "content": "indicates cross-modal transfer helps knowledge acquisition.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 108, + 446, + 244, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 445, + 245, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 245, + 459 + ], + "score": 1.0, + "content": "The key takeaways are as follows:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "From LLMs to MLLMs. Properly handling perception is a necessary step toward artificial general", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "intelligence. The capability of perceiving multimodal input is critical to LLMs. First, multimodal", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "perception enables LLMs to acquire commonsense knowledge beyond text descriptions. Second,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "aligning perception with LLMs opens the door to new tasks, such as robotics, and document intelli-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "gence. Third, the capability of perception unifies various APIs, as graphical user interfaces are the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "most natural and unified way to interact with. For example, MLLMs can directly read the screen or", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 533, + 507, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 507, + 547 + ], + "score": 1.0, + "content": "extract numbers from receipts. We train the KOSMOS-1 models on web-scale multimodal corpora,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "which ensures that the model robustly learns from diverse sources. We not only use a large-scale", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "text corpus but also mine high-quality image-caption pairs and arbitrarily interleaved image and text", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 567, + 210, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 210, + 578 + ], + "score": 1.0, + "content": "documents from the web.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 602 + ], + "score": 1.0, + "content": "Language models as general-purpose interfaces. Following the philosophy proposed in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "METALM [3], we regard language models as a universal task layer. Because of the open-ended output", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "space, we are able to unify various task predictions as texts. Moreover, natural-language instructions", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 623, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 635 + ], + "score": 1.0, + "content": "and action sequences (such as programming language) can be well handled by language models.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "LLMs also serve as basic reasoners [10], which is complementary to perception modules on complex", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "tasks. So it is natural to align world, action, and multimodal perception with the general-purpose", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 655, + 236, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 236, + 668 + ], + "score": 1.0, + "content": "interface, i.e., language models.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "New capabilities of MLLMs. Apart from the capabilities found in previous LLMs [1, 11], MLLMs", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 687, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 703 + ], + "score": 1.0, + "content": "enable new usages and possibilities. First, we can conduct zero- and few-shot multimodal learning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "by using natural language instructions and demonstration examples. Second, we observe promising", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "signals of nonverbal reasoning by evaluating the Raven IQ test, which measures the fluid reasoning", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 122, + 67, + 489, + 238 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 67, + 489, + 238 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 67, + 489, + 238 + ], + "spans": [ + { + "bbox": [ + 122, + 67, + 489, + 238 + ], + "score": 0.426, + "type": "image", + "image_path": "8e3232a1c4f23f7b25cb3ce3da7195272b497bbf8c8cfd492e0f378fc617ec3b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 122, + 67, + 489, + 124.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 122, + 124.0, + 489, + 181.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 181.0, + 489, + 238.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 250, + 506, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "Figure 1: KOSMOS-1 is a multimodal large language model (MLLM) that is capable of perceiving", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 260, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 506, + 274 + ], + "score": 1.0, + "content": "multimodal input, following instructions, and performing in-context learning for not only language", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 271, + 507, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 507, + 285 + ], + "score": 1.0, + "content": "tasks but also multimodal tasks. In this work, we align vision with large language models (LLMs),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 283, + 321, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 321, + 295 + ], + "score": 1.0, + "content": "advancing the trend of going from LLMs to MLLMs.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 314, + 504, + 337 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 313, + 506, + 338 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 341, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 355 + ], + "score": 1.0, + "content": "The KOSMOS-1 model natively supports language, perception-language, and vision tasks. In addition", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "score": 1.0, + "content": "to various natural language tasks, the KOSMOS-1 models natively handle a wide range of perception-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "intensive tasks, spanning visual dialogue, visual explanation, visual question answering, image", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "captioning, simple math equation, OCR, and zero-shot image classification with descriptions. We", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "also build an IQ test benchmark following Raven’s Progressive Matrices [8, 9], which evaluates", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "the capability of nonverbal reasoning for MLLMs. The examples show that the native support of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "multimodal perception enables new opportunities to apply LLMs to new tasks. Moreover, we show", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "score": 1.0, + "content": "that MLLMs achieve better commonsense reasoning performance compared with LLMs, which", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 428, + 346, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 346, + 442 + ], + "score": 1.0, + "content": "indicates cross-modal transfer helps knowledge acquisition.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 341, + 506, + 442 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 446, + 244, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 445, + 245, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 245, + 459 + ], + "score": 1.0, + "content": "The key takeaways are as follows:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 445, + 245, + 459 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "From LLMs to MLLMs. Properly handling perception is a necessary step toward artificial general", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "intelligence. The capability of perceiving multimodal input is critical to LLMs. First, multimodal", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "perception enables LLMs to acquire commonsense knowledge beyond text descriptions. Second,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "aligning perception with LLMs opens the door to new tasks, such as robotics, and document intelli-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "gence. Third, the capability of perception unifies various APIs, as graphical user interfaces are the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "most natural and unified way to interact with. For example, MLLMs can directly read the screen or", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 533, + 507, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 507, + 547 + ], + "score": 1.0, + "content": "extract numbers from receipts. We train the KOSMOS-1 models on web-scale multimodal corpora,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "which ensures that the model robustly learns from diverse sources. We not only use a large-scale", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "text corpus but also mine high-quality image-caption pairs and arbitrarily interleaved image and text", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 567, + 210, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 210, + 578 + ], + "score": 1.0, + "content": "documents from the web.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 468, + 507, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 602 + ], + "score": 1.0, + "content": "Language models as general-purpose interfaces. Following the philosophy proposed in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "METALM [3], we regard language models as a universal task layer. Because of the open-ended output", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "space, we are able to unify various task predictions as texts. Moreover, natural-language instructions", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 623, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 635 + ], + "score": 1.0, + "content": "and action sequences (such as programming language) can be well handled by language models.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "LLMs also serve as basic reasoners [10], which is complementary to perception modules on complex", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "tasks. So it is natural to align world, action, and multimodal perception with the general-purpose", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 655, + 236, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 236, + 668 + ], + "score": 1.0, + "content": "interface, i.e., language models.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 589, + 506, + 668 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "New capabilities of MLLMs. Apart from the capabilities found in previous LLMs [1, 11], MLLMs", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 687, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 703 + ], + "score": 1.0, + "content": "enable new usages and possibilities. First, we can conduct zero- and few-shot multimodal learning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "by using natural language instructions and demonstration examples. Second, we observe promising", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "signals of nonverbal reasoning by evaluating the Raven IQ test, which measures the fluid reasoning", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 71, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 507, + 86 + ], + "score": 1.0, + "content": "ability of humans. Third, MLLMs naturally support multi-turn interactions for general modalities,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 226, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 226, + 96 + ], + "score": 1.0, + "content": "such as multimodal dialogue.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 37.5, + "bbox_fs": [ + 104, + 678, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 505, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 507, + 86 + ], + "score": 1.0, + "content": "ability of humans. Third, MLLMs naturally support multi-turn interactions for general modalities,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 226, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 226, + 96 + ], + "score": 1.0, + "content": "such as multimodal dialogue.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 107, + 110, + 386, + 125 + ], + "lines": [ + { + "bbox": [ + 104, + 109, + 388, + 127 + ], + "spans": [ + { + "bbox": [ + 104, + 109, + 388, + 127 + ], + "score": 1.0, + "content": "2 KOSMOS-1: A Multimodal Large Language Model", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 136, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 136, + 507, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 507, + 149 + ], + "score": 1.0, + "content": "KOSMOS-1 is a multimodal language model that can perceive general modalities, follow instructions,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 148, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 159 + ], + "score": 1.0, + "content": "learn in context, and generate outputs. Given the previous context, the model learns to generate", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 170 + ], + "score": 1.0, + "content": "texts in an auto-regressive manner. Specifically, the backbone of KOSMOS-1 is a Transformer-based", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 169, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 182 + ], + "score": 1.0, + "content": "causal language model. Apart from text, other modalities are embedded and fed into the language", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 193 + ], + "score": 1.0, + "content": "model. The Transformer decoder serves as a general-purpose interface to multimodal input. We", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 192, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 203 + ], + "score": 1.0, + "content": "train KOSMOS-1 on multimodal corpora, including monomodal data, cross-modal paired data, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "interleaved multimodal data. Once the models are trained, we can directly evaluate the models in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 214, + 416, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 416, + 225 + ], + "score": 1.0, + "content": "zero-shot and few-shot settings on both language tasks and multimodal tasks.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 107, + 237, + 221, + 250 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 222, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 222, + 252 + ], + "score": 1.0, + "content": "2.1 Input Representation", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "The Transformer decoder perceives general modalities in a unified way. For input format, we flatten", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 387, + 282 + ], + "score": 1.0, + "content": "input as a sequence decorated with special tokens. Specifically, we use", + "type": "text" + }, + { + "bbox": [ + 387, + 271, + 404, + 280 + ], + "score": 0.77, + "content": "\\mathtt { < s > }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 268, + 422, + 282 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 422, + 270, + 444, + 280 + ], + "score": 0.89, + "content": "< / { \\mathsf { s } } { \\mathsf { > } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "to denote start-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "and end-of-sequence. The special tokens and indicate the beginning and end of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 275, + 304 + ], + "score": 1.0, + "content": "encoded image embeddings. For example,", + "type": "text" + }, + { + "bbox": [ + 275, + 292, + 296, + 301 + ], + "score": 0.56, + "content": "\\ \" < \\mathsf { s } >", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 291, + 336, + 304 + ], + "score": 1.0, + "content": "document", + "type": "text" + }, + { + "bbox": [ + 337, + 291, + 364, + 302 + ], + "score": 0.82, + "content": "< / { \\mathsf { s } } > ^ { , , }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 291, + 438, + 304 + ], + "score": 1.0, + "content": "is a text input, and", + "type": "text" + }, + { + "bbox": [ + 439, + 292, + 460, + 302 + ], + "score": 0.68, + "content": "\\ \" < \\mathsf { s } >", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "paragraph", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 302, + 474, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 309, + 316 + ], + "score": 1.0, + "content": " Image Embedding paragraph", + "type": "text" + }, + { + "bbox": [ + 309, + 303, + 336, + 313 + ], + "score": 0.83, + "content": "< / { \\mathsf { s } } > ^ { , \\mathsf { \\curlyeq } }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 302, + 474, + 316 + ], + "score": 1.0, + "content": "is an interleaved image-text input.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 318, + 505, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "An embedding module is used to encode both text tokens and other input modalities into vectors.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "Then the embeddings are fed into the decoder. For text tokens, we use a lookup table to map them", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "into embeddings. For the modalities of continuous signals (e.g., image, and audio), it is also feasible", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "score": 1.0, + "content": "to represent inputs as discrete code and then regard them as “foreign languages” [4, 12]. In this work,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "following [3], we employ a vision encoder as the embedding module for input images. In addition,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 372, + 507, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 507, + 386 + ], + "score": 1.0, + "content": "Resampler [5] is used as an attentive pooling mechanism to reduce the number of image embeddings.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 107, + 397, + 332, + 410 + ], + "lines": [ + { + "bbox": [ + 104, + 396, + 333, + 413 + ], + "spans": [ + { + "bbox": [ + 104, + 396, + 333, + 413 + ], + "score": 1.0, + "content": "2.2 Multimodal Large Language Models (MLLMs)", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 504, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 504, + 431 + ], + "score": 1.0, + "content": "After obtaining the embeddings of an input sequence, we feed them into the Transformer-based", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "decoder. The left-to-right causal model processes the sequence in an auto-regressive manner, which", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "produces the next token by conditioning on past timesteps. The causal masking is used to mask", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 504, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 504, + 464 + ], + "score": 1.0, + "content": "out future information. A softmax classifier upon Transformer is used to generate tokens over the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 155, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 155, + 476 + ], + "score": 1.0, + "content": "vocabulary.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "MLLMs serve as general-purpose interfaces [3] that can perform interactions with both natural", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "language and multimodal input. The framework is flexible to handle various data types, as long as we", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "can represent input as vectors. MLLMs combine the best of two worlds. First, the language models", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "naturally inherit the capabilities of in-context learning and instruction following. Second, perception", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 522, + 379, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 379, + 535 + ], + "score": 1.0, + "content": "is aligned with language models by training on multimodal corpora.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 538, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "The implementation is based on the library TorchScale [13], which is designed for large-scale", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "model training. Compared with the standard Transformer architecture, we include the following", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "modifications: We use MAGNETO [14], a Transformer variant, as the backbone architecture and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 570, + 391, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 391, + 585 + ], + "score": 1.0, + "content": "XPOS [15] relative position encoding for better long-context modeling.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 107, + 596, + 243, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 244, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 244, + 610 + ], + "score": 1.0, + "content": "2.3 Multimodal Training Data", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 616, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "The models are trained on web-scale multimodal corpora. The training datasets consist of text corpora,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 628, + 355, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 355, + 640 + ], + "score": 1.0, + "content": "image-caption pairs, and interleaved data of images and texts.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 651, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 506, + 664 + ], + "score": 1.0, + "content": "Text Corpora We train our model with The Pile [16] and Common Crawl (CC). The Pile is a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "score": 1.0, + "content": "massive English text dataset built for training large-scale language models. We exclude data splits", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "from GitHub, arXiv, Stack Exchange, and PubMed Central. We also include the Common Crawl", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 684, + 459, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 459, + 697 + ], + "score": 1.0, + "content": "snapshots (2020-50 and 2021-04) datasets, CC-Stories, and RealNews datasets [17, 18].", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "Image-Caption Pairs The image-caption pairs are constructed from several datasets, including", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "English LAION-2B [19], LAION-400M [20], COYO-700M [21], and Conceptual Captions [22, 23].", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 505, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 71, + 507, + 96 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 110, + 386, + 125 + ], + "lines": [ + { + "bbox": [ + 104, + 109, + 388, + 127 + ], + "spans": [ + { + "bbox": [ + 104, + 109, + 388, + 127 + ], + "score": 1.0, + "content": "2 KOSMOS-1: A Multimodal Large Language Model", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 136, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 136, + 507, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 507, + 149 + ], + "score": 1.0, + "content": "KOSMOS-1 is a multimodal language model that can perceive general modalities, follow instructions,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 148, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 159 + ], + "score": 1.0, + "content": "learn in context, and generate outputs. Given the previous context, the model learns to generate", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 170 + ], + "score": 1.0, + "content": "texts in an auto-regressive manner. Specifically, the backbone of KOSMOS-1 is a Transformer-based", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 169, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 182 + ], + "score": 1.0, + "content": "causal language model. Apart from text, other modalities are embedded and fed into the language", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 193 + ], + "score": 1.0, + "content": "model. The Transformer decoder serves as a general-purpose interface to multimodal input. We", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 192, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 203 + ], + "score": 1.0, + "content": "train KOSMOS-1 on multimodal corpora, including monomodal data, cross-modal paired data, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "interleaved multimodal data. Once the models are trained, we can directly evaluate the models in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 214, + 416, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 416, + 225 + ], + "score": 1.0, + "content": "zero-shot and few-shot settings on both language tasks and multimodal tasks.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 136, + 507, + 225 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 237, + 221, + 250 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 222, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 222, + 252 + ], + "score": 1.0, + "content": "2.1 Input Representation", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "The Transformer decoder perceives general modalities in a unified way. For input format, we flatten", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 387, + 282 + ], + "score": 1.0, + "content": "input as a sequence decorated with special tokens. Specifically, we use", + "type": "text" + }, + { + "bbox": [ + 387, + 271, + 404, + 280 + ], + "score": 0.77, + "content": "\\mathtt { < s > }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 268, + 422, + 282 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 422, + 270, + 444, + 280 + ], + "score": 0.89, + "content": "< / { \\mathsf { s } } { \\mathsf { > } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "to denote start-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "and end-of-sequence. The special tokens and indicate the beginning and end of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 275, + 304 + ], + "score": 1.0, + "content": "encoded image embeddings. For example,", + "type": "text" + }, + { + "bbox": [ + 275, + 292, + 296, + 301 + ], + "score": 0.56, + "content": "\\ \" < \\mathsf { s } >", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 291, + 336, + 304 + ], + "score": 1.0, + "content": "document", + "type": "text" + }, + { + "bbox": [ + 337, + 291, + 364, + 302 + ], + "score": 0.82, + "content": "< / { \\mathsf { s } } > ^ { , , }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 291, + 438, + 304 + ], + "score": 1.0, + "content": "is a text input, and", + "type": "text" + }, + { + "bbox": [ + 439, + 292, + 460, + 302 + ], + "score": 0.68, + "content": "\\ \" < \\mathsf { s } >", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "paragraph", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 302, + 474, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 309, + 316 + ], + "score": 1.0, + "content": " Image Embedding paragraph", + "type": "text" + }, + { + "bbox": [ + 309, + 303, + 336, + 313 + ], + "score": 0.83, + "content": "< / { \\mathsf { s } } > ^ { , \\mathsf { \\curlyeq } }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 302, + 474, + 316 + ], + "score": 1.0, + "content": "is an interleaved image-text input.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 258, + 506, + 316 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 318, + 505, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "An embedding module is used to encode both text tokens and other input modalities into vectors.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "Then the embeddings are fed into the decoder. For text tokens, we use a lookup table to map them", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "into embeddings. For the modalities of continuous signals (e.g., image, and audio), it is also feasible", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "score": 1.0, + "content": "to represent inputs as discrete code and then regard them as ���foreign languages” [4, 12]. In this work,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "following [3], we employ a vision encoder as the embedding module for input images. In addition,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 372, + 507, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 507, + 386 + ], + "score": 1.0, + "content": "Resampler [5] is used as an attentive pooling mechanism to reduce the number of image embeddings.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 317, + 507, + 386 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 397, + 332, + 410 + ], + "lines": [ + { + "bbox": [ + 104, + 396, + 333, + 413 + ], + "spans": [ + { + "bbox": [ + 104, + 396, + 333, + 413 + ], + "score": 1.0, + "content": "2.2 Multimodal Large Language Models (MLLMs)", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 504, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 504, + 431 + ], + "score": 1.0, + "content": "After obtaining the embeddings of an input sequence, we feed them into the Transformer-based", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "decoder. The left-to-right causal model processes the sequence in an auto-regressive manner, which", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "produces the next token by conditioning on past timesteps. The causal masking is used to mask", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 504, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 504, + 464 + ], + "score": 1.0, + "content": "out future information. A softmax classifier upon Transformer is used to generate tokens over the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 155, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 155, + 476 + ], + "score": 1.0, + "content": "vocabulary.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 417, + 506, + 476 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "MLLMs serve as general-purpose interfaces [3] that can perform interactions with both natural", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "language and multimodal input. The framework is flexible to handle various data types, as long as we", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "can represent input as vectors. MLLMs combine the best of two worlds. First, the language models", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "naturally inherit the capabilities of in-context learning and instruction following. Second, perception", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 522, + 379, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 379, + 535 + ], + "score": 1.0, + "content": "is aligned with language models by training on multimodal corpora.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 478, + 505, + 535 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 538, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "The implementation is based on the library TorchScale [13], which is designed for large-scale", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "model training. Compared with the standard Transformer architecture, we include the following", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "modifications: We use MAGNETO [14], a Transformer variant, as the backbone architecture and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 570, + 391, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 391, + 585 + ], + "score": 1.0, + "content": "XPOS [15] relative position encoding for better long-context modeling.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 538, + 505, + 585 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 596, + 243, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 244, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 244, + 610 + ], + "score": 1.0, + "content": "2.3 Multimodal Training Data", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 616, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "The models are trained on web-scale multimodal corpora. The training datasets consist of text corpora,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 628, + 355, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 355, + 640 + ], + "score": 1.0, + "content": "image-caption pairs, and interleaved data of images and texts.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 615, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 651, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 506, + 664 + ], + "score": 1.0, + "content": "Text Corpora We train our model with The Pile [16] and Common Crawl (CC). The Pile is a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "score": 1.0, + "content": "massive English text dataset built for training large-scale language models. We exclude data splits", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "from GitHub, arXiv, Stack Exchange, and PubMed Central. We also include the Common Crawl", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 684, + 459, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 459, + 697 + ], + "score": 1.0, + "content": "snapshots (2020-50 and 2021-04) datasets, CC-Stories, and RealNews datasets [17, 18].", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 651, + 506, + 697 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "Image-Caption Pairs The image-caption pairs are constructed from several datasets, including", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "English LAION-2B [19], LAION-400M [20], COYO-700M [21], and Conceptual Captions [22, 23].", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 106, + 699, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "English LAION-2B, LAION-400M, and COYO-700M are collected from web pages of the Common", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 97 + ], + "score": 1.0, + "content": "Crawl web data by extracting image sources and the corresponding alt-text. Conceptual Captions are", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 227, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 227, + 108 + ], + "score": 1.0, + "content": "also from internet web pages.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 113, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 113, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 113, + 505, + 125 + ], + "score": 1.0, + "content": "Interleaved Image-Text Data We collect interleaved multimodal data from the Common Crawl", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 125, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 506, + 137 + ], + "score": 1.0, + "content": "snapshot, which is a publicly available archive of web pages. We use a filtering process to select", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "score": 1.0, + "content": "about 71 millions web pages from the original 2 billions web pages in the snapshot. We then extract", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 145, + 362, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 362, + 160 + ], + "score": 1.0, + "content": "the text and images from the HTML of each selected web page.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 107, + 175, + 210, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 212, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 212, + 190 + ], + "score": 1.0, + "content": "2.4 Training Objective", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 196, + 505, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 197, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 505, + 209 + ], + "score": 1.0, + "content": "The KOSMOS-1 training is conducted on web-scale multimodal corpora, including monomodal data", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "(e.g., text corpus), cross-modal paired data (e.g., image-caption pairs), and interleaved multimodal", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 232 + ], + "score": 1.0, + "content": "data (e.g., documents of arbitrarily interleaved images and texts). To be specific, we use monomodal", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "data for representation learning. For example, language modeling with text data pretrains instruction", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "score": 1.0, + "content": "following, in-context learning, and various language tasks. Moreover, cross-modal pairs and inter-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "leaved data learn to align the perception of general modalities with language models. Interleaved data", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "also naturally fit in the multimodal language modeling task. We present more details of training data", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 274, + 266, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 266, + 286 + ], + "score": 1.0, + "content": "collection in the supplemental material.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 290, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "The models are trained with the next-token prediction task, i.e., learning to generate the next token", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "depending on the previous context. The training objective is to maximize the log-likelihood of tokens", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 309, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 326 + ], + "score": 1.0, + "content": "in examples. Notice that only discrete tokens, such as text tokens, are accounted for in the training", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 321, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 336 + ], + "score": 1.0, + "content": "loss. Multimodal language modeling is a scalable way to train the models. More importantly, the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 333, + 495, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 495, + 346 + ], + "score": 1.0, + "content": "emergence of various capabilities makes the training task favorable for downstream applications.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 107, + 365, + 191, + 378 + ], + "lines": [ + { + "bbox": [ + 103, + 362, + 193, + 382 + ], + "spans": [ + { + "bbox": [ + 103, + 362, + 193, + 382 + ], + "score": 1.0, + "content": "3 Experiments", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 392, + 194, + 404 + ], + "lines": [ + { + "bbox": [ + 104, + 389, + 196, + 407 + ], + "spans": [ + { + "bbox": [ + 104, + 389, + 196, + 407 + ], + "score": 1.0, + "content": "3.1 Training Setup", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 427 + ], + "score": 1.0, + "content": "We train KOSMOS-1 with 1.6 billion parameters using a mix of text corpora, image-caption pairs,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "and interleaved data. We use Magneto’s initialization for optimization stability and a pretrained CLIP", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 375, + 449 + ], + "score": 1.0, + "content": "ViT-L/14 model for image representation. The model is trained for", + "type": "text" + }, + { + "bbox": [ + 376, + 436, + 398, + 447 + ], + "score": 0.28, + "content": "3 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "steps using a batch size of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 445, + 507, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 507, + 460 + ], + "score": 1.0, + "content": "1.2 million tokens and the AdamW optimizer. We adopt a learning rate warm-up and decay schedule,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 458, + 267, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 267, + 470 + ], + "score": 1.0, + "content": "and use SentencePiece for tokenization.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 488 + ], + "score": 1.0, + "content": "To improve instruction-following capabilities, we perform language-only instruction tuning using", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "Unnatural Instructions [24] and FLANv2 [25] datasets. This tuning process is conducted as language", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "modeling, and improvements transfer across modalities. More details about hyperparameters can be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 507, + 250, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 250, + 519 + ], + "score": 1.0, + "content": "found in the supplemental material.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 108, + 523, + 504, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "Table 1 summarizes the corresponding datasets and what capabilities we would like to evaluate. We", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 533, + 437, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 437, + 547 + ], + "score": 1.0, + "content": "evaluate different capabilities related to language, perception-language and vision.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 108, + 563, + 248, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 249, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 249, + 578 + ], + "score": 1.0, + "content": "3.2 Perception-Language Tasks", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 584, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "Image Captioning Table 2a shows the captioning performance on COCO [39] Karpathy test split", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "and Flickr30k [40] test set. KOSMOS-1 achieves remarkable results in zero-shot setting on two image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "captioning datasets. Specifically, our model achieves a CIDEr score of 67.1 on the Flickr30k dataset,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "compared to 60.6 and 61.5 for the Flamingo-3B and Flamingo-9B models, respectively. Notably, our", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "model is able to accomplish this feat with a smaller size of 1.6B, compared to Flamingo models. This", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 380, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 380, + 654 + ], + "score": 1.0, + "content": "demonstrates our model’s superiority in zero-shot image captioning.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "Visual Question Answering Table 2b reports the visual question answering results on VQAv2 [41]", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "and VizWiz [42]. We show that KOSMOS-1 can better handle the diversity and complexity of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "the VizWiz dataset. KOSMOS-1 achieves higher accuracy and robustness than Flamingo-3B and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "Flamingo-9B models on zero-shot settings. In addition, our model is competitive with Flamingo on", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 187, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 187, + 724 + ], + "score": 1.0, + "content": "the VQAv2 dataset.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + } + ], + "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, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "English LAION-2B, LAION-400M, and COYO-700M are collected from web pages of the Common", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 97 + ], + "score": 1.0, + "content": "Crawl web data by extracting image sources and the corresponding alt-text. Conceptual Captions are", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 227, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 227, + 108 + ], + "score": 1.0, + "content": "also from internet web pages.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 72, + 505, + 108 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 113, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 113, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 113, + 505, + 125 + ], + "score": 1.0, + "content": "Interleaved Image-Text Data We collect interleaved multimodal data from the Common Crawl", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 125, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 506, + 137 + ], + "score": 1.0, + "content": "snapshot, which is a publicly available archive of web pages. We use a filtering process to select", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "score": 1.0, + "content": "about 71 millions web pages from the original 2 billions web pages in the snapshot. We then extract", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 145, + 362, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 362, + 160 + ], + "score": 1.0, + "content": "the text and images from the HTML of each selected web page.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 113, + 506, + 160 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 175, + 210, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 212, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 212, + 190 + ], + "score": 1.0, + "content": "2.4 Training Objective", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 196, + 505, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 197, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 505, + 209 + ], + "score": 1.0, + "content": "The KOSMOS-1 training is conducted on web-scale multimodal corpora, including monomodal data", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "(e.g., text corpus), cross-modal paired data (e.g., image-caption pairs), and interleaved multimodal", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 232 + ], + "score": 1.0, + "content": "data (e.g., documents of arbitrarily interleaved images and texts). To be specific, we use monomodal", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "data for representation learning. For example, language modeling with text data pretrains instruction", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "score": 1.0, + "content": "following, in-context learning, and various language tasks. Moreover, cross-modal pairs and inter-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "leaved data learn to align the perception of general modalities with language models. Interleaved data", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "also naturally fit in the multimodal language modeling task. We present more details of training data", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 274, + 266, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 266, + 286 + ], + "score": 1.0, + "content": "collection in the supplemental material.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 197, + 506, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 290, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "The models are trained with the next-token prediction task, i.e., learning to generate the next token", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "depending on the previous context. The training objective is to maximize the log-likelihood of tokens", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 309, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 326 + ], + "score": 1.0, + "content": "in examples. Notice that only discrete tokens, such as text tokens, are accounted for in the training", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 321, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 336 + ], + "score": 1.0, + "content": "loss. Multimodal language modeling is a scalable way to train the models. More importantly, the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 333, + 495, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 495, + 346 + ], + "score": 1.0, + "content": "emergence of various capabilities makes the training task favorable for downstream applications.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 289, + 505, + 346 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 365, + 191, + 378 + ], + "lines": [ + { + "bbox": [ + 103, + 362, + 193, + 382 + ], + "spans": [ + { + "bbox": [ + 103, + 362, + 193, + 382 + ], + "score": 1.0, + "content": "3 Experiments", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 392, + 194, + 404 + ], + "lines": [ + { + "bbox": [ + 104, + 389, + 196, + 407 + ], + "spans": [ + { + "bbox": [ + 104, + 389, + 196, + 407 + ], + "score": 1.0, + "content": "3.1 Training Setup", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 427 + ], + "score": 1.0, + "content": "We train KOSMOS-1 with 1.6 billion parameters using a mix of text corpora, image-caption pairs,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "and interleaved data. We use Magneto’s initialization for optimization stability and a pretrained CLIP", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 375, + 449 + ], + "score": 1.0, + "content": "ViT-L/14 model for image representation. The model is trained for", + "type": "text" + }, + { + "bbox": [ + 376, + 436, + 398, + 447 + ], + "score": 0.28, + "content": "3 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "steps using a batch size of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 445, + 507, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 507, + 460 + ], + "score": 1.0, + "content": "1.2 million tokens and the AdamW optimizer. We adopt a learning rate warm-up and decay schedule,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 458, + 267, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 267, + 470 + ], + "score": 1.0, + "content": "and use SentencePiece for tokenization.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 413, + 507, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 488 + ], + "score": 1.0, + "content": "To improve instruction-following capabilities, we perform language-only instruction tuning using", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "Unnatural Instructions [24] and FLANv2 [25] datasets. This tuning process is conducted as language", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "modeling, and improvements transfer across modalities. More details about hyperparameters can be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 507, + 250, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 250, + 519 + ], + "score": 1.0, + "content": "found in the supplemental material.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 473, + 505, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 523, + 504, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "Table 1 summarizes the corresponding datasets and what capabilities we would like to evaluate. We", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 533, + 437, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 437, + 547 + ], + "score": 1.0, + "content": "evaluate different capabilities related to language, perception-language and vision.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 523, + 505, + 547 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 563, + 248, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 249, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 249, + 578 + ], + "score": 1.0, + "content": "3.2 Perception-Language Tasks", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 584, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "Image Captioning Table 2a shows the captioning performance on COCO [39] Karpathy test split", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "and Flickr30k [40] test set. KOSMOS-1 achieves remarkable results in zero-shot setting on two image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "captioning datasets. Specifically, our model achieves a CIDEr score of 67.1 on the Flickr30k dataset,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "compared to 60.6 and 61.5 for the Flamingo-3B and Flamingo-9B models, respectively. Notably, our", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "model is able to accomplish this feat with a smaller size of 1.6B, compared to Flamingo models. This", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 380, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 380, + 654 + ], + "score": 1.0, + "content": "demonstrates our model’s superiority in zero-shot image captioning.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 585, + 506, + 654 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "Visual Question Answering Table 2b reports the visual question answering results on VQAv2 [41]", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "and VizWiz [42]. We show that KOSMOS-1 can better handle the diversity and complexity of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "the VizWiz dataset. KOSMOS-1 achieves higher accuracy and robustness than Flamingo-3B and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "Flamingo-9B models on zero-shot settings. In addition, our model is competitive with Flamingo on", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 187, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 187, + 724 + ], + "score": 1.0, + "content": "the VQAv2 dataset.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 666, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 70, + 493, + 333 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 117, + 70, + 493, + 333 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 70, + 493, + 333 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 493, + 333 + ], + "score": 0.984, + "html": "
DatasetTask descriptionMetricZero-shotFew-shot
Language tasks
StoryCloze [26]Commonsense reasoningAccuracy
HellaSwag [27]Commonsense NLIAccuracy
Winograd [28]Word ambiguityAccuracy
Winogrande [29]Word ambiguityAccuracy
PIQA [30]Physical commonsenseAccuracy
BoolQ[31]Question answeringAccuracy
CB [32]Textual entailmentAccuracyvvνvvvvv
COPA [33]Causal reasoningAccuracy
Rendered SST-2 [34]OCR-free sentiment classificationAccuracy
HatefulMemes [35]OCR-free meme classificationROC AUC
Cross-modal transfer
RelativeSize [36]Commonsense reasoning (object size)Accuracy√
MemoryColor [37]Commonsense reasoning (object color)Accuracy√
ColorTerms [38]Commonsense reasoning (object color)Accuracy√
Nonverbal reasoning tasks
IQ TestRaven's Progressive MatricesAccuracy√
Perception-language tasks
COCO Caption [39]Image captioningCIDEr, etc.
Flicker30k [40]Image captioningCIDEr, etc.
VQAv2 [41]Visual question answeringVQA acc.<<vv
VizWiz[42]Visual question answeringVQA acc.
WebSRC[43]Web page question answeringF1 score
Vision tasks
CUB [44]Zero-shot image classification with descriptionsAccuracy
", + "type": "table", + "image_path": "fb05bd9e6b45f4f4a15930e2ccc1d1343a5cc9faa880f00294b1ad19b35bb3d1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 70, + 493, + 157.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 157.66666666666669, + 493, + 245.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 245.33333333333337, + 493, + 333.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 108, + 334, + 502, + 355 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 331, + 504, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 504, + 346 + ], + "score": 1.0, + "content": "Table 1: We evaluate the capabilities of KOSMOS-1 on language, perception-language, and vision", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 343, + 321, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 321, + 357 + ], + "score": 1.0, + "content": "tasks under both zero- and few-shot learning settings.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 107, + 383, + 260, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 261, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 261, + 398 + ], + "score": 1.0, + "content": "3.3 IQ Test: Nonverbal Reasoning", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 404, + 505, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "Raven’s Progressive Matrices [9, 8] is one of the most common tests to evaluate nonverbal reasoning.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "The capability of nonverbal reasoning is typically a reflection of an individual’s intelligence quotientWhich option can complete the matrix?", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "(IQ). Figure 2 shows an example. Given eight images, the task is to identify the following elementA B C D E F", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 436, + 219, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 219, + 449 + ], + "score": 1.0, + "content": "from six similar candidates.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "image", + "bbox": [ + 108, + 458, + 331, + 563 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 458, + 331, + 563 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 458, + 331, + 563 + ], + "spans": [ + { + "bbox": [ + 108, + 458, + 331, + 563 + ], + "score": 0.969, + "type": "image", + "image_path": "c9e724bf92595d0da428f24c3231d71fc5836e0699fc5c281bc47e57ff8c2942.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 108, + 458, + 331, + 473.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 473.0, + 331, + 488.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 488.0, + 331, + 503.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 108, + 503.0, + 331, + 518.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 108, + 518.0, + 331, + 533.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 108, + 533.0, + 331, + 548.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 108, + 548.0, + 331, + 563.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 570, + 495, + 592 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 568, + 496, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 496, + 583 + ], + "score": 1.0, + "content": "Figure 2: We append each candidate image to the prompt Table 3: Zero-shot generalization on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 579, + 394, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 292, + 592 + ], + "score": 1.0, + "content": "separately and query the model if it is correct.", + "type": "text" + }, + { + "bbox": [ + 334, + 579, + 394, + 593 + ], + "score": 1.0, + "content": "Raven IQ test.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + } + ], + "index": 16.25 + }, + { + "type": "table", + "bbox": [ + 349, + 519, + 480, + 565 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 349, + 519, + 480, + 565 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 349, + 519, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 349, + 519, + 480, + 565 + ], + "score": 0.945, + "html": "
MethodAccuracy
Random Choice17%
KosMOs-122%
", + "type": "table", + "image_path": "3bbdff5b7c762d7d5c92560e40ebceb42e33279918de308941f953cf6c2dd216.jpg" + } + ] + } + ], + "index": 17.0, + "virtual_lines": [ + { + "bbox": [ + 349, + 519, + 480, + 542.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 349, + 542.0, + 480, + 565.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 17.0 + }, + { + "type": "text", + "bbox": [ + 106, + 607, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "The models need to conduct zero-shot nonverbal reasoning without explicitly fine-tuning. The Raven", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "IQ test is analogous to in-context learning of language models, where the difference is whether the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "context is nonverbal or verbal. In order to infer the answers, the models have to recognize abstract", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "score": 1.0, + "content": "concepts and identify the underlying patterns of given images. So the IQ task is a good testbed to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 650, + 331, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 331, + 664 + ], + "score": 1.0, + "content": "benchmark the nonverbal in-context learning capability.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 426, + 678 + ], + "score": 1.0, + "content": "Table 3 shows the evaluation results on the IQ test dataset. KOSMOS-1 achieves", + "type": "text" + }, + { + "bbox": [ + 426, + 667, + 449, + 678 + ], + "score": 0.86, + "content": "5 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 667, + 505, + 678 + ], + "score": 1.0, + "content": "improvement", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "respectively over the random baseline. The results indicate that KOSMOS-1 is able to perceive abstract", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "conceptual patterns in a nonverbal context, and then deduce the following element across multiple", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "choices. To the best of our knowledge, it is the first time that a model can perform such zero-shot", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "Raven IQ tests. Although there is still a large performance gap between the current model and the", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 70, + 493, + 333 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 117, + 70, + 493, + 333 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 70, + 493, + 333 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 493, + 333 + ], + "score": 0.984, + "html": "
DatasetTask descriptionMetricZero-shotFew-shot
Language tasks
StoryCloze [26]Commonsense reasoningAccuracy
HellaSwag [27]Commonsense NLIAccuracy
Winograd [28]Word ambiguityAccuracy
Winogrande [29]Word ambiguityAccuracy
PIQA [30]Physical commonsenseAccuracy
BoolQ[31]Question answeringAccuracy
CB [32]Textual entailmentAccuracyvvνvvvvv
COPA [33]Causal reasoningAccuracy
Rendered SST-2 [34]OCR-free sentiment classificationAccuracy
HatefulMemes [35]OCR-free meme classificationROC AUC
Cross-modal transfer
RelativeSize [36]Commonsense reasoning (object size)Accuracy√
MemoryColor [37]Commonsense reasoning (object color)Accuracy√
ColorTerms [38]Commonsense reasoning (object color)Accuracy√
Nonverbal reasoning tasks
IQ TestRaven's Progressive MatricesAccuracy√
Perception-language tasks
COCO Caption [39]Image captioningCIDEr, etc.
Flicker30k [40]Image captioningCIDEr, etc.
VQAv2 [41]Visual question answeringVQA acc.<<vv
VizWiz[42]Visual question answeringVQA acc.
WebSRC[43]Web page question answeringF1 score
Vision tasks
CUB [44]Zero-shot image classification with descriptionsAccuracy
", + "type": "table", + "image_path": "fb05bd9e6b45f4f4a15930e2ccc1d1343a5cc9faa880f00294b1ad19b35bb3d1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 70, + 493, + 157.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 157.66666666666669, + 493, + 245.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 245.33333333333337, + 493, + 333.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 108, + 334, + 502, + 355 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 331, + 504, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 504, + 346 + ], + "score": 1.0, + "content": "Table 1: We evaluate the capabilities of KOSMOS-1 on language, perception-language, and vision", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 343, + 321, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 321, + 357 + ], + "score": 1.0, + "content": "tasks under both zero- and few-shot learning settings.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 107, + 383, + 260, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 261, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 261, + 398 + ], + "score": 1.0, + "content": "3.3 IQ Test: Nonverbal Reasoning", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 404, + 505, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "Raven’s Progressive Matrices [9, 8] is one of the most common tests to evaluate nonverbal reasoning.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "The capability of nonverbal reasoning is typically a reflection of an individual’s intelligence quotientWhich option can complete the matrix?", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "(IQ). Figure 2 shows an example. Given eight images, the task is to identify the following elementA B C D E F", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 436, + 219, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 219, + 449 + ], + "score": 1.0, + "content": "from six similar candidates.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 403, + 506, + 449 + ] + }, + { + "type": "image", + "bbox": [ + 108, + 458, + 331, + 563 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 458, + 331, + 563 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 458, + 331, + 563 + ], + "spans": [ + { + "bbox": [ + 108, + 458, + 331, + 563 + ], + "score": 0.969, + "type": "image", + "image_path": "c9e724bf92595d0da428f24c3231d71fc5836e0699fc5c281bc47e57ff8c2942.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 108, + 458, + 331, + 473.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 473.0, + 331, + 488.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 488.0, + 331, + 503.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 108, + 503.0, + 331, + 518.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 108, + 518.0, + 331, + 533.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 108, + 533.0, + 331, + 548.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 108, + 548.0, + 331, + 563.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 570, + 495, + 592 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 568, + 496, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 496, + 583 + ], + "score": 1.0, + "content": "Figure 2: We append each candidate image to the prompt Table 3: Zero-shot generalization on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 579, + 394, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 292, + 592 + ], + "score": 1.0, + "content": "separately and query the model if it is correct.", + "type": "text" + }, + { + "bbox": [ + 334, + 579, + 394, + 593 + ], + "score": 1.0, + "content": "Raven IQ test.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + } + ], + "index": 16.25 + }, + { + "type": "table", + "bbox": [ + 349, + 519, + 480, + 565 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 349, + 519, + 480, + 565 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 349, + 519, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 349, + 519, + 480, + 565 + ], + "score": 0.945, + "html": "
MethodAccuracy
Random Choice17%
KosMOs-122%
", + "type": "table", + "image_path": "3bbdff5b7c762d7d5c92560e40ebceb42e33279918de308941f953cf6c2dd216.jpg" + } + ] + } + ], + "index": 17.0, + "virtual_lines": [ + { + "bbox": [ + 349, + 519, + 480, + 542.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 349, + 542.0, + 480, + 565.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 17.0 + }, + { + "type": "text", + "bbox": [ + 106, + 607, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "The models need to conduct zero-shot nonverbal reasoning without explicitly fine-tuning. The Raven", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "IQ test is analogous to in-context learning of language models, where the difference is whether the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "context is nonverbal or verbal. In order to infer the answers, the models have to recognize abstract", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "score": 1.0, + "content": "concepts and identify the underlying patterns of given images. So the IQ task is a good testbed to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 650, + 331, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 331, + 664 + ], + "score": 1.0, + "content": "benchmark the nonverbal in-context learning capability.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 606, + 506, + 664 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 426, + 678 + ], + "score": 1.0, + "content": "Table 3 shows the evaluation results on the IQ test dataset. KOSMOS-1 achieves", + "type": "text" + }, + { + "bbox": [ + 426, + 667, + 449, + 678 + ], + "score": 0.86, + "content": "5 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 667, + 505, + 678 + ], + "score": 1.0, + "content": "improvement", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "respectively over the random baseline. The results indicate that KOSMOS-1 is able to perceive abstract", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "conceptual patterns in a nonverbal context, and then deduce the following element across multiple", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "choices. To the best of our knowledge, it is the first time that a model can perform such zero-shot", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "Raven IQ tests. Although there is still a large performance gap between the current model and the", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 667, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 70, + 297, + 255 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 70, + 297, + 255 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 70, + 297, + 255 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 297, + 255 + ], + "score": 0.976, + "html": "
ShotModelCoCoFlickr30k
0ZeroCap45]14.61
VLKD [46]58.31
FewVLM[47]-31.0
METALM[3]82.243.4
Flamingo-3B*[5]73.060.6
Flamingo-9B*[5]79.461.5
KOSMOS-1 (1.6B)84.767.1
2Flamingo-3B* [5]=1
Flamingo-9B*[5] KOSMOS-1 (1.6B)- 99.6-
4Flamingo-3B* [5]85.070.0 72.0
Flamingo-9B* [5]93.172.6
KoSMOS-1 (1.6B)101.775.3
Flamingo-3B* [5]90.671.7
8Flamingo-9B* [5]99.073.4
KoSMOS-1 (1.6B)96.768.0
", + "type": "table", + "image_path": "dae83ae565f65ae9db8b83d404b73e1dfef200595bc31335cc307359565ec8ec.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 108, + 70, + 297, + 83.21428571428571 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 83.21428571428571, + 297, + 96.42857142857142 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 96.42857142857142, + 297, + 109.64285714285712 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 109.64285714285712, + 297, + 122.85714285714283 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 122.85714285714283, + 297, + 136.07142857142856 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 136.07142857142856, + 297, + 149.28571428571428 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 149.28571428571428, + 297, + 162.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 162.5, + 297, + 175.71428571428572 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 175.71428571428572, + 297, + 188.92857142857144 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 108, + 188.92857142857144, + 297, + 202.14285714285717 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 108, + 202.14285714285717, + 297, + 215.3571428571429 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 215.3571428571429, + 297, + 228.5714285714286 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 228.5714285714286, + 297, + 241.78571428571433 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 108, + 241.78571428571433, + 297, + 255.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 102, + 258, + 503, + 279 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 258, + 499, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 499, + 269 + ], + "score": 1.0, + "content": "(a) Image captioning results on COCO caption Karpa- (b) Visual question answering results on VQAv2 and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 268, + 462, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 462, + 280 + ], + "score": 1.0, + "content": "thy test and Flickr30k test. We present CIDEr scores. VizWiz. We present VQA accuracy scores.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + } + ], + "index": 17.5 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 495, + 254 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 307, + 71, + 495, + 254 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 307, + 71, + 495, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 495, + 254 + ], + "score": 0.977, + "html": "
ShotModelVQAv2VizWiz
0Frozen29.5
VLKDViT-B/1638.6=
METALM41.1-
Flamingo-3B*49.228.9
Flamingo-9B*51.828.8
KoSMOS-1 (1.6B)51.029.2
2Flamingo-3B* [5]-
Flamingo-9B*[5] KoSMOS-1 (1.6B)-=
451.431.4
Flamingo-3B* [5] Flamingo-9B* [5]53.2 56.334.4 34.9
KoSMOS-1 (1.6B)51.835.3
8Flamingo-3B* [5]55.438.4
Flamingo-9B* [5]58.039.4
KOSMOS-1 (1.6B)51.439.0
", + "type": "table", + "image_path": "6372deadd0b75f30e1a9a4a4f14d65a287471427b6f441afb31badbc5d9c181d.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 307, + 71, + 495, + 84.07142857142857 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 307, + 84.07142857142857, + 495, + 97.14285714285714 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 307, + 97.14285714285714, + 495, + 110.21428571428571 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 307, + 110.21428571428571, + 495, + 123.28571428571428 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 307, + 123.28571428571428, + 495, + 136.35714285714286 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 307, + 136.35714285714286, + 495, + 149.42857142857144 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 307, + 149.42857142857144, + 495, + 162.50000000000003 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 307, + 162.50000000000003, + 495, + 175.5714285714286 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 307, + 175.5714285714286, + 495, + 188.6428571428572 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 307, + 188.6428571428572, + 495, + 201.71428571428578 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 307, + 201.71428571428578, + 495, + 214.78571428571436 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 307, + 214.78571428571436, + 495, + 227.85714285714295 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 307, + 227.85714285714295, + 495, + 240.92857142857153 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 307, + 240.92857142857153, + 495, + 254.0000000000001 + ], + "spans": [], + "index": 27 + } + ] + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 284, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 143, + 298 + ], + "score": 1.0, + "content": "Table 2:", + "type": "text" + }, + { + "bbox": [ + 144, + 285, + 157, + 294 + ], + "score": 0.43, + "content": "\" * \"", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 283, + 505, + 298 + ], + "score": 1.0, + "content": ": Flamingo [5] builds the zero-shot prompt with two examples from the downstream", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "tasks where their corresponding images are removed (i.e., similar to few-shot text prompts) while the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 305, + 264, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 264, + 319 + ], + "score": 1.0, + "content": "others evaluate true zero-shot learning.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 108, + 347, + 504, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "average level of adults, KOSMOS-1 demonstrates the potential of MLLMs to perform zero-shot", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 358, + 374, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 374, + 371 + ], + "score": 1.0, + "content": "nonverbal reasoning by aligning perception with language models.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 108, + 383, + 286, + 395 + ], + "lines": [ + { + "bbox": [ + 104, + 380, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 104, + 380, + 288, + 399 + ], + "score": 1.0, + "content": "3.4 OCR-Free Language Understanding", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 404, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "OCR-free language understanding is a task that focuses on understanding text and images without", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 414, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 428 + ], + "score": 1.0, + "content": "relying on Optical Character Recognition (OCR). For example, during the Rendered SST-2 task [34],", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "sentences from the Stanford Sentiment Treebank [48] dataset are rendered as images. The model is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 435, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 450 + ], + "score": 1.0, + "content": "asked to predict the sentiment of the text within the images. The task evaluates a model’s ability to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 447, + 444, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 444, + 461 + ], + "score": 1.0, + "content": "read and comprehend the meaning of words and sentences directly from the images.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 530 + ], + "lines": [ + { + "bbox": [ + 105, + 463, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 344, + 477 + ], + "score": 1.0, + "content": "As shown in Table 4a, KOSMOS-1 achieves a ROC AUC of", + "type": "text" + }, + { + "bbox": [ + 345, + 464, + 372, + 475 + ], + "score": 0.88, + "content": "6 3 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 463, + 505, + 477 + ], + "score": 1.0, + "content": "for the HatefulMemes validation", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 217, + 487 + ], + "score": 1.0, + "content": "set and a test accuracy of", + "type": "text" + }, + { + "bbox": [ + 218, + 475, + 246, + 486 + ], + "score": 0.88, + "content": "6 7 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "for the Rendered SST-2 test set. It outperforms CLIP ViT-L", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 287, + 498 + ], + "score": 1.0, + "content": "and Flamingo-9B, which achieve AUCs of", + "type": "text" + }, + { + "bbox": [ + 287, + 486, + 315, + 496 + ], + "score": 0.89, + "content": "6 3 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 484, + 334, + 498 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 335, + 486, + 363, + 496 + ], + "score": 0.88, + "content": "5 7 . 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "on the HatefulMemes task. Note", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "score": 1.0, + "content": "that Flamingo explicitly provides OCR text into the prompt, while KOSMOS-1 does not access", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "any external tools or resources. This indicates that KOSMOS-1 has built-in abilities to read and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 519, + 285, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 285, + 532 + ], + "score": 1.0, + "content": "comprehend the text in the rendered images.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + }, + { + "type": "table", + "bbox": [ + 108, + 541, + 338, + 626 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 541, + 338, + 626 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 108, + 541, + 338, + 626 + ], + "spans": [ + { + "bbox": [ + 108, + 541, + 338, + 626 + ], + "score": 0.973, + "html": "
ModelHatefulMemesRendered SST-2
CLIP ViT-B/3257.659.6
CLIP ViT-B/1661.759.8
CLIP ViT-L/1463.364.0
Flamingo-3B53.71
Flamingo-9B57.0-
KOSMOS-1 (1.6B)63.967.1
", + "type": "table", + "image_path": "63fe5a166f31b0c1dea66c3dfb545a6e3e37828aee0bdb636daef6489772ba5f.jpg" + } + ] + } + ], + "index": 52.0, + "virtual_lines": [ + { + "bbox": [ + 108, + 541, + 338, + 555.1666666666666 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 108, + 555.1666666666666, + 338, + 569.3333333333333 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 108, + 569.3333333333333, + 338, + 583.4999999999999 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 108, + 583.4999999999999, + 338, + 597.6666666666665 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 108, + 597.6666666666665, + 338, + 611.8333333333331 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 108, + 611.8333333333331, + 338, + 625.9999999999998 + ], + "spans": [], + "index": 57 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 629, + 331, + 650 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 629, + 332, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 332, + 641 + ], + "score": 1.0, + "content": "(a) Zero-shot generalization on OCR-free language understand-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 640, + 222, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 222, + 651 + ], + "score": 1.0, + "content": "ing. We report accuracy scores.", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 60.0 + } + ], + "index": 56.0 + }, + { + "type": "table", + "bbox": [ + 341, + 541, + 502, + 626 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 341, + 541, + 502, + 626 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 341, + 541, + 502, + 626 + ], + "spans": [ + { + "bbox": [ + 341, + 541, + 502, + 626 + ], + "score": 0.933, + "html": "
ModelEM F1
Using extracted text
LLM 7.617.9
KosMOs-1 15.831.3
Without using extracted text
KosMos-1 3.810.6
", + "type": "table", + "image_path": "82eac6ef90a61eb6f4a12a3eae88d332d6d183bd2d540dd8edeee0a64d7b9854.jpg" + } + ] + } + ], + "index": 53.0, + "virtual_lines": [ + { + "bbox": [ + 341, + 541, + 502, + 555.1666666666666 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 341, + 555.1666666666666, + 502, + 569.3333333333333 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 341, + 569.3333333333333, + 502, + 583.4999999999999 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 341, + 583.4999999999999, + 502, + 597.6666666666665 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 341, + 597.6666666666665, + 502, + 611.8333333333331 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 341, + 611.8333333333331, + 502, + 625.9999999999998 + ], + "spans": [], + "index": 58 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 341, + 629, + 502, + 650 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 340, + 629, + 503, + 640 + ], + "spans": [ + { + "bbox": [ + 340, + 629, + 503, + 640 + ], + "score": 1.0, + "content": "(b) Zero-shot performance on WebSRC task.", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 340, + 639, + 499, + 651 + ], + "spans": [ + { + "bbox": [ + 340, + 639, + 499, + 651 + ], + "score": 1.0, + "content": "We report exact match (EM) and F1 scores.", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 61.0 + } + ], + "index": 57.0 + }, + { + "type": "title", + "bbox": [ + 107, + 679, + 262, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 263, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 263, + 694 + ], + "score": 1.0, + "content": "3.5 Web Page Question Answering", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 63 + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "Web page question answering aims at finding answers to questions from web pages. It requires", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "the model to comprehend both the semantics and the structure of texts (such as tables, lists, and", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 64.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 70, + 297, + 255 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 70, + 297, + 255 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 70, + 297, + 255 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 297, + 255 + ], + "score": 0.976, + "html": "
ShotModelCoCoFlickr30k
0ZeroCap45]14.61
VLKD [46]58.31
FewVLM[47]-31.0
METALM[3]82.243.4
Flamingo-3B*[5]73.060.6
Flamingo-9B*[5]79.461.5
KOSMOS-1 (1.6B)84.767.1
2Flamingo-3B* [5]=1
Flamingo-9B*[5] KOSMOS-1 (1.6B)- 99.6-
4Flamingo-3B* [5]85.070.0 72.0
Flamingo-9B* [5]93.172.6
KoSMOS-1 (1.6B)101.775.3
Flamingo-3B* [5]90.671.7
8Flamingo-9B* [5]99.073.4
KoSMOS-1 (1.6B)96.768.0
", + "type": "table", + "image_path": "dae83ae565f65ae9db8b83d404b73e1dfef200595bc31335cc307359565ec8ec.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 108, + 70, + 297, + 83.21428571428571 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 83.21428571428571, + 297, + 96.42857142857142 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 96.42857142857142, + 297, + 109.64285714285712 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 109.64285714285712, + 297, + 122.85714285714283 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 122.85714285714283, + 297, + 136.07142857142856 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 136.07142857142856, + 297, + 149.28571428571428 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 149.28571428571428, + 297, + 162.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 162.5, + 297, + 175.71428571428572 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 175.71428571428572, + 297, + 188.92857142857144 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 108, + 188.92857142857144, + 297, + 202.14285714285717 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 108, + 202.14285714285717, + 297, + 215.3571428571429 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 215.3571428571429, + 297, + 228.5714285714286 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 228.5714285714286, + 297, + 241.78571428571433 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 108, + 241.78571428571433, + 297, + 255.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 102, + 258, + 503, + 279 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 258, + 499, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 499, + 269 + ], + "score": 1.0, + "content": "(a) Image captioning results on COCO caption Karpa- (b) Visual question answering results on VQAv2 and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 268, + 462, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 462, + 280 + ], + "score": 1.0, + "content": "thy test and Flickr30k test. We present CIDEr scores. VizWiz. We present VQA accuracy scores.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + } + ], + "index": 17.5 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 495, + 254 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 307, + 71, + 495, + 254 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 307, + 71, + 495, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 495, + 254 + ], + "score": 0.977, + "html": "
ShotModelVQAv2VizWiz
0Frozen29.5
VLKDViT-B/1638.6=
METALM41.1-
Flamingo-3B*49.228.9
Flamingo-9B*51.828.8
KoSMOS-1 (1.6B)51.029.2
2Flamingo-3B* [5]-
Flamingo-9B*[5] KoSMOS-1 (1.6B)-=
451.431.4
Flamingo-3B* [5] Flamingo-9B* [5]53.2 56.334.4 34.9
KoSMOS-1 (1.6B)51.835.3
8Flamingo-3B* [5]55.438.4
Flamingo-9B* [5]58.039.4
KOSMOS-1 (1.6B)51.439.0
", + "type": "table", + "image_path": "6372deadd0b75f30e1a9a4a4f14d65a287471427b6f441afb31badbc5d9c181d.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 307, + 71, + 495, + 84.07142857142857 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 307, + 84.07142857142857, + 495, + 97.14285714285714 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 307, + 97.14285714285714, + 495, + 110.21428571428571 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 307, + 110.21428571428571, + 495, + 123.28571428571428 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 307, + 123.28571428571428, + 495, + 136.35714285714286 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 307, + 136.35714285714286, + 495, + 149.42857142857144 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 307, + 149.42857142857144, + 495, + 162.50000000000003 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 307, + 162.50000000000003, + 495, + 175.5714285714286 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 307, + 175.5714285714286, + 495, + 188.6428571428572 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 307, + 188.6428571428572, + 495, + 201.71428571428578 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 307, + 201.71428571428578, + 495, + 214.78571428571436 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 307, + 214.78571428571436, + 495, + 227.85714285714295 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 307, + 227.85714285714295, + 495, + 240.92857142857153 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 307, + 240.92857142857153, + 495, + 254.0000000000001 + ], + "spans": [], + "index": 27 + } + ] + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 284, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 143, + 298 + ], + "score": 1.0, + "content": "Table 2:", + "type": "text" + }, + { + "bbox": [ + 144, + 285, + 157, + 294 + ], + "score": 0.43, + "content": "\" * \"", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 283, + 505, + 298 + ], + "score": 1.0, + "content": ": Flamingo [5] builds the zero-shot prompt with two examples from the downstream", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "tasks where their corresponding images are removed (i.e., similar to few-shot text prompts) while the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 305, + 264, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 264, + 319 + ], + "score": 1.0, + "content": "others evaluate true zero-shot learning.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 283, + 505, + 319 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 347, + 504, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "average level of adults, KOSMOS-1 demonstrates the potential of MLLMs to perform zero-shot", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 358, + 374, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 374, + 371 + ], + "score": 1.0, + "content": "nonverbal reasoning by aligning perception with language models.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 347, + 505, + 371 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 383, + 286, + 395 + ], + "lines": [ + { + "bbox": [ + 104, + 380, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 104, + 380, + 288, + 399 + ], + "score": 1.0, + "content": "3.4 OCR-Free Language Understanding", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 404, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "OCR-free language understanding is a task that focuses on understanding text and images without", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 414, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 428 + ], + "score": 1.0, + "content": "relying on Optical Character Recognition (OCR). For example, during the Rendered SST-2 task [34],", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "sentences from the Stanford Sentiment Treebank [48] dataset are rendered as images. The model is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 435, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 450 + ], + "score": 1.0, + "content": "asked to predict the sentiment of the text within the images. The task evaluates a model’s ability to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 447, + 444, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 444, + 461 + ], + "score": 1.0, + "content": "read and comprehend the meaning of words and sentences directly from the images.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 403, + 506, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 530 + ], + "lines": [ + { + "bbox": [ + 105, + 463, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 344, + 477 + ], + "score": 1.0, + "content": "As shown in Table 4a, KOSMOS-1 achieves a ROC AUC of", + "type": "text" + }, + { + "bbox": [ + 345, + 464, + 372, + 475 + ], + "score": 0.88, + "content": "6 3 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 463, + 505, + 477 + ], + "score": 1.0, + "content": "for the HatefulMemes validation", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 217, + 487 + ], + "score": 1.0, + "content": "set and a test accuracy of", + "type": "text" + }, + { + "bbox": [ + 218, + 475, + 246, + 486 + ], + "score": 0.88, + "content": "6 7 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "for the Rendered SST-2 test set. It outperforms CLIP ViT-L", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 287, + 498 + ], + "score": 1.0, + "content": "and Flamingo-9B, which achieve AUCs of", + "type": "text" + }, + { + "bbox": [ + 287, + 486, + 315, + 496 + ], + "score": 0.89, + "content": "6 3 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 484, + 334, + 498 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 335, + 486, + 363, + 496 + ], + "score": 0.88, + "content": "5 7 . 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "on the HatefulMemes task. Note", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "score": 1.0, + "content": "that Flamingo explicitly provides OCR text into the prompt, while KOSMOS-1 does not access", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "any external tools or resources. This indicates that KOSMOS-1 has built-in abilities to read and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 519, + 285, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 285, + 532 + ], + "score": 1.0, + "content": "comprehend the text in the rendered images.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 463, + 506, + 532 + ] + }, + { + "type": "table", + "bbox": [ + 108, + 541, + 338, + 626 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 541, + 338, + 626 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 108, + 541, + 338, + 626 + ], + "spans": [ + { + "bbox": [ + 108, + 541, + 338, + 626 + ], + "score": 0.973, + "html": "
ModelHatefulMemesRendered SST-2
CLIP ViT-B/3257.659.6
CLIP ViT-B/1661.759.8
CLIP ViT-L/1463.364.0
Flamingo-3B53.71
Flamingo-9B57.0-
KOSMOS-1 (1.6B)63.967.1
", + "type": "table", + "image_path": "63fe5a166f31b0c1dea66c3dfb545a6e3e37828aee0bdb636daef6489772ba5f.jpg" + } + ] + } + ], + "index": 52.0, + "virtual_lines": [ + { + "bbox": [ + 108, + 541, + 338, + 555.1666666666666 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 108, + 555.1666666666666, + 338, + 569.3333333333333 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 108, + 569.3333333333333, + 338, + 583.4999999999999 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 108, + 583.4999999999999, + 338, + 597.6666666666665 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 108, + 597.6666666666665, + 338, + 611.8333333333331 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 108, + 611.8333333333331, + 338, + 625.9999999999998 + ], + "spans": [], + "index": 57 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 629, + 331, + 650 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 629, + 332, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 332, + 641 + ], + "score": 1.0, + "content": "(a) Zero-shot generalization on OCR-free language understand-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 640, + 222, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 222, + 651 + ], + "score": 1.0, + "content": "ing. We report accuracy scores.", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 60.0 + } + ], + "index": 56.0 + }, + { + "type": "table", + "bbox": [ + 341, + 541, + 502, + 626 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 341, + 541, + 502, + 626 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 341, + 541, + 502, + 626 + ], + "spans": [ + { + "bbox": [ + 341, + 541, + 502, + 626 + ], + "score": 0.933, + "html": "
ModelEM F1
Using extracted text
LLM 7.617.9
KosMOs-1 15.831.3
Without using extracted text
KosMos-1 3.810.6
", + "type": "table", + "image_path": "82eac6ef90a61eb6f4a12a3eae88d332d6d183bd2d540dd8edeee0a64d7b9854.jpg" + } + ] + } + ], + "index": 53.0, + "virtual_lines": [ + { + "bbox": [ + 341, + 541, + 502, + 555.1666666666666 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 341, + 555.1666666666666, + 502, + 569.3333333333333 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 341, + 569.3333333333333, + 502, + 583.4999999999999 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 341, + 583.4999999999999, + 502, + 597.6666666666665 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 341, + 597.6666666666665, + 502, + 611.8333333333331 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 341, + 611.8333333333331, + 502, + 625.9999999999998 + ], + "spans": [], + "index": 58 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 341, + 629, + 502, + 650 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 340, + 629, + 503, + 640 + ], + "spans": [ + { + "bbox": [ + 340, + 629, + 503, + 640 + ], + "score": 1.0, + "content": "(b) Zero-shot performance on WebSRC task.", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 340, + 639, + 499, + 651 + ], + "spans": [ + { + "bbox": [ + 340, + 639, + 499, + 651 + ], + "score": 1.0, + "content": "We report exact match (EM) and F1 scores.", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 61.0 + } + ], + "index": 57.0 + }, + { + "type": "title", + "bbox": [ + 107, + 679, + 262, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 263, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 263, + 694 + ], + "score": 1.0, + "content": "3.5 Web Page Question Answering", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 63 + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "Web page question answering aims at finding answers to questions from web pages. It requires", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "the model to comprehend both the semantics and the structure of texts (such as tables, lists, and", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 64.5, + "bbox_fs": [ + 105, + 699, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 71, + 443, + 159 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 167, + 71, + 443, + 159 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 71, + 443, + 159 + ], + "spans": [ + { + "bbox": [ + 167, + 71, + 443, + 159 + ], + "score": 0.946, + "type": "image", + "image_path": "73479714b923275a1b3dac025e6afa4299f70244a26e1a86d93243a4ac2e89a3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 167, + 71, + 443, + 100.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 167, + 100.33333333333333, + 443, + 129.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 167, + 129.66666666666666, + 443, + 159.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 166, + 495, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 164, + 497, + 180 + ], + "spans": [ + { + "bbox": [ + 113, + 164, + 497, + 180 + ], + "score": 1.0, + "content": "Figure 3: In-context verbal descriptions can help KOSMOS-1 recognize visual categories better.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 197, + 505, + 230 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 504, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 504, + 210 + ], + "score": 1.0, + "content": "HTML layout). We compare the performance on the Web-based Structural Reading Comprehension", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "(WebSRC) dataset [43]. For comparisons, we train a language model (LLM) on the same text corpora", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 218, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 294, + 232 + ], + "score": 1.0, + "content": "with the same training setup as in KOSMOS-1.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 236, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 236, + 504, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 504, + 248 + ], + "score": 1.0, + "content": "The experimental results are summarized in Table 4b. We observe that KOSMOS-1 outperforms the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 245, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 259 + ], + "score": 1.0, + "content": "LLM, indicating that KOSMOS-1 can benefit from the layout and style information of web pages in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "images. In addition, we evaluate the performance of KOSMOS-1 without the extracted text in the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 267, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 328, + 283 + ], + "score": 1.0, + "content": "prompt. It shows that extracted text has a contribution of", + "type": "text" + }, + { + "bbox": [ + 329, + 269, + 372, + 279 + ], + "score": 0.71, + "content": "+ 1 2 . 0 / 2 0 . 7", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 267, + 506, + 283 + ], + "score": 1.0, + "content": "EM/F1 to KOSMOS-1, indicating", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 279, + 420, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 420, + 292 + ], + "score": 1.0, + "content": "that the benefit from modeling images does not sacrifice its language abilities.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 107, + 303, + 309, + 315 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 311, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 311, + 319 + ], + "score": 1.0, + "content": "3.6 Multimodal Chain-of-Thought Prompting", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "Chain-of-thought prompting [10] allows large language models to generate a series of reasoning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "steps and decompose a multi-step problem into intermediate steps, which can significantly improve", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "the performance in complex tasks. Motivated by chain-of-thought prompting, we investigate a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 355, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 370 + ], + "score": 1.0, + "content": "multimodal chain-of-thought prompting using KOSMOS-1. We break down perception-language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "tasks into two steps. In the first stage, given an image, we use a prompt to guide the model to generate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "score": 1.0, + "content": "a rationale. The model is then fed the rationale and a task-aware prompt to produce the final results.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 393, + 507, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 507, + 407 + ], + "score": 1.0, + "content": "We conduct experiments to evaluate the performance of the multimodal chain-of-thought prompting.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 104, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "Table 5a shows that multimodal chain-of-thought prompting achieves a score of 72.9, which is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "5.8 points higher than the standard prompting. By generating intermediate content, the model can", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 427, + 458, + 440 + ], + "spans": [ + { + "bbox": [ + 104, + 427, + 458, + 440 + ], + "score": 1.0, + "content": "recognize the text in the images and infer the sentiment of the sentences more correctly.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "table", + "bbox": [ + 326, + 465, + 468, + 509 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 326, + 465, + 468, + 509 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 326, + 465, + 468, + 509 + ], + "spans": [ + { + "bbox": [ + 326, + 465, + 468, + 509 + ], + "score": 0.968, + "html": "
SetingsAccuracy
Without Descriptions61.7
With Descriptions90.0
", + "type": "table", + "image_path": "0a5e6385a121d94fa4b58830ea6ded21d6d22d76b5153a04fdf77d19242b5f75.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 326, + 465, + 468, + 487.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 326, + 487.0, + 468, + 509.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 317, + 513, + 478, + 533 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 316, + 512, + 478, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 478, + 524 + ], + "score": 1.0, + "content": "(b) Results of zero-shot image classification", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 522, + 452, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 452, + 533 + ], + "score": 1.0, + "content": "without and with verbal descriptions.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + } + ], + "index": 29.5 + }, + { + "type": "table", + "bbox": [ + 108, + 447, + 305, + 526 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 447, + 305, + 526 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 447, + 305, + 526 + ], + "spans": [ + { + "bbox": [ + 108, + 447, + 305, + 526 + ], + "score": 0.955, + "html": "
ModelAccuracy
CLIP ViT-B/3259.6
CLIP ViT-B/1659.8
CLIP ViT-L/1464.0
KosMOS-167.1
w/ multimodal CoT prompting72.9
", + "type": "table", + "image_path": "2a1cfae9f6e73b6588079a96a029836b04455d4352ce7b67b85df8d6ab4f355e.jpg" + } + ] + } + ], + "index": 26.0, + "virtual_lines": [ + { + "bbox": [ + 108, + 447, + 305, + 460.1666666666667 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 108, + 460.1666666666667, + 305, + 473.33333333333337 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 108, + 473.33333333333337, + 305, + 486.50000000000006 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 108, + 486.50000000000006, + 305, + 499.66666666666674 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 108, + 499.66666666666674, + 305, + 512.8333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 108, + 512.8333333333334, + 305, + 526.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 529, + 307, + 550 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 529, + 307, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 307, + 542 + ], + "score": 1.0, + "content": "(a) Multimodal chain-of-thought (CoT) prompting on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 540, + 186, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 186, + 549 + ], + "score": 1.0, + "content": "Rendered SST-2 task.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + } + ], + "index": 29.75 + }, + { + "type": "title", + "bbox": [ + 106, + 575, + 339, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 340, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 340, + 590 + ], + "score": 1.0, + "content": "3.7 Zero-Shot Image Classification with Descriptions", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "The standard approach of image classification as above is to prompt the model for the specific name", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "of the object depicted in the image. However, there are also some classification rules customized for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "different users and scenarios, such as the refined classification of complex animal subspecies. We", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 630, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 641 + ], + "score": 1.0, + "content": "can utilize natural language descriptions to guide KOSMOS-1 to distinguish images in the zero-shot", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "score": 1.0, + "content": "setting, which makes the decision process more interpretable. Following CUB [44], we construct a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "bird classification dataset that contains images and natural-language descriptions of categories. The", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 662, + 293, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 293, + 674 + ], + "score": 1.0, + "content": "evaluation procedure is illustrated in Figure 3.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "The evaluation results are shown in Table 5b. We observe that providing descriptions in context can", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "significantly improve the accuracy of image classification. The consistent improvements indicate", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "that KOSMOS-1 can perceive the intentions of instructions and well align the concepts in language", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 711, + 302, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 302, + 724 + ], + "score": 1.0, + "content": "modality with visual features in vision modality.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "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": "image", + "bbox": [ + 167, + 71, + 443, + 159 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 167, + 71, + 443, + 159 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 71, + 443, + 159 + ], + "spans": [ + { + "bbox": [ + 167, + 71, + 443, + 159 + ], + "score": 0.946, + "type": "image", + "image_path": "73479714b923275a1b3dac025e6afa4299f70244a26e1a86d93243a4ac2e89a3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 167, + 71, + 443, + 100.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 167, + 100.33333333333333, + 443, + 129.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 167, + 129.66666666666666, + 443, + 159.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 166, + 495, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 164, + 497, + 180 + ], + "spans": [ + { + "bbox": [ + 113, + 164, + 497, + 180 + ], + "score": 1.0, + "content": "Figure 3: In-context verbal descriptions can help KOSMOS-1 recognize visual categories better.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 197, + 505, + 230 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 504, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 504, + 210 + ], + "score": 1.0, + "content": "HTML layout). We compare the performance on the Web-based Structural Reading Comprehension", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "(WebSRC) dataset [43]. For comparisons, we train a language model (LLM) on the same text corpora", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 218, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 294, + 232 + ], + "score": 1.0, + "content": "with the same training setup as in KOSMOS-1.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 198, + 505, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 236, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 236, + 504, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 504, + 248 + ], + "score": 1.0, + "content": "The experimental results are summarized in Table 4b. We observe that KOSMOS-1 outperforms the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 245, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 259 + ], + "score": 1.0, + "content": "LLM, indicating that KOSMOS-1 can benefit from the layout and style information of web pages in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "images. In addition, we evaluate the performance of KOSMOS-1 without the extracted text in the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 267, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 328, + 283 + ], + "score": 1.0, + "content": "prompt. It shows that extracted text has a contribution of", + "type": "text" + }, + { + "bbox": [ + 329, + 269, + 372, + 279 + ], + "score": 0.71, + "content": "+ 1 2 . 0 / 2 0 . 7", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 267, + 506, + 283 + ], + "score": 1.0, + "content": "EM/F1 to KOSMOS-1, indicating", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 279, + 420, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 420, + 292 + ], + "score": 1.0, + "content": "that the benefit from modeling images does not sacrifice its language abilities.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 236, + 506, + 292 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 303, + 309, + 315 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 311, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 311, + 319 + ], + "score": 1.0, + "content": "3.6 Multimodal Chain-of-Thought Prompting", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "Chain-of-thought prompting [10] allows large language models to generate a series of reasoning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "steps and decompose a multi-step problem into intermediate steps, which can significantly improve", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "the performance in complex tasks. Motivated by chain-of-thought prompting, we investigate a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 355, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 370 + ], + "score": 1.0, + "content": "multimodal chain-of-thought prompting using KOSMOS-1. We break down perception-language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "tasks into two steps. In the first stage, given an image, we use a prompt to guide the model to generate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "score": 1.0, + "content": "a rationale. The model is then fed the rationale and a task-aware prompt to produce the final results.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 323, + 506, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 393, + 507, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 507, + 407 + ], + "score": 1.0, + "content": "We conduct experiments to evaluate the performance of the multimodal chain-of-thought prompting.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 104, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "Table 5a shows that multimodal chain-of-thought prompting achieves a score of 72.9, which is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "5.8 points higher than the standard prompting. By generating intermediate content, the model can", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 427, + 458, + 440 + ], + "spans": [ + { + "bbox": [ + 104, + 427, + 458, + 440 + ], + "score": 1.0, + "content": "recognize the text in the images and infer the sentiment of the sentences more correctly.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 104, + 393, + 507, + 440 + ] + }, + { + "type": "table", + "bbox": [ + 326, + 465, + 468, + 509 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 326, + 465, + 468, + 509 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 326, + 465, + 468, + 509 + ], + "spans": [ + { + "bbox": [ + 326, + 465, + 468, + 509 + ], + "score": 0.968, + "html": "
SetingsAccuracy
Without Descriptions61.7
With Descriptions90.0
", + "type": "table", + "image_path": "0a5e6385a121d94fa4b58830ea6ded21d6d22d76b5153a04fdf77d19242b5f75.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 326, + 465, + 468, + 487.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 326, + 487.0, + 468, + 509.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 317, + 513, + 478, + 533 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 316, + 512, + 478, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 478, + 524 + ], + "score": 1.0, + "content": "(b) Results of zero-shot image classification", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 522, + 452, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 452, + 533 + ], + "score": 1.0, + "content": "without and with verbal descriptions.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + } + ], + "index": 29.5 + }, + { + "type": "table", + "bbox": [ + 108, + 447, + 305, + 526 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 447, + 305, + 526 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 447, + 305, + 526 + ], + "spans": [ + { + "bbox": [ + 108, + 447, + 305, + 526 + ], + "score": 0.955, + "html": "
ModelAccuracy
CLIP ViT-B/3259.6
CLIP ViT-B/1659.8
CLIP ViT-L/1464.0
KosMOS-167.1
w/ multimodal CoT prompting72.9
", + "type": "table", + "image_path": "2a1cfae9f6e73b6588079a96a029836b04455d4352ce7b67b85df8d6ab4f355e.jpg" + } + ] + } + ], + "index": 26.0, + "virtual_lines": [ + { + "bbox": [ + 108, + 447, + 305, + 460.1666666666667 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 108, + 460.1666666666667, + 305, + 473.33333333333337 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 108, + 473.33333333333337, + 305, + 486.50000000000006 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 108, + 486.50000000000006, + 305, + 499.66666666666674 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 108, + 499.66666666666674, + 305, + 512.8333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 108, + 512.8333333333334, + 305, + 526.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 529, + 307, + 550 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 529, + 307, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 307, + 542 + ], + "score": 1.0, + "content": "(a) Multimodal chain-of-thought (CoT) prompting on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 540, + 186, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 186, + 549 + ], + "score": 1.0, + "content": "Rendered SST-2 task.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + } + ], + "index": 29.75 + }, + { + "type": "title", + "bbox": [ + 106, + 575, + 339, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 340, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 340, + 590 + ], + "score": 1.0, + "content": "3.7 Zero-Shot Image Classification with Descriptions", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "The standard approach of image classification as above is to prompt the model for the specific name", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "of the object depicted in the image. However, there are also some classification rules customized for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "different users and scenarios, such as the refined classification of complex animal subspecies. We", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 630, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 641 + ], + "score": 1.0, + "content": "can utilize natural language descriptions to guide KOSMOS-1 to distinguish images in the zero-shot", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "score": 1.0, + "content": "setting, which makes the decision process more interpretable. Following CUB [44], we construct a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "bird classification dataset that contains images and natural-language descriptions of categories. The", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 662, + 293, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 293, + 674 + ], + "score": 1.0, + "content": "evaluation procedure is illustrated in Figure 3.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 596, + 506, + 674 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "The evaluation results are shown in Table 5b. We observe that providing descriptions in context can", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "significantly improve the accuracy of image classification. The consistent improvements indicate", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "that KOSMOS-1 can perceive the intentions of instructions and well align the concepts in language", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 711, + 302, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 302, + 724 + ], + "score": 1.0, + "content": "modality with visual features in vision modality.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 677, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 199, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 200, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 200, + 86 + ], + "score": 1.0, + "content": "3.8 Language Tasks", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 93, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "score": 1.0, + "content": "The models are evaluated on the language tasks given task instructions (i.e., zero-shot) or several", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 116 + ], + "score": 1.0, + "content": "demonstration examples (i.e., few-shot). Text inputs are directly fed into the models as in vanilla", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "language models. We train a language model (LLM) baseline with the same text corpora and training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 423, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 423, + 138 + ], + "score": 1.0, + "content": "setup. We evaluate KOSMOS-1 and the LLM baseline on eight language tasks.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "Table 6 presents the in-context learning performance of language tasks. KOSMOS-1 achieves", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "comparable or even better performance in cloze completion and commonsense reasoning tasks when", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "compared to LLM. In terms of the average result across all these datasets, LLM performs better in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 428, + 187 + ], + "score": 1.0, + "content": "zero-shot and one-shot settings, whereas our model performs better in few-shot", + "type": "text" + }, + { + "bbox": [ + 428, + 175, + 454, + 185 + ], + "score": 0.86, + "content": "k = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 175, + 505, + 187 + ], + "score": 1.0, + "content": ") settings. In", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "addition, Section 3.9.2 shows that MLLMs learn better visual commonsense knowledge compared", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 196, + 156, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 156, + 208 + ], + "score": 1.0, + "content": "with LLMs.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5 + }, + { + "type": "table", + "bbox": [ + 138, + 217, + 473, + 374 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 138, + 217, + 473, + 374 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 473, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 473, + 374 + ], + "score": 0.984, + "html": "
TaskZero-shotOne-shotFew-shot (k = 4)
LLMKosMos-1LLMKoSMOS-1LLMKosMos-1
StoryCloze72.972.172.972.273.172.3
HellaSwag50.450.050.250.050.450.3
Winograd71.669.871.268.470.969.8
Winogrande56.754.856.754.557.055.7
PIQA73.272.973.072.572.672.3
BoolQ56.456.455.157.258.759.2
CB39.344.641.148.242.953.6
COPA68.063.069.064.069.064.0
Average61.160.561.260.961.862.2
", + "type": "table", + "image_path": "8221ed666c3899e6db128f908b4315996ad1bb3e3d8a5bc83172d676a4741fe6.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 138, + 217, + 473, + 269.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 138, + 269.3333333333333, + 473, + 321.66666666666663 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 138, + 321.66666666666663, + 473, + 373.99999999999994 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 108, + 375, + 504, + 407 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "score": 1.0, + "content": "Table 6: Performance comparisons of language tasks between KOSMOS-1 and LLM. We use the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "same textual data and training setup to reimplement a language model. Both models do not use", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 396, + 266, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 266, + 408 + ], + "score": 1.0, + "content": "instruction tuning for fair comparisons.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 107, + 432, + 223, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 225, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 225, + 444 + ], + "score": 1.0, + "content": "3.9 Cross-modal Transfer", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 506, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 452, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 506, + 464 + ], + "score": 1.0, + "content": "Cross-modal transferability allows a model to learn from one modality (such as text, image, audio,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 463, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 506, + 475 + ], + "score": 1.0, + "content": "etc.) and transfer the knowledge to the other modalities. This skill can enable a model to perform", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "various tasks across different modalities. In this part, we evaluate the cross-model transferability of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 486, + 250, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 250, + 496 + ], + "score": 1.0, + "content": "KOSMOS-1 on several benchmarks.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 107, + 508, + 463, + 520 + ], + "lines": [ + { + "bbox": [ + 104, + 506, + 464, + 523 + ], + "spans": [ + { + "bbox": [ + 104, + 506, + 464, + 523 + ], + "score": 1.0, + "content": "3.9.1 Transfer from Language to Multimodal: Language-Only Instruction Tuning", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 527, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "To evaluate the effect of language-only instruction tuning, we conduct an ablation study using four", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "datasets: COCO, Flickr30k, VQAv2, and VizWiz. These datasets consist of image captioning and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 550, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 561 + ], + "score": 1.0, + "content": "visual questions anwsering. The evaluation metrics are: CIDEr scores for COCO/Flickr30k and VQA", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 560, + 227, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 227, + 572 + ], + "score": 1.0, + "content": "accuracy for VQAv2/VizWiz.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 108, + 576, + 504, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "Table 7 shows the experimental results. Language-only instruction tuning boosts our model’s", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "performance by 1.9 points on Flickr30k, 4.3 points on VQAv2, and 1.3 points on VizWiz. Our experi-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "score": 1.0, + "content": "ments show that language-only instruction tuning can significantly improve the model’s instruction-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "following capabilities across modalities. The results also indicate that our model can transfer the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 621, + 375, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 375, + 633 + ], + "score": 1.0, + "content": "instruction-following capability from language to other modalities.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + }, + { + "type": "table", + "bbox": [ + 147, + 642, + 464, + 683 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 147, + 642, + 464, + 683 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 147, + 642, + 464, + 683 + ], + "spans": [ + { + "bbox": [ + 147, + 642, + 464, + 683 + ], + "score": 0.972, + "html": "
ModelCoCoFlickr30kVQAv2VizWiz
Kosmos-184.767.151.029.2
w/o language-only instruction tuning87.665.246.727.9
", + "type": "table", + "image_path": "04a64646c0696b264d1d9985741f6cdecbcc400cb47c621b3523d0acdb8721ab.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 147, + 642, + 464, + 655.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 147, + 655.6666666666666, + 464, + 669.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 147, + 669.3333333333333, + 464, + 682.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 684, + 505, + 705 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 104, + 681, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 104, + 681, + 506, + 696 + ], + "score": 1.0, + "content": "Table 7: Ablation study on language-only instruction tuning. We report CIDEr scores for COCO and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 693, + 356, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 356, + 705 + ], + "score": 1.0, + "content": "Flickr30k, and VQA accuracy scores for VQAv2 and VizWiz.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + } + ], + "index": 34.25 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 740, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 740, + 309, + 752 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 199, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 200, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 200, + 86 + ], + "score": 1.0, + "content": "3.8 Language Tasks", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 93, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "score": 1.0, + "content": "The models are evaluated on the language tasks given task instructions (i.e., zero-shot) or several", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 116 + ], + "score": 1.0, + "content": "demonstration examples (i.e., few-shot). Text inputs are directly fed into the models as in vanilla", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "language models. We train a language model (LLM) baseline with the same text corpora and training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 423, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 423, + 138 + ], + "score": 1.0, + "content": "setup. We evaluate KOSMOS-1 and the LLM baseline on eight language tasks.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 92, + 505, + 138 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "Table 6 presents the in-context learning performance of language tasks. KOSMOS-1 achieves", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "comparable or even better performance in cloze completion and commonsense reasoning tasks when", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "compared to LLM. In terms of the average result across all these datasets, LLM performs better in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 428, + 187 + ], + "score": 1.0, + "content": "zero-shot and one-shot settings, whereas our model performs better in few-shot", + "type": "text" + }, + { + "bbox": [ + 428, + 175, + 454, + 185 + ], + "score": 0.86, + "content": "k = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 175, + 505, + 187 + ], + "score": 1.0, + "content": ") settings. In", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "addition, Section 3.9.2 shows that MLLMs learn better visual commonsense knowledge compared", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 196, + 156, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 156, + 208 + ], + "score": 1.0, + "content": "with LLMs.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 142, + 505, + 208 + ] + }, + { + "type": "table", + "bbox": [ + 138, + 217, + 473, + 374 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 138, + 217, + 473, + 374 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 473, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 473, + 374 + ], + "score": 0.984, + "html": "
TaskZero-shotOne-shotFew-shot (k = 4)
LLMKosMos-1LLMKoSMOS-1LLMKosMos-1
StoryCloze72.972.172.972.273.172.3
HellaSwag50.450.050.250.050.450.3
Winograd71.669.871.268.470.969.8
Winogrande56.754.856.754.557.055.7
PIQA73.272.973.072.572.672.3
BoolQ56.456.455.157.258.759.2
CB39.344.641.148.242.953.6
COPA68.063.069.064.069.064.0
Average61.160.561.260.961.862.2
", + "type": "table", + "image_path": "8221ed666c3899e6db128f908b4315996ad1bb3e3d8a5bc83172d676a4741fe6.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 138, + 217, + 473, + 269.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 138, + 269.3333333333333, + 473, + 321.66666666666663 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 138, + 321.66666666666663, + 473, + 373.99999999999994 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 108, + 375, + 504, + 407 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "score": 1.0, + "content": "Table 6: Performance comparisons of language tasks between KOSMOS-1 and LLM. We use the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "same textual data and training setup to reimplement a language model. Both models do not use", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 396, + 266, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 266, + 408 + ], + "score": 1.0, + "content": "instruction tuning for fair comparisons.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 107, + 432, + 223, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 225, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 225, + 444 + ], + "score": 1.0, + "content": "3.9 Cross-modal Transfer", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 506, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 452, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 506, + 464 + ], + "score": 1.0, + "content": "Cross-modal transferability allows a model to learn from one modality (such as text, image, audio,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 463, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 506, + 475 + ], + "score": 1.0, + "content": "etc.) and transfer the knowledge to the other modalities. This skill can enable a model to perform", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "various tasks across different modalities. In this part, we evaluate the cross-model transferability of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 486, + 250, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 250, + 496 + ], + "score": 1.0, + "content": "KOSMOS-1 on several benchmarks.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 452, + 506, + 496 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 508, + 463, + 520 + ], + "lines": [ + { + "bbox": [ + 104, + 506, + 464, + 523 + ], + "spans": [ + { + "bbox": [ + 104, + 506, + 464, + 523 + ], + "score": 1.0, + "content": "3.9.1 Transfer from Language to Multimodal: Language-Only Instruction Tuning", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 527, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "To evaluate the effect of language-only instruction tuning, we conduct an ablation study using four", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "datasets: COCO, Flickr30k, VQAv2, and VizWiz. These datasets consist of image captioning and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 550, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 561 + ], + "score": 1.0, + "content": "visual questions anwsering. The evaluation metrics are: CIDEr scores for COCO/Flickr30k and VQA", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 560, + 227, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 227, + 572 + ], + "score": 1.0, + "content": "accuracy for VQAv2/VizWiz.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 528, + 506, + 572 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 576, + 504, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "Table 7 shows the experimental results. Language-only instruction tuning boosts our model’s", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "performance by 1.9 points on Flickr30k, 4.3 points on VQAv2, and 1.3 points on VizWiz. Our experi-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "score": 1.0, + "content": "ments show that language-only instruction tuning can significantly improve the model’s instruction-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "following capabilities across modalities. The results also indicate that our model can transfer the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 621, + 375, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 375, + 633 + ], + "score": 1.0, + "content": "instruction-following capability from language to other modalities.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 576, + 506, + 633 + ] + }, + { + "type": "table", + "bbox": [ + 147, + 642, + 464, + 683 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 147, + 642, + 464, + 683 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 147, + 642, + 464, + 683 + ], + "spans": [ + { + "bbox": [ + 147, + 642, + 464, + 683 + ], + "score": 0.972, + "html": "
ModelCoCoFlickr30kVQAv2VizWiz
Kosmos-184.767.151.029.2
w/o language-only instruction tuning87.665.246.727.9
", + "type": "table", + "image_path": "04a64646c0696b264d1d9985741f6cdecbcc400cb47c621b3523d0acdb8721ab.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 147, + 642, + 464, + 655.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 147, + 655.6666666666666, + 464, + 669.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 147, + 669.3333333333333, + 464, + 682.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 684, + 505, + 705 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 104, + 681, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 104, + 681, + 506, + 696 + ], + "score": 1.0, + "content": "Table 7: Ablation study on language-only instruction tuning. We report CIDEr scores for COCO and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 693, + 356, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 356, + 705 + ], + "score": 1.0, + "content": "Flickr30k, and VQA accuracy scores for VQAv2 and VizWiz.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + } + ], + "index": 34.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 72, + 450, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 452, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 452, + 87 + ], + "score": 1.0, + "content": "3.9.2 Transfer from Multimodal to Language: Visual Commonsense Reasoning", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 93, + 506, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "Visual commonsense reasoning tasks require an understanding of the properties of everyday objects", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "in the real world, such as color, size, and shape. These tasks are challenging for language mod-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "els because they may require more information about object properties than what is available in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 138 + ], + "score": 1.0, + "content": "texts. To investigate the visual commonsense capabilities, we compare the zero-shot performance", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "of KOSMOS-1 and LLM on three object commonsense reasoning datasets, RELATIVESIZE [36],", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "score": 1.0, + "content": "MEMORYCOLOR [37] and COLORTERMS [38] datasets. RELATIVESIZE contains 486 object pairs", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "from 41 physical objects. The model is required to predict the size relation between two objects in a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 506, + 183 + ], + "score": 1.0, + "content": "binary question-answering format with “Yes”/“No” answers. MEMORYCOLOR and COLORTERMS", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 505, + 193 + ], + "score": 1.0, + "content": "require the model to predict the color of objects from a set of 11 color labels in a multiple-choice", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 138, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 138, + 204 + ], + "score": 1.0, + "content": "format.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 208, + 506, + 296 + ], + "lines": [ + { + "bbox": [ + 104, + 207, + 507, + 221 + ], + "spans": [ + { + "bbox": [ + 104, + 207, + 507, + 221 + ], + "score": 1.0, + "content": "Table 8 presents the zero-shot performance of KOSMOS-1 and LLM on visual commonsense rea-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 357, + 231 + ], + "score": 1.0, + "content": "soning tasks. KOSMOS-1 significantly outperforms LLM by", + "type": "text" + }, + { + "bbox": [ + 357, + 219, + 380, + 230 + ], + "score": 0.85, + "content": "1 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 219, + 464, + 231 + ], + "score": 1.0, + "content": "on RELATIVESIZE,", + "type": "text" + }, + { + "bbox": [ + 464, + 219, + 491, + 230 + ], + "score": 0.87, + "content": "1 4 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 219, + 506, + 231 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 506, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 200, + 242 + ], + "score": 1.0, + "content": "MEMORYCOLOR, and", + "type": "text" + }, + { + "bbox": [ + 200, + 230, + 222, + 241 + ], + "score": 0.85, + "content": "9 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 230, + 506, + 242 + ], + "score": 1.0, + "content": "on COLORTERMS dataset. The consistent improvements indicate that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "KOSMOS-1 benefits from the visual knowledge to complete the corresponding visual commonsense", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "reasoning. The reason for KOSMOS-1’s superior performance is that it has modality transferability,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 263, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 506, + 275 + ], + "score": 1.0, + "content": "which enables the model to transfer visual knowledge to language tasks. On the contrary, LLM has", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "to rely on textual knowledge and clues to answer visual commonsense questions, which limits its", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 284, + 270, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 270, + 297 + ], + "score": 1.0, + "content": "ability to reason about object properties.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "table", + "bbox": [ + 176, + 311, + 434, + 396 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 176, + 311, + 434, + 396 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 311, + 434, + 396 + ], + "spans": [ + { + "bbox": [ + 176, + 311, + 434, + 396 + ], + "score": 0.977, + "html": "
ModelSize Reasoning RELATIVESIZEColor Reasoning MEMORYCOLORCOLORTERMS
Using retrieved images VALM [49]85.058.652.7
Language-only zero-shot evaluation
LLM92.761.463.4
KosMos-194.276.173.1
", + "type": "table", + "image_path": "db4b9aa66963f69cde4a040bfc352e3d7f50c359cfb7de951bbffcf6db76deea.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 176, + 311, + 434, + 339.3333333333333 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 176, + 339.3333333333333, + 434, + 367.66666666666663 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 176, + 367.66666666666663, + 434, + 395.99999999999994 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 397, + 505, + 418 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "Table 8: Zero-shot visual commonsense reasoning on RELATIVESIZE, MEMORYCOLOR, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 406, + 327, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 327, + 419 + ], + "score": 1.0, + "content": "COLORTERMS datasets. Accuracy scores are reported.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + } + ], + "index": 21.25 + }, + { + "type": "title", + "bbox": [ + 107, + 458, + 197, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 198, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 198, + 473 + ], + "score": 1.0, + "content": "4 Related Work", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "In recent years, vision-language learning and representation models has garnered significant at-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "tention [2, 3, 4, 34, 50, 51, 52, 53, 54]. Previous vision-language models still exhibit limitations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "in instruction following, in-context abilities, and generalization capabilities for unseen tasks. Re-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "searchers have begun exploring more powerful multimodal large language models. Flamingo [5]", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "trained its model from scratch and made it possible to generate text tokens conditioned on both visual", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "and text inputs. Another category of research focus on learning multimodality abilities based on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "score": 1.0, + "content": "LLMs [7, 55, 56]. Meanwhile, some work [57, 58, 59] introduce visual instruction tuning to enhance", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 564, + 242, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 242, + 574 + ], + "score": 1.0, + "content": "instruction following capabilities.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5 + }, + { + "type": "title", + "bbox": [ + 107, + 595, + 183, + 609 + ], + "lines": [ + { + "bbox": [ + 104, + 592, + 185, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 185, + 612 + ], + "score": 1.0, + "content": "5 Conclusion", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "In this work, we introduce KOSMOS-1, a multimodal large language model that can perceive general", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "modalities, follow instructions, and perform in-context learning. The models trained on web-scale", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "multimodal corpora achieve promising results across a wide range of language tasks and multimodal", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "score": 1.0, + "content": "tasks. We show that going from LLMs to MLLMs enables new capabilities and opportunities. In the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "future, we would like to scale up KOSMOS-1 in terms of model size [13, 14, 60], and integrate the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "speech [12] capability into KOSMOS-1. In addition, KOSMOS-1 can be used as a unified interface", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "for multimodal learning, e.g., enabling using instructions and examples to control text-to-image", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "generation. We further discuss the limitations and broader societal impacts of KOSMOS-1 in the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 711, + 200, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 200, + 723 + ], + "score": 1.0, + "content": "supplemental material.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 72, + 450, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 452, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 452, + 87 + ], + "score": 1.0, + "content": "3.9.2 Transfer from Multimodal to Language: Visual Commonsense Reasoning", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 93, + 506, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "Visual commonsense reasoning tasks require an understanding of the properties of everyday objects", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "in the real world, such as color, size, and shape. These tasks are challenging for language mod-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "els because they may require more information about object properties than what is available in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 138 + ], + "score": 1.0, + "content": "texts. To investigate the visual commonsense capabilities, we compare the zero-shot performance", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "of KOSMOS-1 and LLM on three object commonsense reasoning datasets, RELATIVESIZE [36],", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "score": 1.0, + "content": "MEMORYCOLOR [37] and COLORTERMS [38] datasets. RELATIVESIZE contains 486 object pairs", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "from 41 physical objects. The model is required to predict the size relation between two objects in a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 506, + 183 + ], + "score": 1.0, + "content": "binary question-answering format with “Yes”/“No” answers. MEMORYCOLOR and COLORTERMS", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 505, + 193 + ], + "score": 1.0, + "content": "require the model to predict the color of objects from a set of 11 color labels in a multiple-choice", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 138, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 138, + 204 + ], + "score": 1.0, + "content": "format.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 93, + 506, + 204 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 208, + 506, + 296 + ], + "lines": [ + { + "bbox": [ + 104, + 207, + 507, + 221 + ], + "spans": [ + { + "bbox": [ + 104, + 207, + 507, + 221 + ], + "score": 1.0, + "content": "Table 8 presents the zero-shot performance of KOSMOS-1 and LLM on visual commonsense rea-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 357, + 231 + ], + "score": 1.0, + "content": "soning tasks. KOSMOS-1 significantly outperforms LLM by", + "type": "text" + }, + { + "bbox": [ + 357, + 219, + 380, + 230 + ], + "score": 0.85, + "content": "1 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 219, + 464, + 231 + ], + "score": 1.0, + "content": "on RELATIVESIZE,", + "type": "text" + }, + { + "bbox": [ + 464, + 219, + 491, + 230 + ], + "score": 0.87, + "content": "1 4 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 219, + 506, + 231 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 506, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 200, + 242 + ], + "score": 1.0, + "content": "MEMORYCOLOR, and", + "type": "text" + }, + { + "bbox": [ + 200, + 230, + 222, + 241 + ], + "score": 0.85, + "content": "9 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 230, + 506, + 242 + ], + "score": 1.0, + "content": "on COLORTERMS dataset. The consistent improvements indicate that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "KOSMOS-1 benefits from the visual knowledge to complete the corresponding visual commonsense", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "reasoning. The reason for KOSMOS-1’s superior performance is that it has modality transferability,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 263, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 506, + 275 + ], + "score": 1.0, + "content": "which enables the model to transfer visual knowledge to language tasks. On the contrary, LLM has", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "to rely on textual knowledge and clues to answer visual commonsense questions, which limits its", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 284, + 270, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 270, + 297 + ], + "score": 1.0, + "content": "ability to reason about object properties.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 207, + 507, + 297 + ] + }, + { + "type": "table", + "bbox": [ + 176, + 311, + 434, + 396 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 176, + 311, + 434, + 396 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 311, + 434, + 396 + ], + "spans": [ + { + "bbox": [ + 176, + 311, + 434, + 396 + ], + "score": 0.977, + "html": "
ModelSize Reasoning RELATIVESIZEColor Reasoning MEMORYCOLORCOLORTERMS
Using retrieved images VALM [49]85.058.652.7
Language-only zero-shot evaluation
LLM92.761.463.4
KosMos-194.276.173.1
", + "type": "table", + "image_path": "db4b9aa66963f69cde4a040bfc352e3d7f50c359cfb7de951bbffcf6db76deea.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 176, + 311, + 434, + 339.3333333333333 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 176, + 339.3333333333333, + 434, + 367.66666666666663 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 176, + 367.66666666666663, + 434, + 395.99999999999994 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 397, + 505, + 418 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "Table 8: Zero-shot visual commonsense reasoning on RELATIVESIZE, MEMORYCOLOR, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 406, + 327, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 327, + 419 + ], + "score": 1.0, + "content": "COLORTERMS datasets. Accuracy scores are reported.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + } + ], + "index": 21.25 + }, + { + "type": "title", + "bbox": [ + 107, + 458, + 197, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 198, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 198, + 473 + ], + "score": 1.0, + "content": "4 Related Work", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "In recent years, vision-language learning and representation models has garnered significant at-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "tention [2, 3, 4, 34, 50, 51, 52, 53, 54]. Previous vision-language models still exhibit limitations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "in instruction following, in-context abilities, and generalization capabilities for unseen tasks. Re-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "searchers have begun exploring more powerful multimodal large language models. Flamingo [5]", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "trained its model from scratch and made it possible to generate text tokens conditioned on both visual", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "and text inputs. Another category of research focus on learning multimodality abilities based on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "score": 1.0, + "content": "LLMs [7, 55, 56]. Meanwhile, some work [57, 58, 59] introduce visual instruction tuning to enhance", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 564, + 242, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 242, + 574 + ], + "score": 1.0, + "content": "instruction following capabilities.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 487, + 506, + 574 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 595, + 183, + 609 + ], + "lines": [ + { + "bbox": [ + 104, + 592, + 185, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 185, + 612 + ], + "score": 1.0, + "content": "5 Conclusion", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "In this work, we introduce KOSMOS-1, a multimodal large language model that can perceive general", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "modalities, follow instructions, and perform in-context learning. The models trained on web-scale", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "multimodal corpora achieve promising results across a wide range of language tasks and multimodal", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "score": 1.0, + "content": "tasks. We show that going from LLMs to MLLMs enables new capabilities and opportunities. In the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "future, we would like to scale up KOSMOS-1 in terms of model size [13, 14, 60], and integrate the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "speech [12] capability into KOSMOS-1. In addition, KOSMOS-1 can be used as a unified interface", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "for multimodal learning, e.g., enabling using instructions and examples to control text-to-image", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "generation. We further discuss the limitations and broader societal impacts of KOSMOS-1 in the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 711, + 200, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 200, + 723 + ], + "score": 1.0, + "content": "supplemental material.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 624, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 163, + 84 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 111, + 91, + 506, + 168 + ], + "lines": [ + { + "bbox": [ + 110, + 90, + 506, + 103 + ], + "spans": [ + { + "bbox": [ + 110, + 90, + 506, + 103 + ], + "score": 1.0, + "content": "[1] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 128, + 102, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 128, + 102, + 505, + 113 + ], + "score": 1.0, + "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 111, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 126, + 111, + 506, + 126 + ], + "score": 1.0, + "content": "Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 124, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 127, + 124, + 506, + 135 + ], + "score": 1.0, + "content": "Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 133, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 127, + 133, + 506, + 146 + ], + "score": 1.0, + "content": "Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 127, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 128, + 156, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 128, + 156, + 506, + 168 + ], + "score": 1.0, + "content": "Information Processing Systems, volume 33, pages 1877–1901. Curran Associates, Inc., 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 109, + 175, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 110, + 174, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 110, + 174, + 506, + 187 + ], + "score": 1.0, + "content": "[2] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 185, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 127, + 185, + 505, + 200 + ], + "score": 1.0, + "content": "Multimodal few-shot learning with frozen language models. In Neural Information Processing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 197, + 189, + 210 + ], + "spans": [ + { + "bbox": [ + 127, + 197, + 189, + 210 + ], + "score": 1.0, + "content": "Systems, 2021.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 109, + 216, + 505, + 240 + ], + "lines": [ + { + "bbox": [ + 110, + 216, + 507, + 230 + ], + "spans": [ + { + "bbox": [ + 110, + 216, + 507, + 230 + ], + "score": 1.0, + "content": "[3] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 227, + 507, + 240 + ], + "spans": [ + { + "bbox": [ + 127, + 227, + 507, + 240 + ], + "score": 1.0, + "content": "and Furu Wei. Language models are general-purpose interfaces. ArXiv, abs/2206.06336, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 109, + 246, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 111, + 246, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 111, + 246, + 506, + 260 + ], + "score": 1.0, + "content": "[4] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 127, + 257, + 506, + 271 + ], + "score": 1.0, + "content": "Owais Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language:", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 268, + 483, + 282 + ], + "spans": [ + { + "bbox": [ + 127, + 268, + 483, + 282 + ], + "score": 1.0, + "content": "BEiT pretraining for all vision and vision-language tasks. ArXiv, abs/2208.10442, 2022.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 111, + 288, + 506, + 366 + ], + "lines": [ + { + "bbox": [ + 110, + 287, + 507, + 302 + ], + "spans": [ + { + "bbox": [ + 110, + 287, + 507, + 302 + ], + "score": 1.0, + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 127, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 309, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 126, + 309, + 506, + 323 + ], + "score": 1.0, + "content": "Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 322, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 127, + 322, + 506, + 333 + ], + "score": 1.0, + "content": "Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 126, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 126, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "Flamingo: a visual language model for few-shot learning. In Advances in Neural Information", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 354, + 489, + 366 + ], + "spans": [ + { + "bbox": [ + 126, + 354, + 489, + 366 + ], + "score": 1.0, + "content": "Processing Systems, 2022. URL https://openreview.net/forum?id=EbMuimAbPbs.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 109, + 373, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 111, + 373, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 111, + 373, + 506, + 386 + ], + "score": 1.0, + "content": "[6] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 384, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 384, + 506, + 396 + ], + "score": 1.0, + "content": "Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer. CM3: A", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 128, + 396, + 448, + 406 + ], + "spans": [ + { + "bbox": [ + 128, + 396, + 448, + 406 + ], + "score": 1.0, + "content": "causal masked multimodal model of the Internet. ArXiv, abs/2201.07520, 2022.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 109, + 414, + 505, + 447 + ], + "lines": [ + { + "bbox": [ + 110, + 413, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 110, + 413, + 506, + 429 + ], + "score": 1.0, + "content": "[7] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 426, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 127, + 426, + 506, + 437 + ], + "score": 1.0, + "content": "pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 434, + 154, + 449 + ], + "spans": [ + { + "bbox": [ + 126, + 434, + 154, + 449 + ], + "score": 1.0, + "content": "2023.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 110, + 455, + 504, + 489 + ], + "lines": [ + { + "bbox": [ + 110, + 455, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 110, + 455, + 506, + 469 + ], + "score": 1.0, + "content": "[8] John and Jean Raven. Raven Progressive Matrices, pages 223–237. Springer US, Boston,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 465, + 507, + 480 + ], + "spans": [ + { + "bbox": [ + 126, + 465, + 507, + 480 + ], + "score": 1.0, + "content": "MA, 2003. ISBN 978-1-4615-0153-4. doi: 10.1007/978-1-4615-0153-4_11. URL https:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 477, + 333, + 490 + ], + "spans": [ + { + "bbox": [ + 127, + 477, + 333, + 490 + ], + "score": 1.0, + "content": "//doi.org/10.1007/978-1-4615-0153-4_11.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 110, + 497, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 110, + 496, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 110, + 496, + 507, + 511 + ], + "score": 1.0, + "content": "[9] Patricia A Carpenter, Marcel A Just, and Peter Shell. What one intelligence test measures:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 127, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "a theoretical account of the processing in the raven progressive matrices test. Psychological", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 519, + 228, + 531 + ], + "spans": [ + { + "bbox": [ + 127, + 519, + 228, + 531 + ], + "score": 1.0, + "content": "review, 97(3):404, 1990.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 538, + 504, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "score": 1.0, + "content": "[10] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 127, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 560, + 230, + 571 + ], + "spans": [ + { + "bbox": [ + 127, + 560, + 230, + 571 + ], + "score": 1.0, + "content": "arXiv:2201.11903, 2022.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 109, + 579, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 108, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 108, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 590, + 507, + 604 + ], + "spans": [ + { + "bbox": [ + 127, + 590, + 507, + 604 + ], + "score": 1.0, + "content": "Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 600, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 127, + 600, + 507, + 615 + ], + "score": 1.0, + "content": "Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek B Rao, Parker Barnes, Yi Tay,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 611, + 507, + 625 + ], + "spans": [ + { + "bbox": [ + 127, + 611, + 507, + 625 + ], + "score": 1.0, + "content": "Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 623, + 507, + 636 + ], + "spans": [ + { + "bbox": [ + 127, + 623, + 507, + 636 + ], + "score": 1.0, + "content": "Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 634, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 127, + 634, + 506, + 647 + ], + "score": 1.0, + "content": "Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 644, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 127, + 644, + 505, + 658 + ], + "score": 1.0, + "content": "García, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 127, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 127, + 668, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 127, + 668, + 506, + 679 + ], + "score": 1.0, + "content": "Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 126, + 676, + 507, + 692 + ], + "spans": [ + { + "bbox": [ + 126, + 676, + 507, + 692 + ], + "score": 1.0, + "content": "Aitor Lewkowycz, Erica Oliveira Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 127, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 127, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Díaz, Orhan Firat, Michele Catasta, Jason", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 127, + 699, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 127, + 699, + 507, + 712 + ], + "score": 1.0, + "content": "Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM:", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 127, + 711, + 424, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 424, + 723 + ], + "score": 1.0, + "content": "Scaling language modeling with pathways. ArXiv, abs/2204.02311, 2022.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 44 + } + ], + "page_idx": 9, + "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": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 163, + 84 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 111, + 91, + 506, + 168 + ], + "lines": [ + { + "bbox": [ + 110, + 90, + 506, + 103 + ], + "spans": [ + { + "bbox": [ + 110, + 90, + 506, + 103 + ], + "score": 1.0, + "content": "[1] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 128, + 102, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 128, + 102, + 505, + 113 + ], + "score": 1.0, + "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 111, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 126, + 111, + 506, + 126 + ], + "score": 1.0, + "content": "Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 124, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 127, + 124, + 506, + 135 + ], + "score": 1.0, + "content": "Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 133, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 127, + 133, + 506, + 146 + ], + "score": 1.0, + "content": "Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 127, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 128, + 156, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 128, + 156, + 506, + 168 + ], + "score": 1.0, + "content": "Information Processing Systems, volume 33, pages 1877–1901. Curran Associates, Inc., 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4, + "bbox_fs": [ + 110, + 90, + 506, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 175, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 110, + 174, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 110, + 174, + 506, + 187 + ], + "score": 1.0, + "content": "[2] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 185, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 127, + 185, + 505, + 200 + ], + "score": 1.0, + "content": "Multimodal few-shot learning with frozen language models. In Neural Information Processing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 197, + 189, + 210 + ], + "spans": [ + { + "bbox": [ + 127, + 197, + 189, + 210 + ], + "score": 1.0, + "content": "Systems, 2021.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 110, + 174, + 506, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 216, + 505, + 240 + ], + "lines": [ + { + "bbox": [ + 110, + 216, + 507, + 230 + ], + "spans": [ + { + "bbox": [ + 110, + 216, + 507, + 230 + ], + "score": 1.0, + "content": "[3] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 227, + 507, + 240 + ], + "spans": [ + { + "bbox": [ + 127, + 227, + 507, + 240 + ], + "score": 1.0, + "content": "and Furu Wei. Language models are general-purpose interfaces. ArXiv, abs/2206.06336, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 110, + 216, + 507, + 240 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 246, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 111, + 246, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 111, + 246, + 506, + 260 + ], + "score": 1.0, + "content": "[4] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 127, + 257, + 506, + 271 + ], + "score": 1.0, + "content": "Owais Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language:", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 268, + 483, + 282 + ], + "spans": [ + { + "bbox": [ + 127, + 268, + 483, + 282 + ], + "score": 1.0, + "content": "BEiT pretraining for all vision and vision-language tasks. ArXiv, abs/2208.10442, 2022.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 111, + 246, + 506, + 282 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 288, + 506, + 366 + ], + "lines": [ + { + "bbox": [ + 110, + 287, + 507, + 302 + ], + "spans": [ + { + "bbox": [ + 110, + 287, + 507, + 302 + ], + "score": 1.0, + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 127, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 309, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 126, + 309, + 506, + 323 + ], + "score": 1.0, + "content": "Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 322, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 127, + 322, + 506, + 333 + ], + "score": 1.0, + "content": "Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 126, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 126, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "Flamingo: a visual language model for few-shot learning. In Advances in Neural Information", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 354, + 489, + 366 + ], + "spans": [ + { + "bbox": [ + 126, + 354, + 489, + 366 + ], + "score": 1.0, + "content": "Processing Systems, 2022. URL https://openreview.net/forum?id=EbMuimAbPbs.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 110, + 287, + 507, + 366 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 373, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 111, + 373, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 111, + 373, + 506, + 386 + ], + "score": 1.0, + "content": "[6] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 384, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 384, + 506, + 396 + ], + "score": 1.0, + "content": "Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer. CM3: A", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 128, + 396, + 448, + 406 + ], + "spans": [ + { + "bbox": [ + 128, + 396, + 448, + 406 + ], + "score": 1.0, + "content": "causal masked multimodal model of the Internet. ArXiv, abs/2201.07520, 2022.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 111, + 373, + 506, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 414, + 505, + 447 + ], + "lines": [ + { + "bbox": [ + 110, + 413, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 110, + 413, + 506, + 429 + ], + "score": 1.0, + "content": "[7] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 426, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 127, + 426, + 506, + 437 + ], + "score": 1.0, + "content": "pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 434, + 154, + 449 + ], + "spans": [ + { + "bbox": [ + 126, + 434, + 154, + 449 + ], + "score": 1.0, + "content": "2023.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 110, + 413, + 506, + 449 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 455, + 504, + 489 + ], + "lines": [ + { + "bbox": [ + 110, + 455, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 110, + 455, + 506, + 469 + ], + "score": 1.0, + "content": "[8] John and Jean Raven. Raven Progressive Matrices, pages 223–237. Springer US, Boston,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 465, + 507, + 480 + ], + "spans": [ + { + "bbox": [ + 126, + 465, + 507, + 480 + ], + "score": 1.0, + "content": "MA, 2003. ISBN 978-1-4615-0153-4. doi: 10.1007/978-1-4615-0153-4_11. URL https:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 477, + 333, + 490 + ], + "spans": [ + { + "bbox": [ + 127, + 477, + 333, + 490 + ], + "score": 1.0, + "content": "//doi.org/10.1007/978-1-4615-0153-4_11.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 110, + 455, + 507, + 490 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 497, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 110, + 496, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 110, + 496, + 507, + 511 + ], + "score": 1.0, + "content": "[9] Patricia A Carpenter, Marcel A Just, and Peter Shell. What one intelligence test measures:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 127, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "a theoretical account of the processing in the raven progressive matrices test. Psychological", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 519, + 228, + 531 + ], + "spans": [ + { + "bbox": [ + 127, + 519, + 228, + 531 + ], + "score": 1.0, + "content": "review, 97(3):404, 1990.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 110, + 496, + 507, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 538, + 504, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "score": 1.0, + "content": "[10] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 127, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 560, + 230, + 571 + ], + "spans": [ + { + "bbox": [ + 127, + 560, + 230, + 571 + ], + "score": 1.0, + "content": "arXiv:2201.11903, 2022.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 537, + 505, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 579, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 108, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 108, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 590, + 507, + 604 + ], + "spans": [ + { + "bbox": [ + 127, + 590, + 507, + 604 + ], + "score": 1.0, + "content": "Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 600, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 127, + 600, + 507, + 615 + ], + "score": 1.0, + "content": "Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek B Rao, Parker Barnes, Yi Tay,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 611, + 507, + 625 + ], + "spans": [ + { + "bbox": [ + 127, + 611, + 507, + 625 + ], + "score": 1.0, + "content": "Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 623, + 507, + 636 + ], + "spans": [ + { + "bbox": [ + 127, + 623, + 507, + 636 + ], + "score": 1.0, + "content": "Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 634, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 127, + 634, + 506, + 647 + ], + "score": 1.0, + "content": "Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 644, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 127, + 644, + 505, + 658 + ], + "score": 1.0, + "content": "García, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 127, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 127, + 668, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 127, + 668, + 506, + 679 + ], + "score": 1.0, + "content": "Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 126, + 676, + 507, + 692 + ], + "spans": [ + { + "bbox": [ + 126, + 676, + 507, + 692 + ], + "score": 1.0, + "content": "Aitor Lewkowycz, Erica Oliveira Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 127, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 127, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Díaz, Orhan Firat, Michele Catasta, Jason", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 127, + 699, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 127, + 699, + 507, + 712 + ], + "score": 1.0, + "content": "Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM:", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 127, + 711, + 424, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 424, + 723 + ], + "score": 1.0, + "content": "Scaling language modeling with pathways. ArXiv, abs/2204.02311, 2022.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 44, + "bbox_fs": [ + 108, + 579, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[12] Chengyi Wang, Sanyuan Chen, Yu Wu, Zi-Hua Zhang, Long Zhou, Shujie Liu, Zhuo Chen,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "Yanqing Liu, Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, and Furu Wei. Neural codec", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 486, + 107 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 486, + 107 + ], + "score": 1.0, + "content": "language models are zero-shot text to speech synthesizers. ArXiv, abs/2301.02111, 2023.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 105, + 113, + 504, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "[13] Shuming Ma, Hongyu Wang, Shaohan Huang, Wenhui Wang, Zewen Chi, Li Dong, Alon", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 124, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 127, + 124, + 505, + 137 + ], + "score": 1.0, + "content": "Benhaim, Barun Patra, Vishrav Chaudhary, Xia Song, and Furu Wei. TorchScale: Transformers", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 136, + 286, + 147 + ], + "spans": [ + { + "bbox": [ + 127, + 136, + 286, + 147 + ], + "score": 1.0, + "content": "at scale. CoRR, abs/2211.13184, 2022.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 105, + 154, + 503, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "[14] Hongyu Wang, Shuming Ma, Shaohan Huang, Li Dong, Wenhui Wang, Zhiliang Peng, Yu Wu,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 127, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "Payal Bajaj, Saksham Singhal, Alon Benhaim, Barun Patra, Zhun Liu, Vishrav Chaudhary, Xia", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 176, + 440, + 189 + ], + "spans": [ + { + "bbox": [ + 127, + 176, + 440, + 189 + ], + "score": 1.0, + "content": "Song, and Furu Wei. Foundation transformers. CoRR, abs/2210.06423, 2022.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 105, + 195, + 505, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 208 + ], + "score": 1.0, + "content": "[15] Yutao Sun, Li Dong, Barun Patra, Shuming Ma, Shaohan Huang, Alon Benhaim, Vishrav", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 128, + 206, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 128, + 206, + 505, + 218 + ], + "score": 1.0, + "content": "Chaudhary, Xia Song, and Furu Wei. A length-extrapolatable transformer. arXiv preprint", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 217, + 231, + 228 + ], + "spans": [ + { + "bbox": [ + 127, + 217, + 231, + 228 + ], + "score": 1.0, + "content": "arXiv:2212.10554, 2022.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 105, + 236, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "score": 1.0, + "content": "[16] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 127, + 246, + 405, + 261 + ], + "score": 1.0, + "content": "Phang, Horace He, Anish Thite, Noa Nabeshima, et al. The pile: An", + "type": "text" + }, + { + "bbox": [ + 405, + 247, + 432, + 258 + ], + "score": 0.28, + "content": "8 0 0 \\mathrm { g b }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 246, + 506, + 261 + ], + "score": 1.0, + "content": "dataset of diverse", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 257, + 403, + 271 + ], + "spans": [ + { + "bbox": [ + 127, + 257, + 403, + 271 + ], + "score": 1.0, + "content": "text for language modeling. arXiv preprint arXiv:2101.00027, 2020.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 105, + 277, + 503, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 277, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "[17] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 127, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "Catanzaro. Megatron-lm: Training multi-billion parameter language models using model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 298, + 341, + 312 + ], + "spans": [ + { + "bbox": [ + 127, + 298, + 341, + 312 + ], + "score": 1.0, + "content": "parallelism. arXiv preprint arXiv:1909.08053, 2019.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 506, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "[18] Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 328, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 127, + 328, + 506, + 343 + ], + "score": 1.0, + "content": "Jared Casper, Zhun Liu, Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, Elton Zhang,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 127, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "Rewon Child, Reza Yazdani Aminabadi, Julie Bernauer, Xia Song, Mohammad Shoeybi,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 350, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 127, + 350, + 505, + 364 + ], + "score": 1.0, + "content": "Yuxiong He, Michael Houston, Saurabh Tiwary, and Bryan Catanzaro. Using DeepSpeed and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 361, + 507, + 376 + ], + "spans": [ + { + "bbox": [ + 127, + 361, + 507, + 376 + ], + "score": 1.0, + "content": "Megatron to train Megatron-Turing NLG 530B, a large-scale generative language model, 2022.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 507, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 507, + 394 + ], + "score": 1.0, + "content": "[19] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 127, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 127, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "5b: An open large-scale dataset for training next generation image-text models. arXiv preprint", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 414, + 231, + 425 + ], + "spans": [ + { + "bbox": [ + 127, + 414, + 231, + 425 + ], + "score": 1.0, + "content": "arXiv:2210.08402, 2022.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 503, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 446 + ], + "score": 1.0, + "content": "[20] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 127, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 455, + 500, + 467 + ], + "spans": [ + { + "bbox": [ + 127, + 455, + 500, + 467 + ], + "score": 1.0, + "content": "dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 504, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "[21] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 128, + 484, + 326, + 497 + ], + "spans": [ + { + "bbox": [ + 128, + 484, + 326, + 497 + ], + "score": 1.0, + "content": "Kim. Coyo-700m: Image-text pair dataset, 2022.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 506, + 559 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "[22] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 513, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 127, + 513, + 505, + 528 + ], + "score": 1.0, + "content": "cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 127, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 127, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 127, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 2556–2565. Association", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 547, + 279, + 560 + ], + "spans": [ + { + "bbox": [ + 127, + 547, + 279, + 560 + ], + "score": 1.0, + "content": "for Computational Linguistics, 2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 566, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 581 + ], + "score": 1.0, + "content": "[23] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 127, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "web-scale image-text pre-training to recognize long-tail visual concepts. In Proceedings of the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 127, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558–3568, 2021.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 105, + 606, + 506, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 621 + ], + "score": 1.0, + "content": "[24] Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 618, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 127, + 618, + 507, + 630 + ], + "score": 1.0, + "content": "language models with (almost) no human labor, 2022. URL https://arxiv.org/abs/2212.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 627, + 160, + 642 + ], + "spans": [ + { + "bbox": [ + 126, + 627, + 160, + 642 + ], + "score": 1.0, + "content": "09689.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 105, + 648, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 507, + 662 + ], + "score": 1.0, + "content": "[25] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 658, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 127, + 658, + 506, + 671 + ], + "score": 1.0, + "content": "Quoc V Le, Barret Zoph, Jason Wei, et al. The flan collection: Designing data and methods for", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 128, + 670, + 403, + 682 + ], + "spans": [ + { + "bbox": [ + 128, + 670, + 403, + 682 + ], + "score": 1.0, + "content": "effective instruction tuning. arXiv preprint arXiv:2301.13688, 2023.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "[26] Nasrin Mostafazadeh, Michael Roth, Annie Louis, Nathanael Chambers, and James Allen.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 127, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "Lsdsem 2017 shared task: The story cloze test. In Proceedings of the 2nd Workshop on Linking", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 127, + 711, + 455, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 455, + 723 + ], + "score": 1.0, + "content": "Models of Lexical, Sentential and Discourse-level Semantics, pages 46–51, 2017.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + } + ], + "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": "text", + "bbox": [ + 105, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[12] Chengyi Wang, Sanyuan Chen, Yu Wu, Zi-Hua Zhang, Long Zhou, Shujie Liu, Zhuo Chen,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "Yanqing Liu, Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, and Furu Wei. Neural codec", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 486, + 107 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 486, + 107 + ], + "score": 1.0, + "content": "language models are zero-shot text to speech synthesizers. ArXiv, abs/2301.02111, 2023.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 72, + 507, + 107 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 113, + 504, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "[13] Shuming Ma, Hongyu Wang, Shaohan Huang, Wenhui Wang, Zewen Chi, Li Dong, Alon", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 124, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 127, + 124, + 505, + 137 + ], + "score": 1.0, + "content": "Benhaim, Barun Patra, Vishrav Chaudhary, Xia Song, and Furu Wei. TorchScale: Transformers", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 136, + 286, + 147 + ], + "spans": [ + { + "bbox": [ + 127, + 136, + 286, + 147 + ], + "score": 1.0, + "content": "at scale. CoRR, abs/2211.13184, 2022.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 113, + 505, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 154, + 503, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "[14] Hongyu Wang, Shuming Ma, Shaohan Huang, Li Dong, Wenhui Wang, Zhiliang Peng, Yu Wu,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 127, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "Payal Bajaj, Saksham Singhal, Alon Benhaim, Barun Patra, Zhun Liu, Vishrav Chaudhary, Xia", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 176, + 440, + 189 + ], + "spans": [ + { + "bbox": [ + 127, + 176, + 440, + 189 + ], + "score": 1.0, + "content": "Song, and Furu Wei. Foundation transformers. CoRR, abs/2210.06423, 2022.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 153, + 506, + 189 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 195, + 505, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 208 + ], + "score": 1.0, + "content": "[15] Yutao Sun, Li Dong, Barun Patra, Shuming Ma, Shaohan Huang, Alon Benhaim, Vishrav", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 128, + 206, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 128, + 206, + 505, + 218 + ], + "score": 1.0, + "content": "Chaudhary, Xia Song, and Furu Wei. A length-extrapolatable transformer. arXiv preprint", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 217, + 231, + 228 + ], + "spans": [ + { + "bbox": [ + 127, + 217, + 231, + 228 + ], + "score": 1.0, + "content": "arXiv:2212.10554, 2022.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 194, + 505, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 236, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "score": 1.0, + "content": "[16] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 127, + 246, + 405, + 261 + ], + "score": 1.0, + "content": "Phang, Horace He, Anish Thite, Noa Nabeshima, et al. The pile: An", + "type": "text" + }, + { + "bbox": [ + 405, + 247, + 432, + 258 + ], + "score": 0.28, + "content": "8 0 0 \\mathrm { g b }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 246, + 506, + 261 + ], + "score": 1.0, + "content": "dataset of diverse", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 257, + 403, + 271 + ], + "spans": [ + { + "bbox": [ + 127, + 257, + 403, + 271 + ], + "score": 1.0, + "content": "text for language modeling. arXiv preprint arXiv:2101.00027, 2020.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 235, + 506, + 271 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 277, + 503, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 277, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "[17] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 127, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "Catanzaro. Megatron-lm: Training multi-billion parameter language models using model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 298, + 341, + 312 + ], + "spans": [ + { + "bbox": [ + 127, + 298, + 341, + 312 + ], + "score": 1.0, + "content": "parallelism. arXiv preprint arXiv:1909.08053, 2019.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 277, + 505, + 312 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 506, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "[18] Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 328, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 127, + 328, + 506, + 343 + ], + "score": 1.0, + "content": "Jared Casper, Zhun Liu, Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, Elton Zhang,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 127, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "Rewon Child, Reza Yazdani Aminabadi, Julie Bernauer, Xia Song, Mohammad Shoeybi,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 350, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 127, + 350, + 505, + 364 + ], + "score": 1.0, + "content": "Yuxiong He, Michael Houston, Saurabh Tiwary, and Bryan Catanzaro. Using DeepSpeed and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 361, + 507, + 376 + ], + "spans": [ + { + "bbox": [ + 127, + 361, + 507, + 376 + ], + "score": 1.0, + "content": "Megatron to train Megatron-Turing NLG 530B, a large-scale generative language model, 2022.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 318, + 507, + 376 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 507, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 507, + 394 + ], + "score": 1.0, + "content": "[19] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 127, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 127, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "5b: An open large-scale dataset for training next generation image-text models. arXiv preprint", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 414, + 231, + 425 + ], + "spans": [ + { + "bbox": [ + 127, + 414, + 231, + 425 + ], + "score": 1.0, + "content": "arXiv:2210.08402, 2022.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 380, + 507, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 503, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 446 + ], + "score": 1.0, + "content": "[20] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 127, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 455, + 500, + 467 + ], + "spans": [ + { + "bbox": [ + 127, + 455, + 500, + 467 + ], + "score": 1.0, + "content": "dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 431, + 505, + 467 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 504, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "[21] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 128, + 484, + 326, + 497 + ], + "spans": [ + { + "bbox": [ + 128, + 484, + 326, + 497 + ], + "score": 1.0, + "content": "Kim. Coyo-700m: Image-text pair dataset, 2022.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 473, + 505, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 506, + 559 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "[22] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 513, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 127, + 513, + 505, + 528 + ], + "score": 1.0, + "content": "cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 127, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 127, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 127, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 2556–2565. Association", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 547, + 279, + 560 + ], + "spans": [ + { + "bbox": [ + 127, + 547, + 279, + 560 + ], + "score": 1.0, + "content": "for Computational Linguistics, 2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 503, + 506, + 560 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 566, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 581 + ], + "score": 1.0, + "content": "[23] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 127, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "web-scale image-text pre-training to recognize long-tail visual concepts. In Proceedings of the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 127, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558–3568, 2021.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 564, + 506, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 606, + 506, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 621 + ], + "score": 1.0, + "content": "[24] Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 618, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 127, + 618, + 507, + 630 + ], + "score": 1.0, + "content": "language models with (almost) no human labor, 2022. URL https://arxiv.org/abs/2212.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 627, + 160, + 642 + ], + "spans": [ + { + "bbox": [ + 126, + 627, + 160, + 642 + ], + "score": 1.0, + "content": "09689.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 605, + 507, + 642 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 648, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 507, + 662 + ], + "score": 1.0, + "content": "[25] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 658, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 127, + 658, + 506, + 671 + ], + "score": 1.0, + "content": "Quoc V Le, Barret Zoph, Jason Wei, et al. The flan collection: Designing data and methods for", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 128, + 670, + 403, + 682 + ], + "spans": [ + { + "bbox": [ + 128, + 670, + 403, + 682 + ], + "score": 1.0, + "content": "effective instruction tuning. arXiv preprint arXiv:2301.13688, 2023.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 646, + 507, + 682 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "[26] Nasrin Mostafazadeh, Michael Roth, Annie Louis, Nathanael Chambers, and James Allen.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 127, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "Lsdsem 2017 shared task: The story cloze test. In Proceedings of the 2nd Workshop on Linking", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 127, + 711, + 455, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 455, + 723 + ], + "score": 1.0, + "content": "Models of Lexical, Sentential and Discourse-level Semantics, pages 46–51, 2017.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47, + "bbox_fs": [ + 106, + 689, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "[27] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 326, + 106 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 326, + 106 + ], + "score": 1.0, + "content": "Association for Computational Linguistics, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 112, + 504, + 145 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 127 + ], + "score": 1.0, + "content": "[28] Hector Levesque, Ernest Davis, and Leora Morgenstern. The winograd schema challenge.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 127, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "In Thirteenth International Conference on the Principles of Knowledge Representation and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 133, + 200, + 146 + ], + "spans": [ + { + "bbox": [ + 127, + 133, + 200, + 146 + ], + "score": 1.0, + "content": "Reasoning, 2012.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 105, + 151, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 506, + 165 + ], + "score": 1.0, + "content": "[29] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WinoGrande: An", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 163, + 460, + 176 + ], + "spans": [ + { + "bbox": [ + 127, + 163, + 460, + 176 + ], + "score": 1.0, + "content": "adversarial winograd schema challenge at scale. In AAAI, pages 8732–8740, 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 180, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 195 + ], + "score": 1.0, + "content": "[30] Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. Piqa: Reasoning", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 127, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "about physical commonsense in natural language. In Thirty-Fourth AAAI Conference on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 202, + 242, + 214 + ], + "spans": [ + { + "bbox": [ + 127, + 202, + 242, + 214 + ], + "score": 1.0, + "content": "Artificial Intelligence, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 219, + 506, + 287 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 233 + ], + "score": 1.0, + "content": "[31] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 127, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "Kristina Toutanova. BoolQ: Exploring the surprising difficulty of natural yes/no questions.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 127, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "In Proceedings of the 2019 Conference of the North American Chapter of the Association", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 126, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 127, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "Papers), pages 2924–2936, Minneapolis, Minnesota, June 2019. Association for Computational", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 274, + 498, + 287 + ], + "spans": [ + { + "bbox": [ + 127, + 274, + 498, + 287 + ], + "score": 1.0, + "content": "Linguistics. doi: 10.18653/v1/N19-1300. URL https://aclanthology.org/N19-1300.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 292, + 504, + 326 + ], + "lines": [ + { + "bbox": [ + 105, + 292, + 507, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 507, + 304 + ], + "score": 1.0, + "content": "[32] Marie-Catherine de Marneffe, Mandy Simons, and Judith Tonhauser. The CommitmentBank:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 302, + 507, + 317 + ], + "spans": [ + { + "bbox": [ + 126, + 302, + 507, + 317 + ], + "score": 1.0, + "content": "Investigating projection in naturally occurring discourse. Proceedings of Sinn und Bedeutung,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 128, + 314, + 234, + 326 + ], + "spans": [ + { + "bbox": [ + 128, + 314, + 234, + 326 + ], + "score": 1.0, + "content": "23(2):107–124, Jul. 2019.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 105, + 332, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "[33] Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S. Gordon. Choice of plausible", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 126, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "alternatives: An evaluation of commonsense causal reasoning. In AAAI Spring Symposium,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 352, + 155, + 366 + ], + "spans": [ + { + "bbox": [ + 126, + 352, + 155, + 366 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 384 + ], + "score": 1.0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 383, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 127, + 383, + 506, + 395 + ], + "score": 1.0, + "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 393, + 507, + 407 + ], + "spans": [ + { + "bbox": [ + 126, + 393, + 507, + 407 + ], + "score": 1.0, + "content": "models from natural language supervision. In International conference on machine learning,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 405, + 261, + 416 + ], + "spans": [ + { + "bbox": [ + 126, + 405, + 261, + 416 + ], + "score": 1.0, + "content": "pages 8748–8763. PMLR, 2021.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 422, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "[35] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 127, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 127, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "multimodal memes. In Advances in Neural Information Processing Systems, volume 33, pages", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 128, + 456, + 202, + 466 + ], + "spans": [ + { + "bbox": [ + 128, + 456, + 202, + 466 + ], + "score": 1.0, + "content": "2611–2624, 2020.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 105, + 472, + 502, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 504, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 504, + 487 + ], + "score": 1.0, + "content": "[36] Hessam Bagherinezhad, Hannaneh Hajishirzi, Yejin Choi, and Ali Farhadi. Are elephants bigger", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 128, + 484, + 449, + 496 + ], + "spans": [ + { + "bbox": [ + 128, + 484, + 449, + 496 + ], + "score": 1.0, + "content": "than butterflies? reasoning about sizes of objects. ArXiv, abs/1602.00753, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 501, + 503, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "[37] Tobias Norlund, Lovisa Hagström, and Richard Johansson. Transferring knowledge from vision", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 513, + 470, + 524 + ], + "spans": [ + { + "bbox": [ + 127, + 513, + 470, + 524 + ], + "score": 1.0, + "content": "to language: How to achieve it and how to measure it? ArXiv, abs/2109.11321, 2021.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 105, + 530, + 505, + 553 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 543 + ], + "score": 1.0, + "content": "[38] Elia Bruni, Gemma Boleda, Marco Baroni, and Nam Khanh Tran. Distributional semantics in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 541, + 237, + 552 + ], + "spans": [ + { + "bbox": [ + 127, + 541, + 237, + 552 + ], + "score": 1.0, + "content": "technicolor. In ACL, 2012.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 559, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 573 + ], + "score": 1.0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 127, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 581, + 219, + 592 + ], + "spans": [ + { + "bbox": [ + 126, + 581, + 219, + 592 + ], + "score": 1.0, + "content": "pages 740–755, 2014.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 504, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "[40] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 127, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "visual denotations: New similarity metrics for semantic inference over event descriptions. TACL,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 621, + 191, + 632 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 191, + 632 + ], + "score": 1.0, + "content": "2:67–78, 2014.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 638, + 504, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "[41] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 126, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 126, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "v in vqa matter: Elevating the role of image understanding in visual question answering. In", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 660, + 258, + 673 + ], + "spans": [ + { + "bbox": [ + 127, + 660, + 258, + 673 + ], + "score": 1.0, + "content": "CVPR, pages 6325–6334, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "[42] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 127, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 126, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 126, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 128, + 711, + 202, + 722 + ], + "spans": [ + { + "bbox": [ + 128, + 711, + 202, + 722 + ], + "score": 1.0, + "content": "3608–3617, 2018.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5 + } + ], + "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": [ + 106, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "[27] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 326, + 106 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 326, + 106 + ], + "score": 1.0, + "content": "Association for Computational Linguistics, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 72, + 506, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 112, + 504, + 145 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 127 + ], + "score": 1.0, + "content": "[28] Hector Levesque, Ernest Davis, and Leora Morgenstern. The winograd schema challenge.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 127, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "In Thirteenth International Conference on the Principles of Knowledge Representation and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 133, + 200, + 146 + ], + "spans": [ + { + "bbox": [ + 127, + 133, + 200, + 146 + ], + "score": 1.0, + "content": "Reasoning, 2012.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 110, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 151, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 506, + 165 + ], + "score": 1.0, + "content": "[29] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WinoGrande: An", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 163, + 460, + 176 + ], + "spans": [ + { + "bbox": [ + 127, + 163, + 460, + 176 + ], + "score": 1.0, + "content": "adversarial winograd schema challenge at scale. In AAAI, pages 8732–8740, 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 151, + 506, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 180, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 195 + ], + "score": 1.0, + "content": "[30] Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. Piqa: Reasoning", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 127, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "about physical commonsense in natural language. In Thirty-Fourth AAAI Conference on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 202, + 242, + 214 + ], + "spans": [ + { + "bbox": [ + 127, + 202, + 242, + 214 + ], + "score": 1.0, + "content": "Artificial Intelligence, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 179, + 505, + 214 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 219, + 506, + 287 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 233 + ], + "score": 1.0, + "content": "[31] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 127, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "Kristina Toutanova. BoolQ: Exploring the surprising difficulty of natural yes/no questions.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 127, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "In Proceedings of the 2019 Conference of the North American Chapter of the Association", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 126, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 127, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "Papers), pages 2924–2936, Minneapolis, Minnesota, June 2019. Association for Computational", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 274, + 498, + 287 + ], + "spans": [ + { + "bbox": [ + 127, + 274, + 498, + 287 + ], + "score": 1.0, + "content": "Linguistics. doi: 10.18653/v1/N19-1300. URL https://aclanthology.org/N19-1300.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 219, + 506, + 287 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 292, + 504, + 326 + ], + "lines": [ + { + "bbox": [ + 105, + 292, + 507, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 507, + 304 + ], + "score": 1.0, + "content": "[32] Marie-Catherine de Marneffe, Mandy Simons, and Judith Tonhauser. The CommitmentBank:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 302, + 507, + 317 + ], + "spans": [ + { + "bbox": [ + 126, + 302, + 507, + 317 + ], + "score": 1.0, + "content": "Investigating projection in naturally occurring discourse. Proceedings of Sinn und Bedeutung,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 128, + 314, + 234, + 326 + ], + "spans": [ + { + "bbox": [ + 128, + 314, + 234, + 326 + ], + "score": 1.0, + "content": "23(2):107–124, Jul. 2019.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 292, + 507, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 332, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "[33] Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S. Gordon. Choice of plausible", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 126, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "alternatives: An evaluation of commonsense causal reasoning. In AAAI Spring Symposium,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 352, + 155, + 366 + ], + "spans": [ + { + "bbox": [ + 126, + 352, + 155, + 366 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 331, + 506, + 366 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 384 + ], + "score": 1.0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 383, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 127, + 383, + 506, + 395 + ], + "score": 1.0, + "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 393, + 507, + 407 + ], + "spans": [ + { + "bbox": [ + 126, + 393, + 507, + 407 + ], + "score": 1.0, + "content": "models from natural language supervision. In International conference on machine learning,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 405, + 261, + 416 + ], + "spans": [ + { + "bbox": [ + 126, + 405, + 261, + 416 + ], + "score": 1.0, + "content": "pages 8748–8763. PMLR, 2021.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 372, + 507, + 416 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 422, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "[35] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 127, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 127, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "multimodal memes. In Advances in Neural Information Processing Systems, volume 33, pages", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 128, + 456, + 202, + 466 + ], + "spans": [ + { + "bbox": [ + 128, + 456, + 202, + 466 + ], + "score": 1.0, + "content": "2611–2624, 2020.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 422, + 506, + 466 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 472, + 502, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 504, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 504, + 487 + ], + "score": 1.0, + "content": "[36] Hessam Bagherinezhad, Hannaneh Hajishirzi, Yejin Choi, and Ali Farhadi. Are elephants bigger", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 128, + 484, + 449, + 496 + ], + "spans": [ + { + "bbox": [ + 128, + 484, + 449, + 496 + ], + "score": 1.0, + "content": "than butterflies? reasoning about sizes of objects. ArXiv, abs/1602.00753, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 471, + 504, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 501, + 503, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "[37] Tobias Norlund, Lovisa Hagström, and Richard Johansson. Transferring knowledge from vision", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 513, + 470, + 524 + ], + "spans": [ + { + "bbox": [ + 127, + 513, + 470, + 524 + ], + "score": 1.0, + "content": "to language: How to achieve it and how to measure it? ArXiv, abs/2109.11321, 2021.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 501, + 505, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 530, + 505, + 553 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 543 + ], + "score": 1.0, + "content": "[38] Elia Bruni, Gemma Boleda, Marco Baroni, and Nam Khanh Tran. Distributional semantics in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 541, + 237, + 552 + ], + "spans": [ + { + "bbox": [ + 127, + 541, + 237, + 552 + ], + "score": 1.0, + "content": "technicolor. In ACL, 2012.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 530, + 505, + 552 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 559, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 573 + ], + "score": 1.0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 127, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 581, + 219, + 592 + ], + "spans": [ + { + "bbox": [ + 126, + 581, + 219, + 592 + ], + "score": 1.0, + "content": "pages 740–755, 2014.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 558, + 506, + 592 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 504, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "[40] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 127, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "visual denotations: New similarity metrics for semantic inference over event descriptions. TACL,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 621, + 191, + 632 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 191, + 632 + ], + "score": 1.0, + "content": "2:67–78, 2014.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 597, + 506, + 632 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 638, + 504, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "[41] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 126, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 126, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "v in vqa matter: Elevating the role of image understanding in visual question answering. In", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 660, + 258, + 673 + ], + "spans": [ + { + "bbox": [ + 127, + 660, + 258, + 673 + ], + "score": 1.0, + "content": "CVPR, pages 6325–6334, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 638, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "[42] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 127, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 126, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 126, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 128, + 711, + 202, + 722 + ], + "spans": [ + { + "bbox": [ + 128, + 711, + 202, + 722 + ], + "score": 1.0, + "content": "3608–3617, 2018.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5, + "bbox_fs": [ + 106, + 677, + 506, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "score": 1.0, + "content": "[43] Xingyu Chen, Zihan Zhao, Lu Chen, JiaBao Ji, Danyang Zhang, Ao Luo, Yuxuan Xiong, and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 127, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "Kai Yu. WebSRC: A dataset for web-based structural reading comprehension. In Proceedings of", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 127, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4173–4185,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 127, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "Online and Punta Cana, Dominican Republic, November 2021. Association for Computational", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 115, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 127, + 115, + 505, + 129 + ], + "score": 1.0, + "content": "Linguistics. doi: 10.18653/v1/2021.emnlp-main.343. URL https://aclanthology.org/", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 127, + 233, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 127, + 233, + 140 + ], + "score": 1.0, + "content": "2021.emnlp-main.343.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 105, + 145, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "[44] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 156, + 303, + 168 + ], + "spans": [ + { + "bbox": [ + 127, + 156, + 303, + 168 + ], + "score": 1.0, + "content": "caltech-ucsd birds-200-2011 dataset. 2011.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 188 + ], + "score": 1.0, + "content": "[45] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 127, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "generation for visual-semantic arithmetic. 2022 IEEE/CVF Conference on Computer Vision", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 196, + 374, + 210 + ], + "spans": [ + { + "bbox": [ + 127, + 196, + 374, + 210 + ], + "score": 1.0, + "content": "and Pattern Recognition (CVPR), pages 17897–17907, 2021.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 506, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 506, + 228 + ], + "score": 1.0, + "content": "[46] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 127, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "multimodal generation on CLIP via vision-language knowledge distillation. In Findings of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 127, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "the Association for Computational Linguistics: ACL 2022, pages 2383–2395, Dublin, Ireland,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 128, + 248, + 507, + 259 + ], + "spans": [ + { + "bbox": [ + 128, + 248, + 507, + 259 + ], + "score": 1.0, + "content": "May 2022. Association for Computational Linguistics. doi: 10.18653/v1/2022.findings-acl.187.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 257, + 396, + 272 + ], + "spans": [ + { + "bbox": [ + 127, + 257, + 396, + 272 + ], + "score": 1.0, + "content": "URL https://aclanthology.org/2022.findings-acl.187.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 506, + 343 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "[47] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 286, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 126, + 286, + 506, + 302 + ], + "score": 1.0, + "content": "is worth millions of parameters: Low-resource prompt-based learning for vision-language", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 127, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 127, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "models. In Proceedings of the 60th Annual Meeting of the Association for Computa-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 127, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "tional Linguistics (Volume 1: Long Papers), pages 2763–2775, Dublin, Ireland, May 2022.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 320, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 127, + 320, + 507, + 333 + ], + "score": 1.0, + "content": "Association for Computational Linguistics. doi: 10.18653/v1/2022.acl-long.197. URL", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 331, + 354, + 344 + ], + "spans": [ + { + "bbox": [ + 127, + 331, + 354, + 344 + ], + "score": 1.0, + "content": "https://aclanthology.org/2022.acl-long.197.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 349, + 505, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "[48] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 359, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 127, + 359, + 506, + 374 + ], + "score": 1.0, + "content": "Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 127, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 382, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 127, + 382, + 506, + 395 + ], + "score": 1.0, + "content": "Language Processing, pages 1631–1642, Seattle, Washington, USA, October 2013. Association", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 393, + 496, + 406 + ], + "spans": [ + { + "bbox": [ + 127, + 393, + 496, + 406 + ], + "score": 1.0, + "content": "for Computational Linguistics. URL https://www.aclweb.org/anthology/D13-1170.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 411, + 503, + 445 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "[49] Weizhi Wang, Li Dong, Hao Cheng, Haoyu Song, Xiaodong Liu, Xifeng Yan, Jianfeng Gao, and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 421, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 126, + 421, + 505, + 436 + ], + "score": 1.0, + "content": "Furu Wei. Visually-augmented language modeling. In International Conference on Learning", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 434, + 222, + 446 + ], + "spans": [ + { + "bbox": [ + 127, + 434, + 222, + 446 + ], + "score": 1.0, + "content": "Representations, 2023.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 105, + 451, + 503, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 451, + 504, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 504, + 464 + ], + "score": 1.0, + "content": "[50] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 462, + 429, + 475 + ], + "spans": [ + { + "bbox": [ + 127, + 462, + 429, + 475 + ], + "score": 1.0, + "content": "Wu. Coca: Contrastive captioners are image-text foundation models, 2022.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 480, + 504, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "[51] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 492, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 127, + 492, + 506, + 504 + ], + "score": 1.0, + "content": "Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 501, + 154, + 515 + ], + "spans": [ + { + "bbox": [ + 126, + 501, + 154, + 515 + ], + "score": 1.0, + "content": "2023.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 507, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 507, + 534 + ], + "score": 1.0, + "content": "[52] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 128, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 128, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 542, + 380, + 555 + ], + "spans": [ + { + "bbox": [ + 127, + 542, + 380, + 555 + ], + "score": 1.0, + "content": "multilingual language-image model. arXiv:2209.06794, 2022.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 506, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 507, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 507, + 575 + ], + "score": 1.0, + "content": "[53] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 127, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 582, + 455, + 595 + ], + "spans": [ + { + "bbox": [ + 127, + 582, + 455, + 595 + ], + "score": 1.0, + "content": "multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 108, + 601, + 283, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 284, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 284, + 614 + ], + "score": 1.0, + "content": "[54] OpenAI. Gpt-4 technical report, 2023.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 105, + 619, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "[55] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 128, + 631, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 128, + 631, + 505, + 643 + ], + "score": 1.0, + "content": "Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 128, + 641, + 254, + 652 + ], + "spans": [ + { + "bbox": [ + 128, + 641, + 254, + 652 + ], + "score": 1.0, + "content": "visual instruction model, 2023.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 105, + 659, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "[56] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 126, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 681, + 408, + 694 + ], + "spans": [ + { + "bbox": [ + 127, + 681, + 408, + 694 + ], + "score": 1.0, + "content": "multimodal language model. arXiv preprint arXiv:2303.03378, 2023.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "[57] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 126, + 711, + 266, + 722 + ], + "spans": [ + { + "bbox": [ + 126, + 711, + 266, + 722 + ], + "score": 1.0, + "content": "preprint arXiv:2304.08485, 2023.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + } + ], + "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": "text", + "bbox": [ + 106, + 72, + 506, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "score": 1.0, + "content": "[43] Xingyu Chen, Zihan Zhao, Lu Chen, JiaBao Ji, Danyang Zhang, Ao Luo, Yuxuan Xiong, and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 127, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "Kai Yu. WebSRC: A dataset for web-based structural reading comprehension. In Proceedings of", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 127, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4173–4185,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 127, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "Online and Punta Cana, Dominican Republic, November 2021. Association for Computational", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 115, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 127, + 115, + 505, + 129 + ], + "score": 1.0, + "content": "Linguistics. doi: 10.18653/v1/2021.emnlp-main.343. URL https://aclanthology.org/", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 127, + 233, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 127, + 233, + 140 + ], + "score": 1.0, + "content": "2021.emnlp-main.343.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 72, + 507, + 140 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 145, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "[44] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 156, + 303, + 168 + ], + "spans": [ + { + "bbox": [ + 127, + 156, + 303, + 168 + ], + "score": 1.0, + "content": "caltech-ucsd birds-200-2011 dataset. 2011.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 145, + 505, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 188 + ], + "score": 1.0, + "content": "[45] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 127, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "generation for visual-semantic arithmetic. 2022 IEEE/CVF Conference on Computer Vision", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 196, + 374, + 210 + ], + "spans": [ + { + "bbox": [ + 127, + 196, + 374, + 210 + ], + "score": 1.0, + "content": "and Pattern Recognition (CVPR), pages 17897–17907, 2021.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 174, + 505, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 506, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 506, + 228 + ], + "score": 1.0, + "content": "[46] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 127, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "multimodal generation on CLIP via vision-language knowledge distillation. In Findings of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 127, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "the Association for Computational Linguistics: ACL 2022, pages 2383–2395, Dublin, Ireland,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 128, + 248, + 507, + 259 + ], + "spans": [ + { + "bbox": [ + 128, + 248, + 507, + 259 + ], + "score": 1.0, + "content": "May 2022. Association for Computational Linguistics. doi: 10.18653/v1/2022.findings-acl.187.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 257, + 396, + 272 + ], + "spans": [ + { + "bbox": [ + 127, + 257, + 396, + 272 + ], + "score": 1.0, + "content": "URL https://aclanthology.org/2022.findings-acl.187.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 214, + 507, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 506, + 343 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "[47] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 286, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 126, + 286, + 506, + 302 + ], + "score": 1.0, + "content": "is worth millions of parameters: Low-resource prompt-based learning for vision-language", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 127, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 127, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "models. In Proceedings of the 60th Annual Meeting of the Association for Computa-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 127, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "tional Linguistics (Volume 1: Long Papers), pages 2763–2775, Dublin, Ireland, May 2022.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 320, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 127, + 320, + 507, + 333 + ], + "score": 1.0, + "content": "Association for Computational Linguistics. doi: 10.18653/v1/2022.acl-long.197. URL", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 331, + 354, + 344 + ], + "spans": [ + { + "bbox": [ + 127, + 331, + 354, + 344 + ], + "score": 1.0, + "content": "https://aclanthology.org/2022.acl-long.197.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 276, + 507, + 344 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 349, + 505, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "[48] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 359, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 127, + 359, + 506, + 374 + ], + "score": 1.0, + "content": "Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 127, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 382, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 127, + 382, + 506, + 395 + ], + "score": 1.0, + "content": "Language Processing, pages 1631–1642, Seattle, Washington, USA, October 2013. Association", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 393, + 496, + 406 + ], + "spans": [ + { + "bbox": [ + 127, + 393, + 496, + 406 + ], + "score": 1.0, + "content": "for Computational Linguistics. URL https://www.aclweb.org/anthology/D13-1170.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 349, + 506, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 411, + 503, + 445 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "[49] Weizhi Wang, Li Dong, Hao Cheng, Haoyu Song, Xiaodong Liu, Xifeng Yan, Jianfeng Gao, and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 421, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 126, + 421, + 505, + 436 + ], + "score": 1.0, + "content": "Furu Wei. Visually-augmented language modeling. In International Conference on Learning", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 434, + 222, + 446 + ], + "spans": [ + { + "bbox": [ + 127, + 434, + 222, + 446 + ], + "score": 1.0, + "content": "Representations, 2023.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 411, + 505, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 451, + 503, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 451, + 504, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 504, + 464 + ], + "score": 1.0, + "content": "[50] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 462, + 429, + 475 + ], + "spans": [ + { + "bbox": [ + 127, + 462, + 429, + 475 + ], + "score": 1.0, + "content": "Wu. Coca: Contrastive captioners are image-text foundation models, 2022.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 451, + 504, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 480, + 504, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "[51] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 492, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 127, + 492, + 506, + 504 + ], + "score": 1.0, + "content": "Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 501, + 154, + 515 + ], + "spans": [ + { + "bbox": [ + 126, + 501, + 154, + 515 + ], + "score": 1.0, + "content": "2023.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 480, + 506, + 515 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 507, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 507, + 534 + ], + "score": 1.0, + "content": "[52] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 128, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 128, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 542, + 380, + 555 + ], + "spans": [ + { + "bbox": [ + 127, + 542, + 380, + 555 + ], + "score": 1.0, + "content": "multilingual language-image model. arXiv:2209.06794, 2022.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 520, + 507, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 506, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 507, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 507, + 575 + ], + "score": 1.0, + "content": "[53] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 127, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 582, + 455, + 595 + ], + "spans": [ + { + "bbox": [ + 127, + 582, + 455, + 595 + ], + "score": 1.0, + "content": "multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 560, + 507, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 601, + 283, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 284, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 284, + 614 + ], + "score": 1.0, + "content": "[54] OpenAI. Gpt-4 technical report, 2023.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 600, + 284, + 614 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 619, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "[55] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 128, + 631, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 128, + 631, + 505, + 643 + ], + "score": 1.0, + "content": "Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 128, + 641, + 254, + 652 + ], + "spans": [ + { + "bbox": [ + 128, + 641, + 254, + 652 + ], + "score": 1.0, + "content": "visual instruction model, 2023.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 619, + 506, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 659, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "[56] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 126, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 681, + 408, + 694 + ], + "spans": [ + { + "bbox": [ + 127, + 681, + 408, + 694 + ], + "score": 1.0, + "content": "multimodal language model. arXiv preprint arXiv:2303.03378, 2023.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 659, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "[57] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 126, + 711, + 266, + 722 + ], + "spans": [ + { + "bbox": [ + 126, + 711, + 266, + 722 + ], + "score": 1.0, + "content": "preprint arXiv:2304.08485, 2023.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 699, + 505, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 71, + 507, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: En-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 84, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 84, + 506, + 97 + ], + "score": 1.0, + "content": "hancing vision-language understanding with advanced large language models. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 128, + 95, + 230, + 106 + ], + "spans": [ + { + "bbox": [ + 128, + 95, + 230, + 106 + ], + "score": 1.0, + "content": "arXiv:2304.10592, 2023.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 111, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 128 + ], + "score": 1.0, + "content": "[59] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 123, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 123, + 506, + 138 + ], + "score": 1.0, + "content": "Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 135, + 424, + 148 + ], + "spans": [ + { + "bbox": [ + 127, + 135, + 424, + 148 + ], + "score": 1.0, + "content": "vision-language models with instruction tuning. arXiv:2305.06500, 2023.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "[60] Zewen Chi, Li Dong, Shaohan Huang, Damai Dai, Shuming Ma, Barun Patra, Saksham Singhal,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 127, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "Payal Bajaj, Xia Song, Xian-Ling Mao, Heyan Huang, and Furu Wei. On the representation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 175, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 127, + 175, + 506, + 190 + ], + "score": 1.0, + "content": "collapse of sparse mixture of experts. In Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 187, + 401, + 200 + ], + "spans": [ + { + "bbox": [ + 127, + 187, + 401, + 200 + ], + "score": 1.0, + "content": "2022. URL https://openreview.net/forum?id=mWaYC6CZf5.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 220 + ], + "score": 1.0, + "content": "[61] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 215, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 126, + 215, + 505, + 231 + ], + "score": 1.0, + "content": "Mike Lewis, Luke Zettlemoyer, and Wen tau Yih. Retrieval-augmented multimodal language", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 128, + 228, + 293, + 239 + ], + "spans": [ + { + "bbox": [ + 128, + 228, + 293, + 239 + ], + "score": 1.0, + "content": "modeling. ArXiv, abs/2211.12561, 2022.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "[62] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 128, + 257, + 402, + 270 + ], + "spans": [ + { + "bbox": [ + 128, + 257, + 402, + 270 + ], + "score": 1.0, + "content": "for multimodal generation. arXiv preprint arXiv:2301.13823, 2023.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7 + } + ], + "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": [ + 105, + 71, + 507, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: En-", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 84, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 84, + 506, + 97 + ], + "score": 1.0, + "content": "hancing vision-language understanding with advanced large language models. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 128, + 95, + 230, + 106 + ], + "spans": [ + { + "bbox": [ + 128, + 95, + 230, + 106 + ], + "score": 1.0, + "content": "arXiv:2304.10592, 2023.", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 111, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 128 + ], + "score": 1.0, + "content": "[59] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 123, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 123, + 506, + 138 + ], + "score": 1.0, + "content": "Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 135, + 424, + 148 + ], + "spans": [ + { + "bbox": [ + 127, + 135, + 424, + 148 + ], + "score": 1.0, + "content": "vision-language models with instruction tuning. arXiv:2305.06500, 2023.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "[60] Zewen Chi, Li Dong, Shaohan Huang, Damai Dai, Shuming Ma, Barun Patra, Saksham Singhal,", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 127, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "Payal Bajaj, Xia Song, Xian-Ling Mao, Heyan Huang, and Furu Wei. On the representation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 175, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 127, + 175, + 506, + 190 + ], + "score": 1.0, + "content": "collapse of sparse mixture of experts. In Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 187, + 401, + 200 + ], + "spans": [ + { + "bbox": [ + 127, + 187, + 401, + 200 + ], + "score": 1.0, + "content": "2022. URL https://openreview.net/forum?id=mWaYC6CZf5.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 204, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 220 + ], + "score": 1.0, + "content": "[61] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang,", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 215, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 126, + 215, + 505, + 231 + ], + "score": 1.0, + "content": "Mike Lewis, Luke Zettlemoyer, and Wen tau Yih. Retrieval-augmented multimodal language", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 128, + 228, + 293, + 239 + ], + "spans": [ + { + "bbox": [ + 128, + 228, + 293, + 239 + ], + "score": 1.0, + "content": "modeling. ArXiv, abs/2211.12561, 2022.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "[62] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 257, + 402, + 270 + ], + "spans": [ + { + "bbox": [ + 128, + 257, + 402, + 270 + ], + "score": 1.0, + "content": "for multimodal generation. arXiv preprint arXiv:2301.13823, 2023.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + } + ], + "index": 7, + "bbox_fs": [ + 105, + 73, + 506, + 270 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file