| BLOOM176B | OPT175B | GLM-130B* |
| qnli (valid, median of 5 prompts) | 50.9 | 55.4 | 86.7 |
| mnli (valid, median of 15 prompts) | 35.5 | 36.0 | 85.7 |
| mnli_mismatched (valid, median of 15 prompts) | 35.5 | 36.0 | 84.6 |
| wnli (valid, median of 5 prompts) | 57.7 | 53.5 | 67.6 |
| glue/cola (valid, median of 5 prompts) | 39.0 | 44.4 | 57.6 |
| glue/mrpc (valid, median of 5 prompts) | 31.6 | 44.6 | 87.3 |
",
"type": "table",
"image_path": "fa3f6f860515a5520653a4035bb9e172b2f893a9dc71772b981db648bc914436.jpg"
}
]
}
],
"index": 32,
"virtual_lines": [
{
"bbox": [
110,
518,
498,
548.3333333333334
],
"spans": [],
"index": 31
},
{
"bbox": [
110,
548.3333333333334,
498,
578.6666666666667
],
"spans": [],
"index": 32
},
{
"bbox": [
110,
578.6666666666667,
498,
609.0000000000001
],
"spans": [],
"index": 33
}
]
}
],
"index": 30.5
},
{
"type": "title",
"bbox": [
107,
632,
196,
644
],
"lines": [
{
"bbox": [
105,
631,
198,
646
],
"spans": [
{
"bbox": [
105,
631,
198,
646
],
"score": 1.0,
"content": "C.13 SUPERGLUE",
"type": "text"
}
],
"index": 34
}
],
"index": 34
},
{
"type": "text",
"bbox": [
107,
654,
505,
732
],
"lines": [
{
"bbox": [
106,
654,
505,
667
],
"spans": [
{
"bbox": [
106,
654,
505,
667
],
"score": 1.0,
"content": "We also report our evaluation of GLM-130B on the SuperGLUE (Wang et al., 2019) benchmark,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
666,
505,
678
],
"spans": [
{
"bbox": [
105,
666,
505,
678
],
"score": 1.0,
"content": "which consists 8 different natural language understanding challenges. Noted that these results",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
677,
505,
689
],
"spans": [
{
"bbox": [
105,
677,
505,
689
],
"score": 1.0,
"content": "are neither zero/few-shot nor fine-tuned results, because 7 out of 8 tasks’ training sets have been",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
505,
700
],
"score": 1.0,
"content": "included in GLM-130B’s MIP training (except for ReCoRD) together with other 67 multi-task",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
698,
505,
711
],
"spans": [
{
"bbox": [
105,
698,
505,
711
],
"score": 1.0,
"content": "datasets; however, GLM-130B is also not individually fine-tuned on any of them. Therefore, these",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
709,
505,
722
],
"spans": [
{
"bbox": [
105,
709,
505,
722
],
"score": 1.0,
"content": "results are not for relative comparison for any other models’, but only for readers’ reference on",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
720,
228,
733
],
"spans": [
{
"bbox": [
106,
720,
228,
733
],
"score": 1.0,
"content": "GLM-130B’s absolute ability.",
"type": "text"
}
],
"index": 41
}
],
"index": 38
}
],
"page_idx": 43,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
108,
27,
293,
37
],
"lines": [
{
"bbox": [
106,
26,
293,
38
],
"spans": [
{
"bbox": [
106,
26,
293,
38
],
"score": 1.0,
"content": "Published as a conference paper at ICLR 2023",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
752,
311,
759
],
"lines": []
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
107,
82,
504,
127
],
"lines": [
{
"bbox": [
106,
82,
505,
94
],
"spans": [
{
"bbox": [
106,
82,
505,
94
],
"score": 1.0,
"content": "200B English tokens and thus does not memorize the detailed knowledge very well. Since CBQA",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
94,
505,
106
],
"spans": [
{
"bbox": [
105,
94,
505,
106
],
"score": 1.0,
"content": "seems to be a task that especially stresses memorization, as is indicated by Chinchilla (Hoffmann",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
104,
104,
506,
117
],
"spans": [
{
"bbox": [
104,
104,
506,
117
],
"score": 1.0,
"content": "et al., 2022)’s a strong performance, we think with sufficient training later, GLM-130B can perform",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
115,
135,
128
],
"spans": [
{
"bbox": [
105,
115,
135,
128
],
"score": 1.0,
"content": "better.",
"type": "text"
}
],
"index": 3
}
],
"index": 1.5,
"bbox_fs": [
104,
82,
506,
128
]
},
{
"type": "title",
"bbox": [
108,
144,
261,
155
],
"lines": [
{
"bbox": [
106,
144,
262,
157
],
"spans": [
{
"bbox": [
106,
144,
262,
157
],
"score": 1.0,
"content": "C.11 COMMONSENSE REASONING",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
107,
166,
505,
255
],
"lines": [
{
"bbox": [
105,
165,
505,
178
],
"spans": [
{
"bbox": [
105,
165,
505,
178
],
"score": 1.0,
"content": "Here we evaluate GLM-130B and some other LLMs on commonsense reasoning abilities. As we",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
177,
505,
189
],
"spans": [
{
"bbox": [
105,
177,
505,
189
],
"score": 1.0,
"content": "have included PIQA (Bisk et al., 2020), ARC (Clark et al., 2018), and OpenbookQA (Mihaylov",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
188,
505,
201
],
"spans": [
{
"bbox": [
105,
188,
505,
201
],
"score": 1.0,
"content": "et al., 2018) in the MIP training, we select another two widely adopted commonsense reasoning",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
198,
505,
211
],
"spans": [
{
"bbox": [
105,
198,
505,
211
],
"score": 1.0,
"content": "datasets in our evaluation: Commonsense QA (Talmor et al., 2019) and Multiple-choice Temporal",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
210,
505,
222
],
"spans": [
{
"bbox": [
105,
210,
505,
222
],
"score": 1.0,
"content": "Commonsense (MC-TACO, Zhou et al. (2019)). For Commonsense QA, we test the GPT-3 via",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
220,
506,
234
],
"spans": [
{
"bbox": [
105,
220,
506,
234
],
"score": 1.0,
"content": "OpenAI Davinci API, BLOOM-176B via its Huggingface Implementation, and GLM-130B using",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
232,
506,
245
],
"spans": [
{
"bbox": [
105,
232,
506,
245
],
"score": 1.0,
"content": "the prompt “answer_given_question_without_options” from promptsource (Bach et al., 2022). For",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
243,
445,
255
],
"spans": [
{
"bbox": [
105,
243,
445,
255
],
"score": 1.0,
"content": "StrategyQA, we follow the EM computation method provided in (Zhou et al., 2019).",
"type": "text"
}
],
"index": 12
}
],
"index": 8.5,
"bbox_fs": [
105,
165,
506,
255
]
},
{
"type": "text",
"bbox": [
108,
260,
505,
304
],
"lines": [
{
"bbox": [
106,
260,
505,
272
],
"spans": [
{
"bbox": [
106,
260,
505,
272
],
"score": 1.0,
"content": "The results are shown in Table 19. As we can see, GLM-130B performs the best on both Com-",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
271,
506,
284
],
"spans": [
{
"bbox": [
105,
271,
506,
284
],
"score": 1.0,
"content": "monsense QA and MC-TACO across evaluated LLMs, demonstrating that GLM-130B has a good",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
282,
506,
294
],
"spans": [
{
"bbox": [
105,
282,
506,
294
],
"score": 1.0,
"content": "grasp of commonsense knowledge. OPT’s results are not included due to the reason described in",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
292,
168,
306
],
"spans": [
{
"bbox": [
105,
292,
168,
306
],
"score": 1.0,
"content": "Appendix C.3.",
"type": "text"
}
],
"index": 16
}
],
"index": 14.5,
"bbox_fs": [
105,
260,
506,
306
]
},
{
"type": "text",
"bbox": [
108,
322,
475,
333
],
"lines": [
{
"bbox": [
106,
322,
477,
334
],
"spans": [
{
"bbox": [
106,
322,
477,
334
],
"score": 1.0,
"content": "C.12 FIXED LABEL DATASETS: A CASE STUDY IN NATURAL LANGUAGE INFERENCE",
"type": "text"
}
],
"index": 17
}
],
"index": 17,
"bbox_fs": [
106,
322,
477,
334
]
},
{
"type": "text",
"bbox": [
107,
344,
504,
399
],
"lines": [
{
"bbox": [
106,
344,
505,
356
],
"spans": [
{
"bbox": [
106,
344,
505,
356
],
"score": 1.0,
"content": "As is discussed in Section 5, we adopt a rather strict criterion for selecting datasets for zero/few-shot",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
355,
504,
367
],
"spans": [
{
"bbox": [
106,
355,
504,
367
],
"score": 1.0,
"content": "learning in GLM-130B’s evaluation due to the use of MIP. Nevertheless, the criterion significantly",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
366,
505,
379
],
"spans": [
{
"bbox": [
106,
366,
505,
379
],
"score": 1.0,
"content": "reduces the dataset we could currently evaluate, and especially some readers have doubted whether",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
375,
506,
392
],
"spans": [
{
"bbox": [
105,
375,
506,
392
],
"score": 1.0,
"content": "the restriction of not evaluating on MIP-seen fixed-label datasets is necessary (e.g., natural language",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
388,
505,
401
],
"spans": [
{
"bbox": [
106,
388,
505,
401
],
"score": 1.0,
"content": "inference (NLI)), and suggest that we may report them in an independent section to avoid confusion.",
"type": "text"
}
],
"index": 22
}
],
"index": 20,
"bbox_fs": [
105,
344,
506,
401
]
},
{
"type": "text",
"bbox": [
107,
405,
505,
460
],
"lines": [
{
"bbox": [
105,
404,
505,
418
],
"spans": [
{
"bbox": [
105,
404,
505,
418
],
"score": 1.0,
"content": "Frankly speaking, in such a setting GLM-130B’s zero/few-shot learning could be quite advanta-",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
416,
505,
429
],
"spans": [
{
"bbox": [
105,
416,
505,
429
],
"score": 1.0,
"content": "geous. Below, we take NLI as a typical example to show GLM-130B’s outperformance in the",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
427,
505,
439
],
"spans": [
{
"bbox": [
105,
427,
505,
439
],
"score": 1.0,
"content": "scenarios. We include 6 widely-used NLI datasets–which are not incorporated in GLM-130B’s MIP",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
438,
506,
451
],
"spans": [
{
"bbox": [
105,
438,
506,
451
],
"score": 1.0,
"content": "training, as the benchmarks. The results are presented in Table 20, which shows that GLM-130B’s",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
448,
396,
462
],
"spans": [
{
"bbox": [
105,
448,
396,
462
],
"score": 1.0,
"content": "“zero-shot” performance could be much better due to the seen task type.",
"type": "text"
}
],
"index": 27
}
],
"index": 25,
"bbox_fs": [
105,
404,
506,
462
]
},
{
"type": "table",
"bbox": [
110,
518,
498,
609
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
476,
505,
510
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
475,
505,
489
],
"spans": [
{
"bbox": [
105,
475,
505,
489
],
"score": 1.0,
"content": "Table 20: “Zero-shot” results of GLM-130B on 6 typical natural language inference (NLI) datasets.",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
486,
505,
499
],
"spans": [
{
"bbox": [
106,
487,
112,
497
],
"score": 0.29,
"content": "^ *",
"type": "inline_equation"
},
{
"bbox": [
112,
486,
505,
499
],
"score": 1.0,
"content": "DISCLAIMER: Despite the datasets are never seen, some other NLI datasets have been in-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
497,
502,
511
],
"spans": [
{
"bbox": [
105,
497,
502,
511
],
"score": 1.0,
"content": "cluded in GLM-130B’s MIP, making it different from the existing standard zero-shot setting.",
"type": "text"
}
],
"index": 30
}
],
"index": 29
},
{
"type": "table_body",
"bbox": [
110,
518,
498,
609
],
"group_id": 0,
"lines": [
{
"bbox": [
110,
518,
498,
609
],
"spans": [
{
"bbox": [
110,
518,
498,
609
],
"score": 0.983,
"html": " | BLOOM176B | OPT175B | GLM-130B* |
| qnli (valid, median of 5 prompts) | 50.9 | 55.4 | 86.7 |
| mnli (valid, median of 15 prompts) | 35.5 | 36.0 | 85.7 |
| mnli_mismatched (valid, median of 15 prompts) | 35.5 | 36.0 | 84.6 |
| wnli (valid, median of 5 prompts) | 57.7 | 53.5 | 67.6 |
| glue/cola (valid, median of 5 prompts) | 39.0 | 44.4 | 57.6 |
| glue/mrpc (valid, median of 5 prompts) | 31.6 | 44.6 | 87.3 |
",
"type": "table",
"image_path": "fa3f6f860515a5520653a4035bb9e172b2f893a9dc71772b981db648bc914436.jpg"
}
]
}
],
"index": 32,
"virtual_lines": [
{
"bbox": [
110,
518,
498,
548.3333333333334
],
"spans": [],
"index": 31
},
{
"bbox": [
110,
548.3333333333334,
498,
578.6666666666667
],
"spans": [],
"index": 32
},
{
"bbox": [
110,
578.6666666666667,
498,
609.0000000000001
],
"spans": [],
"index": 33
}
]
}
],
"index": 30.5
},
{
"type": "title",
"bbox": [
107,
632,
196,
644
],
"lines": [
{
"bbox": [
105,
631,
198,
646
],
"spans": [
{
"bbox": [
105,
631,
198,
646
],
"score": 1.0,
"content": "C.13 SUPERGLUE",
"type": "text"
}
],
"index": 34
}
],
"index": 34
},
{
"type": "text",
"bbox": [
107,
654,
505,
732
],
"lines": [
{
"bbox": [
106,
654,
505,
667
],
"spans": [
{
"bbox": [
106,
654,
505,
667
],
"score": 1.0,
"content": "We also report our evaluation of GLM-130B on the SuperGLUE (Wang et al., 2019) benchmark,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
666,
505,
678
],
"spans": [
{
"bbox": [
105,
666,
505,
678
],
"score": 1.0,
"content": "which consists 8 different natural language understanding challenges. Noted that these results",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
677,
505,
689
],
"spans": [
{
"bbox": [
105,
677,
505,
689
],
"score": 1.0,
"content": "are neither zero/few-shot nor fine-tuned results, because 7 out of 8 tasks’ training sets have been",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
505,
700
],
"score": 1.0,
"content": "included in GLM-130B’s MIP training (except for ReCoRD) together with other 67 multi-task",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
698,
505,
711
],
"spans": [
{
"bbox": [
105,
698,
505,
711
],
"score": 1.0,
"content": "datasets; however, GLM-130B is also not individually fine-tuned on any of them. Therefore, these",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
709,
505,
722
],
"spans": [
{
"bbox": [
105,
709,
505,
722
],
"score": 1.0,
"content": "results are not for relative comparison for any other models’, but only for readers’ reference on",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
720,
228,
733
],
"spans": [
{
"bbox": [
106,
720,
228,
733
],
"score": 1.0,
"content": "GLM-130B’s absolute ability.",
"type": "text"
}
],
"index": 41
}
],
"index": 38,
"bbox_fs": [
105,
654,
505,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
106,
79,
503,
171
],
"blocks": [
{
"type": "image_body",
"bbox": [
106,
79,
503,
171
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
79,
503,
171
],
"spans": [
{
"bbox": [
106,
79,
503,
171
],
"score": 0.944,
"type": "image",
"image_path": "f55bb0dce90be1d84f3a233001579f48bc4c0a4525e3554ecddd4b386950fcb8.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
106,
79,
503,
109.66666666666667
],
"spans": [],
"index": 0
},
{
"bbox": [
106,
109.66666666666667,
503,
140.33333333333334
],
"spans": [],
"index": 1
},
{
"bbox": [
106,
140.33333333333334,
503,
171.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
178,
505,
223
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
177,
505,
191
],
"spans": [
{
"bbox": [
105,
177,
505,
191
],
"score": 1.0,
"content": "Figure 17: GLM-130B (uni and bi)’s untuned results on SuperGLUE development set, using prompt-",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
190,
505,
202
],
"spans": [
{
"bbox": [
105,
190,
505,
202
],
"score": 1.0,
"content": "source (Bach et al., 2022) prompts and task formulation. DISCLAIMER: Noted that some of the",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
200,
506,
214
],
"spans": [
{
"bbox": [
105,
200,
506,
214
],
"score": 1.0,
"content": "SuperGLUE training sets have been included in the MIP training. We report the results here",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
210,
225,
225
],
"spans": [
{
"bbox": [
105,
210,
225,
225
],
"score": 1.0,
"content": "only for readers’ reference.",
"type": "text"
}
],
"index": 6
}
],
"index": 4.5
}
],
"index": 2.75
},
{
"type": "image",
"bbox": [
144,
234,
465,
330
],
"blocks": [
{
"type": "image_body",
"bbox": [
144,
234,
465,
330
],
"group_id": 1,
"lines": [
{
"bbox": [
144,
234,
465,
330
],
"spans": [
{
"bbox": [
144,
234,
465,
330
],
"score": 0.967,
"type": "image",
"image_path": "7a0df67091ca37c3cd95d35f352b4c21534f3fcee8cabf3b659eaf309a079cfe.jpg"
}
]
}
],
"index": 8,
"virtual_lines": [
{
"bbox": [
144,
234,
465,
266.0
],
"spans": [],
"index": 7
},
{
"bbox": [
144,
266.0,
465,
298.0
],
"spans": [],
"index": 8
},
{
"bbox": [
144,
298.0,
465,
330.0
],
"spans": [],
"index": 9
}
]
},
{
"type": "image_caption",
"bbox": [
107,
336,
504,
360
],
"group_id": 1,
"lines": [
{
"bbox": [
105,
334,
506,
351
],
"spans": [
{
"bbox": [
105,
334,
506,
351
],
"score": 1.0,
"content": "Figure 18: Chain-of-thought prompting can also improve GLM-130B’s performance on reasoning",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
346,
263,
362
],
"spans": [
{
"bbox": [
105,
346,
263,
362
],
"score": 1.0,
"content": "tasks compared to standard prompting.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.5
}
],
"index": 9.25
},
{
"type": "table",
"bbox": [
119,
380,
492,
416
],
"blocks": [
{
"type": "table_body",
"bbox": [
119,
380,
492,
416
],
"group_id": 0,
"lines": [
{
"bbox": [
119,
380,
492,
416
],
"spans": [
{
"bbox": [
119,
380,
492,
416
],
"score": 0.958,
"html": "| Configuration Key | Value |
| adam_beta1 | 0.9 |
| adam_beta2 | 0.95 |
| adam_eps | 1e-08 |
| aggregated_samples_per_sequence | 4 |
| 0.1 |
| attention_dropout attention_softmax_in_fp32 | |
| True |
| average_block_length | 3 |
| bias_dropout_fusion | True True |
| checkpoint_activations | |
| checkpoint_in_cpu | False |
| checkpoint_num_layers | |
| clip_grad contigious_checkpointing | 1.0 |
| False |
| cpu_optimizer | False |
| data_parallel_size | 24 |
| deepnorm | True |
| distributed_backend | nccl |
| eval_interval | 1000 |
| eval_iters ffn_hidden_size | 3 |
| fp16 | 32768 |
| True |
| global_batch_size glu_activation | 4224 |
| gpt_prob | geglu |
| hidden_dropout | 0.7 |
| hidden_size | 0.1 |
| hysteresis | 12288 |
| init_method_std | 2 0.0052 |
| init_method_xavier_uniform | |
| initial_loss_scale | False 65536 |
| layernorm_epsilon | 1E-05 |
| learnable_rotary_embedding | False |
| length_per_sample | 2000 |
| log_interval | 1 |
| loss_scale | 0 |
| loss_scale_window | 2000 |
| lr | 8e-05 |
| lr_decay_iters | None |
| lr_decay_samples | 197753905 |
| lr_decay_style | cosine |
| lr_warmup_samples | 1098632 |
| make_vocab_size_divisible_by | 768 |
| mask_prob | 0.15 |
| masked_softmax_fusion | True |
| micro_batch_size | 1 |
| min_gmask_ratio | 0.2 |
| min_loss_scale | 1.0 |
| min_lr | 8e-06 |
| multitask_ratio | |
| num_attention_heads | 0.05 |
| num_layers | 96 70 |
| onnx_safe | None |
| optimizer | |
| partition_activations | adam |
| True |
| pipeline_model_parallel_size position_embedding_type | 8 |
| rampup_batch_size | rotary 192, 24,5493164 |
| save_interval | 250 |
| seed | 1234 |
| seq_length | 2048 |
| short_seq_prob | 0.02 |
| shrink_embedding_gradient_alpha | 0.1 |
| single_span_prob | 0.02 |
| split tensor_model_parallel_size | 949,50,1 |
| tokenizer_type | 4 |
| weight_decay | IceTokenizer 0.1 |
| zero_contigious_gradients | |
| zero_reduce_bucket_size | False |
| zero_reduce_scatter | 500000000 |
| zero_stage | False |
| |
| zero-optimization.allgather_bucket_size | 500000000 |
| tokenizer_type | IceTokenizer |
| weight_decay | 0.1 |
| world_size | 768 |
| zero_contigious_gradients | FALSE |
| |
| zero_reduce_bucket_size | 500000000 |
| zero_reduce_scatter zero_stage | FALSE 1 |
",
"type": "table",
"image_path": "f84fd01d5a5fed9f1de50406b3ece3f64d4dabe76bb46944ba23ab2555f60a2c.jpg"
}
]
}
],
"index": 20,
"virtual_lines": [
{
"bbox": [
210,
79,
400,
96.02564102564102
],
"spans": [],
"index": 1
},
{
"bbox": [
210,
96.02564102564102,
400,
113.05128205128204
],
"spans": [],
"index": 2
},
{
"bbox": [
210,
113.05128205128204,
400,
130.07692307692307
],
"spans": [],
"index": 3
},
{
"bbox": [
210,
130.07692307692307,
400,
147.1025641025641
],
"spans": [],
"index": 4
},
{
"bbox": [
210,
147.1025641025641,
400,
164.1282051282051
],
"spans": [],
"index": 5
},
{
"bbox": [
210,
164.1282051282051,
400,
181.15384615384613
],
"spans": [],
"index": 6
},
{
"bbox": [
210,
181.15384615384613,
400,
198.17948717948715
],
"spans": [],
"index": 7
},
{
"bbox": [
210,
198.17948717948715,
400,
215.20512820512818
],
"spans": [],
"index": 8
},
{
"bbox": [
210,
215.20512820512818,
400,
232.2307692307692
],
"spans": [],
"index": 9
},
{
"bbox": [
210,
232.2307692307692,
400,
249.25641025641022
],
"spans": [],
"index": 10
},
{
"bbox": [
210,
249.25641025641022,
400,
266.28205128205127
],
"spans": [],
"index": 11
},
{
"bbox": [
210,
266.28205128205127,
400,
283.3076923076923
],
"spans": [],
"index": 12
},
{
"bbox": [
210,
283.3076923076923,
400,
300.33333333333337
],
"spans": [],
"index": 13
},
{
"bbox": [
210,
300.33333333333337,
400,
317.3589743589744
],
"spans": [],
"index": 14
},
{
"bbox": [
210,
317.3589743589744,
400,
334.3846153846155
],
"spans": [],
"index": 15
},
{
"bbox": [
210,
334.3846153846155,
400,
351.4102564102565
],
"spans": [],
"index": 16
},
{
"bbox": [
210,
351.4102564102565,
400,
368.4358974358976
],
"spans": [],
"index": 17
},
{
"bbox": [
210,
368.4358974358976,
400,
385.4615384615386
],
"spans": [],
"index": 18
},
{
"bbox": [
210,
385.4615384615386,
400,
402.4871794871797
],
"spans": [],
"index": 19
},
{
"bbox": [
210,
402.4871794871797,
400,
419.5128205128207
],
"spans": [],
"index": 20
},
{
"bbox": [
210,
419.5128205128207,
400,
436.5384615384618
],
"spans": [],
"index": 21
},
{
"bbox": [
210,
436.5384615384618,
400,
453.5641025641028
],
"spans": [],
"index": 22
},
{
"bbox": [
210,
453.5641025641028,
400,
470.5897435897439
],
"spans": [],
"index": 23
},
{
"bbox": [
210,
470.5897435897439,
400,
487.6153846153849
],
"spans": [],
"index": 24
},
{
"bbox": [
210,
487.6153846153849,
400,
504.641025641026
],
"spans": [],
"index": 25
},
{
"bbox": [
210,
504.641025641026,
400,
521.666666666667
],
"spans": [],
"index": 26
},
{
"bbox": [
210,
521.666666666667,
400,
538.692307692308
],
"spans": [],
"index": 27
},
{
"bbox": [
210,
538.692307692308,
400,
555.717948717949
],
"spans": [],
"index": 28
},
{
"bbox": [
210,
555.717948717949,
400,
572.74358974359
],
"spans": [],
"index": 29
},
{
"bbox": [
210,
572.74358974359,
400,
589.769230769231
],
"spans": [],
"index": 30
},
{
"bbox": [
210,
589.769230769231,
400,
606.7948717948719
],
"spans": [],
"index": 31
},
{
"bbox": [
210,
606.7948717948719,
400,
623.8205128205129
],
"spans": [],
"index": 32
},
{
"bbox": [
210,
623.8205128205129,
400,
640.8461538461539
],
"spans": [],
"index": 33
},
{
"bbox": [
210,
640.8461538461539,
400,
657.8717948717949
],
"spans": [],
"index": 34
},
{
"bbox": [
210,
657.8717948717949,
400,
674.8974358974359
],
"spans": [],
"index": 35
},
{
"bbox": [
210,
674.8974358974359,
400,
691.9230769230769
],
"spans": [],
"index": 36
},
{
"bbox": [
210,
691.9230769230769,
400,
708.9487179487179
],
"spans": [],
"index": 37
},
{
"bbox": [
210,
708.9487179487179,
400,
725.9743589743589
],
"spans": [],
"index": 38
},
{
"bbox": [
210,
725.9743589743589,
400,
742.9999999999999
],
"spans": [],
"index": 39
}
]
}
],
"index": 10.0
}
],
"page_idx": 47,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
293,
37
],
"lines": [
{
"bbox": [
106,
26,
293,
38
],
"spans": [
{
"bbox": [
106,
26,
293,
38
],
"score": 1.0,
"content": "Published as a conference paper at ICLR 2023",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
752,
311,
760
],
"lines": [
{
"bbox": [
298,
750,
312,
763
],
"spans": [
{
"bbox": [
298,
750,
312,
763
],
"score": 1.0,
"content": "",
"type": "text",
"height": 13,
"width": 14
}
]
}
]
}
],
"para_blocks": [
{
"type": "table",
"bbox": [
210,
79,
400,
743
],
"blocks": [
{
"type": "table_caption",
"bbox": [
198,
60,
413,
72
],
"group_id": 0,
"lines": [
{
"bbox": [
196,
57,
415,
74
],
"spans": [
{
"bbox": [
196,
57,
415,
74
],
"score": 1.0,
"content": "Table 11: Full configurations for GLM-130B training",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "table_body",
"bbox": [
210,
79,
400,
743
],
"group_id": 0,
"lines": [
{
"bbox": [
210,
79,
400,
743
],
"spans": [
{
"bbox": [
210,
79,
400,
743
],
"score": 0.954,
"html": "| Configuration Key | Value |
| adam_beta1 | 0.9 |
| adam_beta2 | 0.95 |
| adam_eps | 1e-08 |
| aggregated_samples_per_sequence | 4 |
| 0.1 |
| attention_dropout attention_softmax_in_fp32 | |
| True |
| average_block_length | 3 |
| bias_dropout_fusion | True True |
| checkpoint_activations | |
| checkpoint_in_cpu | False |
| checkpoint_num_layers | |
| clip_grad contigious_checkpointing | 1.0 |
| False |
| cpu_optimizer | False |
| data_parallel_size | 24 |
| deepnorm | True |
| distributed_backend | nccl |
| eval_interval | 1000 |
| eval_iters ffn_hidden_size | 3 |
| fp16 | 32768 |
| True |
| global_batch_size glu_activation | 4224 |
| gpt_prob | geglu |
| hidden_dropout | 0.7 |
| hidden_size | 0.1 |
| hysteresis | 12288 |
| init_method_std | 2 0.0052 |
| init_method_xavier_uniform | |
| initial_loss_scale | False 65536 |
| layernorm_epsilon | 1E-05 |
| learnable_rotary_embedding | False |
| length_per_sample | 2000 |
| log_interval | 1 |
| loss_scale | 0 |
| loss_scale_window | 2000 |
| lr | 8e-05 |
| lr_decay_iters | None |
| lr_decay_samples | 197753905 |
| lr_decay_style | cosine |
| lr_warmup_samples | 1098632 |
| make_vocab_size_divisible_by | 768 |
| mask_prob | 0.15 |
| masked_softmax_fusion | True |
| micro_batch_size | 1 |
| min_gmask_ratio | 0.2 |
| min_loss_scale | 1.0 |
| min_lr | 8e-06 |
| multitask_ratio | |
| num_attention_heads | 0.05 |
| num_layers | 96 70 |
| onnx_safe | None |
| optimizer | |
| partition_activations | adam |
| True |
| pipeline_model_parallel_size position_embedding_type | 8 |
| rampup_batch_size | rotary 192, 24,5493164 |
| save_interval | 250 |
| seed | 1234 |
| seq_length | 2048 |
| short_seq_prob | 0.02 |
| shrink_embedding_gradient_alpha | 0.1 |
| single_span_prob | 0.02 |
| split tensor_model_parallel_size | 949,50,1 |
| tokenizer_type | 4 |
| weight_decay | IceTokenizer 0.1 |
| zero_contigious_gradients | |
| zero_reduce_bucket_size | False |
| zero_reduce_scatter | 500000000 |
| zero_stage | False |
| |
| zero-optimization.allgather_bucket_size | 500000000 |
| tokenizer_type | IceTokenizer |
| weight_decay | 0.1 |
| world_size | 768 |
| zero_contigious_gradients | FALSE |
| |
| zero_reduce_bucket_size | 500000000 |
| zero_reduce_scatter zero_stage | FALSE 1 |
",
"type": "table",
"image_path": "f84fd01d5a5fed9f1de50406b3ece3f64d4dabe76bb46944ba23ab2555f60a2c.jpg"
}
]
}
],
"index": 20,
"virtual_lines": [
{
"bbox": [
210,
79,
400,
96.02564102564102
],
"spans": [],
"index": 1
},
{
"bbox": [
210,
96.02564102564102,
400,
113.05128205128204
],
"spans": [],
"index": 2
},
{
"bbox": [
210,
113.05128205128204,
400,
130.07692307692307
],
"spans": [],
"index": 3
},
{
"bbox": [
210,
130.07692307692307,
400,
147.1025641025641
],
"spans": [],
"index": 4
},
{
"bbox": [
210,
147.1025641025641,
400,
164.1282051282051
],
"spans": [],
"index": 5
},
{
"bbox": [
210,
164.1282051282051,
400,
181.15384615384613
],
"spans": [],
"index": 6
},
{
"bbox": [
210,
181.15384615384613,
400,
198.17948717948715
],
"spans": [],
"index": 7
},
{
"bbox": [
210,
198.17948717948715,
400,
215.20512820512818
],
"spans": [],
"index": 8
},
{
"bbox": [
210,
215.20512820512818,
400,
232.2307692307692
],
"spans": [],
"index": 9
},
{
"bbox": [
210,
232.2307692307692,
400,
249.25641025641022
],
"spans": [],
"index": 10
},
{
"bbox": [
210,
249.25641025641022,
400,
266.28205128205127
],
"spans": [],
"index": 11
},
{
"bbox": [
210,
266.28205128205127,
400,
283.3076923076923
],
"spans": [],
"index": 12
},
{
"bbox": [
210,
283.3076923076923,
400,
300.33333333333337
],
"spans": [],
"index": 13
},
{
"bbox": [
210,
300.33333333333337,
400,
317.3589743589744
],
"spans": [],
"index": 14
},
{
"bbox": [
210,
317.3589743589744,
400,
334.3846153846155
],
"spans": [],
"index": 15
},
{
"bbox": [
210,
334.3846153846155,
400,
351.4102564102565
],
"spans": [],
"index": 16
},
{
"bbox": [
210,
351.4102564102565,
400,
368.4358974358976
],
"spans": [],
"index": 17
},
{
"bbox": [
210,
368.4358974358976,
400,
385.4615384615386
],
"spans": [],
"index": 18
},
{
"bbox": [
210,
385.4615384615386,
400,
402.4871794871797
],
"spans": [],
"index": 19
},
{
"bbox": [
210,
402.4871794871797,
400,
419.5128205128207
],
"spans": [],
"index": 20
},
{
"bbox": [
210,
419.5128205128207,
400,
436.5384615384618
],
"spans": [],
"index": 21
},
{
"bbox": [
210,
436.5384615384618,
400,
453.5641025641028
],
"spans": [],
"index": 22
},
{
"bbox": [
210,
453.5641025641028,
400,
470.5897435897439
],
"spans": [],
"index": 23
},
{
"bbox": [
210,
470.5897435897439,
400,
487.6153846153849
],
"spans": [],
"index": 24
},
{
"bbox": [
210,
487.6153846153849,
400,
504.641025641026
],
"spans": [],
"index": 25
},
{
"bbox": [
210,
504.641025641026,
400,
521.666666666667
],
"spans": [],
"index": 26
},
{
"bbox": [
210,
521.666666666667,
400,
538.692307692308
],
"spans": [],
"index": 27
},
{
"bbox": [
210,
538.692307692308,
400,
555.717948717949
],
"spans": [],
"index": 28
},
{
"bbox": [
210,
555.717948717949,
400,
572.74358974359
],
"spans": [],
"index": 29
},
{
"bbox": [
210,
572.74358974359,
400,
589.769230769231
],
"spans": [],
"index": 30
},
{
"bbox": [
210,
589.769230769231,
400,
606.7948717948719
],
"spans": [],
"index": 31
},
{
"bbox": [
210,
606.7948717948719,
400,
623.8205128205129
],
"spans": [],
"index": 32
},
{
"bbox": [
210,
623.8205128205129,
400,
640.8461538461539
],
"spans": [],
"index": 33
},
{
"bbox": [
210,
640.8461538461539,
400,
657.8717948717949
],
"spans": [],
"index": 34
},
{
"bbox": [
210,
657.8717948717949,
400,
674.8974358974359
],
"spans": [],
"index": 35
},
{
"bbox": [
210,
674.8974358974359,
400,
691.9230769230769
],
"spans": [],
"index": 36
},
{
"bbox": [
210,
691.9230769230769,
400,
708.9487179487179
],
"spans": [],
"index": 37
},
{
"bbox": [
210,
708.9487179487179,
400,
725.9743589743589
],
"spans": [],
"index": 38
},
{
"bbox": [
210,
725.9743589743589,
400,
742.9999999999999
],
"spans": [],
"index": 39
}
]
}
],
"index": 10.0
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
107,
207,
509,
658
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
164,
505,
198
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
163,
504,
176
],
"spans": [
{
"bbox": [
105,
163,
504,
176
],
"score": 1.0,
"content": "Table 12: The 74 datasets involved in Multi-task Instruction Pre-training (MIP). Datasets from T0-",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
175,
505,
187
],
"spans": [
{
"bbox": [
106,
175,
505,
187
],
"score": 1.0,
"content": "PromptSource (Sanh et al., 2022; Bach et al., 2022) are named in their Hugging Face datasets iden-",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
186,
449,
198
],
"spans": [
{
"bbox": [
106,
186,
449,
198
],
"score": 1.0,
"content": "tifiers. Datasets from DeepStruct (Wang et al., 2022a) are described in Appendix C.2.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
107,
207,
509,
658
],
"group_id": 0,
"lines": [
{
"bbox": [
107,
207,
509,
658
],
"spans": [
{
"bbox": [
107,
207,
509,
658
],
"score": 0.983,
"html": "| Discipline | GLM-130B BLOOM176B |
| STEM |
| abstract_algebra anatomy | 24.00 48.90 | 24.00 38.52 |
| astronomy | 48.03 | 34.87 |
| | 37.50 |
| colledge_biology | 47.22 | |
| college_chemistry | 34.00 | 19.00 |
| colledge_computer_science | 44.00 | 1.00 |
| colledge_mathematcis | 27.00 | 31.00 |
| colledge_physics | 30.39 61.00 | 24.50 40.00 |
| computer_security conceptual_physics | 38.72 | 31.49 |
| electrical_engineering | 45.52 | 32.41 |
| elementary_mathematics | 31.75 | 29.63 |
| high_school_biology | 51.29 | 27.42 |
| high_school_chemistry | 34.98 | 27.09 |
| high_school_computer_science | 53.00 | 30.00 |
| high_school_mathematics | 28.15 | 25.93 |
| high_school_physics | 29.80 | 30.46 |
| high_school_statistics machine_learning | 38.43 40.18 | 26.39 29.46 |
| Social Science | | 26.32 | 26.32 |
| econometrics | | |
| high_school_geography | 53.54 | 36.36 |
| high_school_government_and_politics | 62.18 | 40.41 |
| high_school_macroeconomics | 42.56 | 30.77 |
| high_school_microeconomics | 45.80 | 26.89 |
| high_school_psychology | 54.13 51.15 | 39.27 |
| human_sexuality | 42.48 | 35.11 31.54 |
| professional_psychology | 55.46 | 33.64 |
| public_relations security_studies | 44.90 | 34.29 |
| sociology | | 51.74 | 31.84 |
| Humanities | us_foreign_policy | 61.00 | 46.00 |
| formal_logic | 27.78 | 23.02 |
| high_school_european_history | 58.18 | 35.76 |
| high_school_us_history | 58.33 | 40.69 |
| high_school_world_history | 67.09 | 32.07 |
| international_law | 56.20 | 42.15 |
| jurisprudence | 43.52 | 35.19 |
| logical_fallacies | 57.06 | 31.29 |
| moral_disputes | 47.11 | 36.71 |
| moral_scenarios philosophy | 24.25 | 24.36 |
| prehistory | | | 35.37 |
| | 45.34 50.93 | 40.43 |
| professional_law world_religions | 37.94 | 29.53 |
| 55.56 | 42.11 |
| business_ethics clinical_knowledge | 51.00 | 34.00 |
| colledge_medicine | 48.68 | 35.85 28.90 |
| glocal_facts | 43.35 35.00 | |
| human_aging Other | | | 23.00 |
| management | 45.29 56.31 | 32.29 |
| marketing | 67.52 | 27.18 |
| | 39.74 |
| medical_genetics | 48.00 | 45.00 |
| miscellaneous | 61.18 | 40.23 |
| nutrition | 50.65 | 32.35 |
| professional_accounting | 35.46 | 28.72 |
| professional_medicine | 43.38 | 18.01 |
| | |
| virology | 39.16 | 28.31 |
",
"type": "table",
"image_path": "263ed91e2f5136ee90d9f0fb44bda0e81da65c26b9506dedb880140ace88b2e4.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
142,
116,
469,
320.33333333333337
],
"spans": [],
"index": 3
},
{
"bbox": [
142,
320.33333333333337,
469,
524.6666666666667
],
"spans": [],
"index": 4
},
{
"bbox": [
142,
524.6666666666667,
469,
729.0000000000001
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
}
],
"page_idx": 50,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
293,
37
],
"lines": [
{
"bbox": [
106,
26,
293,
38
],
"spans": [
{
"bbox": [
106,
26,
293,
38
],
"score": 1.0,
"content": "Published as a conference paper at ICLR 2023",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
310,
761
],
"lines": [
{
"bbox": [
298,
750,
312,
764
],
"spans": [
{
"bbox": [
298,
750,
312,
764
],
"score": 1.0,
"content": "51",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "table",
"bbox": [
142,
116,
469,
729
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
79,
506,
114
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
78,
504,
93
],
"spans": [
{
"bbox": [
105,
78,
504,
93
],
"score": 1.0,
"content": "Table 15: Detailed results of GLM-130B and BLOOM 176B (Scao et al., 2022) on",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
91,
506,
104
],
"spans": [
{
"bbox": [
106,
91,
506,
104
],
"score": 1.0,
"content": "MMLU (Hendrycks et al., 2021). We find that no existing literature has reported GPT-3 175B’s",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
102,
474,
115
],
"spans": [
{
"bbox": [
106,
102,
474,
115
],
"score": 1.0,
"content": "numerical accuracy. BLOOM is evaluated using Huggingface Transformer implementation.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
142,
116,
469,
729
],
"group_id": 0,
"lines": [
{
"bbox": [
142,
116,
469,
729
],
"spans": [
{
"bbox": [
142,
116,
469,
729
],
"score": 0.973,
"html": "| Discipline | GLM-130B BLOOM176B |
| STEM |
| abstract_algebra anatomy | 24.00 48.90 | 24.00 38.52 |
| astronomy | 48.03 | 34.87 |
| | 37.50 |
| colledge_biology | 47.22 | |
| college_chemistry | 34.00 | 19.00 |
| colledge_computer_science | 44.00 | 1.00 |
| colledge_mathematcis | 27.00 | 31.00 |
| colledge_physics | 30.39 61.00 | 24.50 40.00 |
| computer_security conceptual_physics | 38.72 | 31.49 |
| electrical_engineering | 45.52 | 32.41 |
| elementary_mathematics | 31.75 | 29.63 |
| high_school_biology | 51.29 | 27.42 |
| high_school_chemistry | 34.98 | 27.09 |
| high_school_computer_science | 53.00 | 30.00 |
| high_school_mathematics | 28.15 | 25.93 |
| high_school_physics | 29.80 | 30.46 |
| high_school_statistics machine_learning | 38.43 40.18 | 26.39 29.46 |
| Social Science | | 26.32 | 26.32 |
| econometrics | | |
| high_school_geography | 53.54 | 36.36 |
| high_school_government_and_politics | 62.18 | 40.41 |
| high_school_macroeconomics | 42.56 | 30.77 |
| high_school_microeconomics | 45.80 | 26.89 |
| high_school_psychology | 54.13 51.15 | 39.27 |
| human_sexuality | 42.48 | 35.11 31.54 |
| professional_psychology | 55.46 | 33.64 |
| public_relations security_studies | 44.90 | 34.29 |
| sociology | | 51.74 | 31.84 |
| Humanities | us_foreign_policy | 61.00 | 46.00 |
| formal_logic | 27.78 | 23.02 |
| high_school_european_history | 58.18 | 35.76 |
| high_school_us_history | 58.33 | 40.69 |
| high_school_world_history | 67.09 | 32.07 |
| international_law | 56.20 | 42.15 |
| jurisprudence | 43.52 | 35.19 |
| logical_fallacies | 57.06 | 31.29 |
| moral_disputes | 47.11 | 36.71 |
| moral_scenarios philosophy | 24.25 | 24.36 |
| prehistory | | | 35.37 |
| | 45.34 50.93 | 40.43 |
| professional_law world_religions | 37.94 | 29.53 |
| 55.56 | 42.11 |
| business_ethics clinical_knowledge | 51.00 | 34.00 |
| colledge_medicine | 48.68 | 35.85 28.90 |
| glocal_facts | 43.35 35.00 | |
| human_aging Other | | | 23.00 |
| management | 45.29 56.31 | 32.29 |
| marketing | 67.52 | 27.18 |
| | 39.74 |
| medical_genetics | 48.00 | 45.00 |
| miscellaneous | 61.18 | 40.23 |
| nutrition | 50.65 | 32.35 |
| professional_accounting | 35.46 | 28.72 |
| professional_medicine | 43.38 | 18.01 |
| | |
| virology | 39.16 | 28.31 |
",
"type": "table",
"image_path": "263ed91e2f5136ee90d9f0fb44bda0e81da65c26b9506dedb880140ace88b2e4.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
142,
116,
469,
320.33333333333337
],
"spans": [],
"index": 3
},
{
"bbox": [
142,
320.33333333333337,
469,
524.6666666666667
],
"spans": [],
"index": 4
},
{
"bbox": [
142,
524.6666666666667,
469,
729.0000000000001
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
108,
81,
214,
93
],
"lines": [
{
"bbox": [
104,
79,
216,
96
],
"spans": [
{
"bbox": [
104,
79,
216,
96
],
"score": 1.0,
"content": "E CONTRIBUTIONS",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
105,
505,
183
],
"lines": [
{
"bbox": [
106,
106,
505,
118
],
"spans": [
{
"bbox": [
106,
106,
505,
118
],
"score": 1.0,
"content": "The GLM-130B project was conceived in Dec. 2021 with its pre-training part completed in July 3rd,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
116,
506,
129
],
"spans": [
{
"bbox": [
105,
116,
506,
129
],
"score": 1.0,
"content": "2022 and its evaluation and applications still ongoing. Over the course, we have experienced various",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
128,
505,
140
],
"spans": [
{
"bbox": [
106,
128,
505,
140
],
"score": 1.0,
"content": "technical and engineering challenges (Cf. Appendix F and Figure 21 for details). It would not be",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
104,
138,
506,
152
],
"spans": [
{
"bbox": [
104,
138,
506,
152
],
"score": 1.0,
"content": "possible to reach its current status if without the collaboration of multiple teams—the Knowledge",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
149,
505,
163
],
"spans": [
{
"bbox": [
105,
149,
505,
163
],
"score": 1.0,
"content": "Engineering Group (KEG), Parallel Architecture & Compiler technology of Mobile, Accelerated,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
160,
506,
174
],
"spans": [
{
"bbox": [
105,
160,
506,
174
],
"score": 1.0,
"content": "and Networked systems Group (PACMAN), and Natural Language Processing Group (THUNLP) at",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
172,
450,
184
],
"spans": [
{
"bbox": [
106,
172,
450,
184
],
"score": 1.0,
"content": "Tsinghua University, as well as Zhipu.AI. The detailed contributions are listed below.",
"type": "text"
}
],
"index": 7
}
],
"index": 4
},
{
"type": "title",
"bbox": [
107,
196,
192,
207
],
"lines": [
{
"bbox": [
105,
195,
194,
209
],
"spans": [
{
"bbox": [
105,
195,
194,
209
],
"score": 1.0,
"content": "E.1 PREPARATION",
"type": "text"
}
],
"index": 8
}
],
"index": 8
},
{
"type": "text",
"bbox": [
106,
216,
482,
303
],
"lines": [
{
"bbox": [
105,
216,
333,
230
],
"spans": [
{
"bbox": [
105,
216,
333,
230
],
"score": 1.0,
"content": "• Model Implementation: Aohan Zeng, Zhengxiao Du",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
229,
365,
246
],
"spans": [
{
"bbox": [
105,
229,
365,
246
],
"score": 1.0,
"content": "• Self-Supervised Data Processing: Ming Ding, Wendi Zheng",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
246,
313,
259
],
"spans": [
{
"bbox": [
106,
246,
313,
259
],
"score": 1.0,
"content": "• Multitask Data Processing: Xiao Liu, Xiao Xia",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
261,
406,
274
],
"spans": [
{
"bbox": [
106,
261,
406,
274
],
"score": 1.0,
"content": "• Model Architecture: Aohan Zeng, Xiao Liu, Zhengxiao Du, Hanyu Lai",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
275,
339,
290
],
"spans": [
{
"bbox": [
105,
275,
339,
290
],
"score": 1.0,
"content": "• Training Stability: Aohan Zeng, Xiao Liu, Ming Ding",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
290,
481,
304
],
"spans": [
{
"bbox": [
105,
290,
481,
304
],
"score": 1.0,
"content": "• 3D-Parallelism and Training Efficiency: Aohan Zeng, Zixuan Ma, Jiaao He, Zhenbo Sun",
"type": "text"
}
],
"index": 14
}
],
"index": 11.5
},
{
"type": "text",
"bbox": [
107,
316,
212,
327
],
"lines": [
{
"bbox": [
106,
315,
212,
329
],
"spans": [
{
"bbox": [
106,
315,
212,
329
],
"score": 1.0,
"content": "E.2 MODEL TRAINING",
"type": "text"
}
],
"index": 15
}
],
"index": 15
},
{
"type": "text",
"bbox": [
106,
336,
365,
363
],
"lines": [
{
"bbox": [
106,
335,
366,
349
],
"spans": [
{
"bbox": [
106,
335,
366,
349
],
"score": 1.0,
"content": "• Large-Scale Training & Monitoring: Aohan Zeng, Xiao Liu",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
350,
304,
364
],
"spans": [
{
"bbox": [
105,
350,
304,
364
],
"score": 1.0,
"content": "• Model Performance Validation: Aohan Zeng",
"type": "text"
}
],
"index": 17
}
],
"index": 16.5
},
{
"type": "text",
"bbox": [
107,
376,
202,
387
],
"lines": [
{
"bbox": [
106,
375,
203,
388
],
"spans": [
{
"bbox": [
106,
375,
203,
388
],
"score": 1.0,
"content": "E.3 POST TRAINING",
"type": "text"
}
],
"index": 18
}
],
"index": 18
},
{
"type": "text",
"bbox": [
106,
396,
505,
579
],
"lines": [
{
"bbox": [
105,
395,
334,
409
],
"spans": [
{
"bbox": [
105,
395,
334,
409
],
"score": 1.0,
"content": "• Evaluation Framework: Aohan Zeng, Zhengxiao Du",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
410,
307,
424
],
"spans": [
{
"bbox": [
105,
410,
307,
424
],
"score": 1.0,
"content": "• Language Modeling Evaluation: Aohan Zeng",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
424,
317,
439
],
"spans": [
{
"bbox": [
105,
424,
317,
439
],
"score": 1.0,
"content": "• MMLU & BIG-Bench Evaluation: Aohan Zeng",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
104,
439,
349,
455
],
"spans": [
{
"bbox": [
104,
439,
349,
455
],
"score": 1.0,
"content": "• CLUE & FewCLUE Evaluation: Xiao Liu, Aohan Zeng",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
455,
389,
469
],
"spans": [
{
"bbox": [
105,
455,
389,
469
],
"score": 1.0,
"content": "• Ethical Evaluation: Yifan Xu, Aohan Zeng, Xiao Liu, Zihan Wang",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
470,
351,
483
],
"spans": [
{
"bbox": [
105,
470,
351,
483
],
"score": 1.0,
"content": "• Baseline Evaluation: Xiao Liu, Jifan Yu, Weng Lam Tam",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
485,
396,
498
],
"spans": [
{
"bbox": [
105,
485,
396,
498
],
"score": 1.0,
"content": "• INT4 Quantization: Aohan Zeng, Zihan Wang, Xiao Liu, Hanyu Lai",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
499,
324,
514
],
"spans": [
{
"bbox": [
105,
499,
324,
514
],
"score": 1.0,
"content": "• Inference Acceleration: Zihan Wang, Aohan Zeng",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
515,
432,
528
],
"spans": [
{
"bbox": [
105,
515,
432,
528
],
"score": 1.0,
"content": "• Low-Resource Inference: Gouyang Zeng, Xu Han, Weilin Zhao, Zhiyuan Liu",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
530,
504,
544
],
"spans": [
{
"bbox": [
105,
530,
504,
544
],
"score": 1.0,
"content": "• Demo and API: Hanyu Lai, Jifan Yu, Xiaohan Zhang, Yufei Xue, Shan Wang, Jiecai Shan, Hao-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
113,
542,
220,
555
],
"spans": [
{
"bbox": [
113,
542,
220,
555
],
"score": 1.0,
"content": "han Jiang, Zhengang Guo",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
555,
506,
570
],
"spans": [
{
"bbox": [
105,
555,
506,
570
],
"score": 1.0,
"content": "• Manuscript Writing: Xiao Liu, Yuxiao Dong, and Jie Tang wrote the main paper, and Xiao Liu,",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
114,
568,
330,
581
],
"spans": [
{
"bbox": [
114,
568,
330,
581
],
"score": 1.0,
"content": "Aohan Zeng, and Zhengxiao Du wrote the Appendix.",
"type": "text"
}
],
"index": 31
}
],
"index": 25
},
{
"type": "title",
"bbox": [
107,
592,
237,
603
],
"lines": [
{
"bbox": [
106,
591,
238,
604
],
"spans": [
{
"bbox": [
106,
591,
238,
604
],
"score": 1.0,
"content": "E.4 PROJECT MANAGEMENT",
"type": "text"
}
],
"index": 32
}
],
"index": 32
},
{
"type": "text",
"bbox": [
106,
613,
506,
665
],
"lines": [
{
"bbox": [
105,
611,
284,
624
],
"spans": [
{
"bbox": [
105,
611,
284,
624
],
"score": 1.0,
"content": "• Student Leaders: Aohan Zeng, Xiao Liu",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
104,
624,
506,
642
],
"spans": [
{
"bbox": [
104,
624,
506,
642
],
"score": 1.0,
"content": "• Technical Advisors: Yuxiao Dong, Jidong Zhai, Wenguang Chen, Zhiyuan Liu, Peng Zhang, Jie",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
111,
636,
139,
653
],
"spans": [
{
"bbox": [
111,
636,
139,
653
],
"score": 1.0,
"content": "Tang",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
651,
222,
668
],
"spans": [
{
"bbox": [
105,
651,
222,
668
],
"score": 1.0,
"content": "• Project Leader: Jie Tang",
"type": "text"
}
],
"index": 36
}
],
"index": 34.5
},
{
"type": "title",
"bbox": [
107,
678,
239,
689
],
"lines": [
{
"bbox": [
105,
677,
240,
691
],
"spans": [
{
"bbox": [
105,
677,
240,
691
],
"score": 1.0,
"content": "E.5 COMPUTATION SPONSOR",
"type": "text"
}
],
"index": 37
}
],
"index": 37
},
{
"type": "text",
"bbox": [
106,
698,
218,
711
],
"lines": [
{
"bbox": [
105,
696,
220,
713
],
"spans": [
{
"bbox": [
105,
696,
220,
713
],
"score": 1.0,
"content": "• GPU Sponsor: Zhipu.AI",
"type": "text"
}
],
"index": 38
}
],
"index": 38
}
],
"page_idx": 51,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
293,
37
],
"lines": [
{
"bbox": [
106,
26,
293,
38
],
"spans": [
{
"bbox": [
106,
26,
293,
38
],
"score": 1.0,
"content": "Published as a conference paper at ICLR 2023",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
298,
749,
313,
764
],
"spans": [
{
"bbox": [
298,
749,
313,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 15,
"width": 15
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
108,
81,
214,
93
],
"lines": [
{
"bbox": [
104,
79,
216,
96
],
"spans": [
{
"bbox": [
104,
79,
216,
96
],
"score": 1.0,
"content": "E CONTRIBUTIONS",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
105,
505,
183
],
"lines": [
{
"bbox": [
106,
106,
505,
118
],
"spans": [
{
"bbox": [
106,
106,
505,
118
],
"score": 1.0,
"content": "The GLM-130B project was conceived in Dec. 2021 with its pre-training part completed in July 3rd,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
116,
506,
129
],
"spans": [
{
"bbox": [
105,
116,
506,
129
],
"score": 1.0,
"content": "2022 and its evaluation and applications still ongoing. Over the course, we have experienced various",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
128,
505,
140
],
"spans": [
{
"bbox": [
106,
128,
505,
140
],
"score": 1.0,
"content": "technical and engineering challenges (Cf. Appendix F and Figure 21 for details). It would not be",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
104,
138,
506,
152
],
"spans": [
{
"bbox": [
104,
138,
506,
152
],
"score": 1.0,
"content": "possible to reach its current status if without the collaboration of multiple teams—the Knowledge",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
149,
505,
163
],
"spans": [
{
"bbox": [
105,
149,
505,
163
],
"score": 1.0,
"content": "Engineering Group (KEG), Parallel Architecture & Compiler technology of Mobile, Accelerated,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
160,
506,
174
],
"spans": [
{
"bbox": [
105,
160,
506,
174
],
"score": 1.0,
"content": "and Networked systems Group (PACMAN), and Natural Language Processing Group (THUNLP) at",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
172,
450,
184
],
"spans": [
{
"bbox": [
106,
172,
450,
184
],
"score": 1.0,
"content": "Tsinghua University, as well as Zhipu.AI. The detailed contributions are listed below.",
"type": "text"
}
],
"index": 7
}
],
"index": 4,
"bbox_fs": [
104,
106,
506,
184
]
},
{
"type": "title",
"bbox": [
107,
196,
192,
207
],
"lines": [
{
"bbox": [
105,
195,
194,
209
],
"spans": [
{
"bbox": [
105,
195,
194,
209
],
"score": 1.0,
"content": "E.1 PREPARATION",
"type": "text"
}
],
"index": 8
}
],
"index": 8
},
{
"type": "list",
"bbox": [
106,
216,
482,
303
],
"lines": [
{
"bbox": [
105,
216,
333,
230
],
"spans": [
{
"bbox": [
105,
216,
333,
230
],
"score": 1.0,
"content": "• Model Implementation: Aohan Zeng, Zhengxiao Du",
"type": "text"
}
],
"index": 9,
"is_list_start_line": true
},
{
"bbox": [
105,
229,
365,
246
],
"spans": [
{
"bbox": [
105,
229,
365,
246
],
"score": 1.0,
"content": "• Self-Supervised Data Processing: Ming Ding, Wendi Zheng",
"type": "text"
}
],
"index": 10,
"is_list_start_line": true
},
{
"bbox": [
106,
246,
313,
259
],
"spans": [
{
"bbox": [
106,
246,
313,
259
],
"score": 1.0,
"content": "• Multitask Data Processing: Xiao Liu, Xiao Xia",
"type": "text"
}
],
"index": 11,
"is_list_start_line": true
},
{
"bbox": [
106,
261,
406,
274
],
"spans": [
{
"bbox": [
106,
261,
406,
274
],
"score": 1.0,
"content": "• Model Architecture: Aohan Zeng, Xiao Liu, Zhengxiao Du, Hanyu Lai",
"type": "text"
}
],
"index": 12,
"is_list_start_line": true
},
{
"bbox": [
105,
275,
339,
290
],
"spans": [
{
"bbox": [
105,
275,
339,
290
],
"score": 1.0,
"content": "• Training Stability: Aohan Zeng, Xiao Liu, Ming Ding",
"type": "text"
}
],
"index": 13,
"is_list_start_line": true
},
{
"bbox": [
105,
290,
481,
304
],
"spans": [
{
"bbox": [
105,
290,
481,
304
],
"score": 1.0,
"content": "• 3D-Parallelism and Training Efficiency: Aohan Zeng, Zixuan Ma, Jiaao He, Zhenbo Sun",
"type": "text"
}
],
"index": 14,
"is_list_start_line": true
}
],
"index": 11.5,
"bbox_fs": [
105,
216,
481,
304
]
},
{
"type": "text",
"bbox": [
107,
316,
212,
327
],
"lines": [
{
"bbox": [
106,
315,
212,
329
],
"spans": [
{
"bbox": [
106,
315,
212,
329
],
"score": 1.0,
"content": "E.2 MODEL TRAINING",
"type": "text"
}
],
"index": 15
}
],
"index": 15,
"bbox_fs": [
106,
315,
212,
329
]
},
{
"type": "text",
"bbox": [
106,
336,
365,
363
],
"lines": [
{
"bbox": [
106,
335,
366,
349
],
"spans": [
{
"bbox": [
106,
335,
366,
349
],
"score": 1.0,
"content": "• Large-Scale Training & Monitoring: Aohan Zeng, Xiao Liu",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
350,
304,
364
],
"spans": [
{
"bbox": [
105,
350,
304,
364
],
"score": 1.0,
"content": "• Model Performance Validation: Aohan Zeng",
"type": "text"
}
],
"index": 17
}
],
"index": 16.5,
"bbox_fs": [
105,
335,
366,
364
]
},
{
"type": "text",
"bbox": [
107,
376,
202,
387
],
"lines": [
{
"bbox": [
106,
375,
203,
388
],
"spans": [
{
"bbox": [
106,
375,
203,
388
],
"score": 1.0,
"content": "E.3 POST TRAINING",
"type": "text"
}
],
"index": 18
}
],
"index": 18,
"bbox_fs": [
106,
375,
203,
388
]
},
{
"type": "list",
"bbox": [
106,
396,
505,
579
],
"lines": [
{
"bbox": [
105,
395,
334,
409
],
"spans": [
{
"bbox": [
105,
395,
334,
409
],
"score": 1.0,
"content": "• Evaluation Framework: Aohan Zeng, Zhengxiao Du",
"type": "text"
}
],
"index": 19,
"is_list_end_line": true
},
{
"bbox": [
105,
410,
307,
424
],
"spans": [
{
"bbox": [
105,
410,
307,
424
],
"score": 1.0,
"content": "• Language Modeling Evaluation: Aohan Zeng",
"type": "text"
}
],
"index": 20,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
105,
424,
317,
439
],
"spans": [
{
"bbox": [
105,
424,
317,
439
],
"score": 1.0,
"content": "• MMLU & BIG-Bench Evaluation: Aohan Zeng",
"type": "text"
}
],
"index": 21,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
104,
439,
349,
455
],
"spans": [
{
"bbox": [
104,
439,
349,
455
],
"score": 1.0,
"content": "• CLUE & FewCLUE Evaluation: Xiao Liu, Aohan Zeng",
"type": "text"
}
],
"index": 22,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
105,
455,
389,
469
],
"spans": [
{
"bbox": [
105,
455,
389,
469
],
"score": 1.0,
"content": "• Ethical Evaluation: Yifan Xu, Aohan Zeng, Xiao Liu, Zihan Wang",
"type": "text"
}
],
"index": 23,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
105,
470,
351,
483
],
"spans": [
{
"bbox": [
105,
470,
351,
483
],
"score": 1.0,
"content": "• Baseline Evaluation: Xiao Liu, Jifan Yu, Weng Lam Tam",
"type": "text"
}
],
"index": 24,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
105,
485,
396,
498
],
"spans": [
{
"bbox": [
105,
485,
396,
498
],
"score": 1.0,
"content": "• INT4 Quantization: Aohan Zeng, Zihan Wang, Xiao Liu, Hanyu Lai",
"type": "text"
}
],
"index": 25,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
105,
499,
324,
514
],
"spans": [
{
"bbox": [
105,
499,
324,
514
],
"score": 1.0,
"content": "• Inference Acceleration: Zihan Wang, Aohan Zeng",
"type": "text"
}
],
"index": 26,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
105,
515,
432,
528
],
"spans": [
{
"bbox": [
105,
515,
432,
528
],
"score": 1.0,
"content": "• Low-Resource Inference: Gouyang Zeng, Xu Han, Weilin Zhao, Zhiyuan Liu",
"type": "text"
}
],
"index": 27,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
105,
530,
504,
544
],
"spans": [
{
"bbox": [
105,
530,
504,
544
],
"score": 1.0,
"content": "• Demo and API: Hanyu Lai, Jifan Yu, Xiaohan Zhang, Yufei Xue, Shan Wang, Jiecai Shan, Hao-",
"type": "text"
}
],
"index": 28,
"is_list_start_line": true
},
{
"bbox": [
113,
542,
220,
555
],
"spans": [
{
"bbox": [
113,
542,
220,
555
],
"score": 1.0,
"content": "han Jiang, Zhengang Guo",
"type": "text"
}
],
"index": 29,
"is_list_end_line": true
},
{
"bbox": [
105,
555,
506,
570
],
"spans": [
{
"bbox": [
105,
555,
506,
570
],
"score": 1.0,
"content": "• Manuscript Writing: Xiao Liu, Yuxiao Dong, and Jie Tang wrote the main paper, and Xiao Liu,",
"type": "text"
}
],
"index": 30,
"is_list_start_line": true
},
{
"bbox": [
114,
568,
330,
581
],
"spans": [
{
"bbox": [
114,
568,
330,
581
],
"score": 1.0,
"content": "Aohan Zeng, and Zhengxiao Du wrote the Appendix.",
"type": "text"
}
],
"index": 31,
"is_list_end_line": true
}
],
"index": 25,
"bbox_fs": [
104,
395,
506,
581
]
},
{
"type": "title",
"bbox": [
107,
592,
237,
603
],
"lines": [
{
"bbox": [
106,
591,
238,
604
],
"spans": [
{
"bbox": [
106,
591,
238,
604
],
"score": 1.0,
"content": "E.4 PROJECT MANAGEMENT",
"type": "text"
}
],
"index": 32
}
],
"index": 32
},
{
"type": "list",
"bbox": [
106,
613,
506,
665
],
"lines": [
{
"bbox": [
105,
611,
284,
624
],
"spans": [
{
"bbox": [
105,
611,
284,
624
],
"score": 1.0,
"content": "• Student Leaders: Aohan Zeng, Xiao Liu",
"type": "text"
}
],
"index": 33,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
104,
624,
506,
642
],
"spans": [
{
"bbox": [
104,
624,
506,
642
],
"score": 1.0,
"content": "• Technical Advisors: Yuxiao Dong, Jidong Zhai, Wenguang Chen, Zhiyuan Liu, Peng Zhang, Jie",
"type": "text"
}
],
"index": 34,
"is_list_start_line": true
},
{
"bbox": [
111,
636,
139,
653
],
"spans": [
{
"bbox": [
111,
636,
139,
653
],
"score": 1.0,
"content": "Tang",
"type": "text"
}
],
"index": 35,
"is_list_end_line": true
},
{
"bbox": [
105,
651,
222,
668
],
"spans": [
{
"bbox": [
105,
651,
222,
668
],
"score": 1.0,
"content": "• Project Leader: Jie Tang",
"type": "text"
}
],
"index": 36,
"is_list_start_line": true,
"is_list_end_line": true
}
],
"index": 34.5,
"bbox_fs": [
104,
611,
506,
668
]
},
{
"type": "title",
"bbox": [
107,
678,
239,
689
],
"lines": [
{
"bbox": [
105,
677,
240,
691
],
"spans": [
{
"bbox": [
105,
677,
240,
691
],
"score": 1.0,
"content": "E.5 COMPUTATION SPONSOR",
"type": "text"
}
],
"index": 37
}
],
"index": 37
},
{
"type": "text",
"bbox": [
106,
698,
218,
711
],
"lines": [
{
"bbox": [
105,
696,
220,
713
],
"spans": [
{
"bbox": [
105,
696,
220,
713
],
"score": 1.0,
"content": "• GPU Sponsor: Zhipu.AI",
"type": "text"
}
],
"index": 38
}
],
"index": 38,
"bbox_fs": [
105,
696,
220,
713
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
107,
81,
301,
94
],
"lines": [
{
"bbox": [
106,
81,
301,
95
],
"spans": [
{
"bbox": [
106,
81,
301,
95
],
"score": 1.0,
"content": "F A BRIEF HISTORY OF GLM-130B",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
108,
505,
208
],
"lines": [
{
"bbox": [
105,
106,
505,
122
],
"spans": [
{
"bbox": [
105,
106,
505,
122
],
"score": 1.0,
"content": "The GLM-130B project16 was conceived in Dec. 2021 in a brainstorming meeting at Tsinghua KEG.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
119,
505,
131
],
"spans": [
{
"bbox": [
106,
119,
505,
131
],
"score": 1.0,
"content": "We firmly believe that it is of value to pre-train a highly accurate language model, in particular for",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
131,
505,
142
],
"spans": [
{
"bbox": [
106,
131,
505,
142
],
"score": 1.0,
"content": "both Chinese and English. Though GPT-3 (Brown et al., 2020) is the pioneer for this effort, it is not",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
142,
505,
154
],
"spans": [
{
"bbox": [
105,
142,
505,
154
],
"score": 1.0,
"content": "available to most people in the world. In addition, it supports English only. We therefore decide to",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
153,
506,
164
],
"spans": [
{
"bbox": [
105,
153,
506,
164
],
"score": 1.0,
"content": "initialize the project GLM-130B. Please note that the WuDao 1.75T model we built last year is a",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
163,
505,
176
],
"spans": [
{
"bbox": [
105,
163,
505,
176
],
"score": 1.0,
"content": "sparse model with 480 mixture-of-experts (MoE), rather than a dense one as GPT-3. Our goal then",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
175,
505,
186
],
"spans": [
{
"bbox": [
105,
175,
505,
186
],
"score": 1.0,
"content": "is to train a bilingual pre-trained dense model with high accuracy on downstream tasks, and to make",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
185,
505,
198
],
"spans": [
{
"bbox": [
105,
185,
505,
198
],
"score": 1.0,
"content": "it open to everyone in the world-anyone, anywhere can download it and use it on a single server with",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
197,
183,
208
],
"spans": [
{
"bbox": [
105,
197,
183,
208
],
"score": 1.0,
"content": "appropriate GPUs.",
"type": "text"
}
],
"index": 9
}
],
"index": 5
},
{
"type": "text",
"bbox": [
107,
213,
360,
225
],
"lines": [
{
"bbox": [
105,
211,
361,
227
],
"spans": [
{
"bbox": [
105,
211,
361,
227
],
"score": 1.0,
"content": "The ambitious project soon faced several important challenges:",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
106,
230,
506,
300
],
"lines": [
{
"bbox": [
105,
228,
505,
244
],
"spans": [
{
"bbox": [
105,
228,
505,
244
],
"score": 1.0,
"content": "• Lack of computational resources: No organization is willing to sponsor such a big project and",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
112,
240,
202,
254
],
"spans": [
{
"bbox": [
112,
240,
202,
254
],
"score": 1.0,
"content": "freely make it public.",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
253,
506,
267
],
"spans": [
{
"bbox": [
105,
253,
506,
267
],
"score": 1.0,
"content": "• Lack of a robust pre-training algorithm: Despite GPT-3’s success on English corpus, it is",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
113,
265,
447,
277
],
"spans": [
{
"bbox": [
113,
265,
447,
277
],
"score": 1.0,
"content": "unclear how to train a high-accurate bilingual model for both English and Chinese.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
277,
506,
291
],
"spans": [
{
"bbox": [
106,
277,
506,
291
],
"score": 1.0,
"content": "• Lack of fast inference solutions: Since the goal is to have the model public to everyone, we need",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
114,
289,
447,
302
],
"spans": [
{
"bbox": [
114,
289,
447,
302
],
"score": 1.0,
"content": "to design fast inference solutions with low resource requirements to run the model.",
"type": "text"
}
],
"index": 16
}
],
"index": 13.5
},
{
"type": "text",
"bbox": [
107,
306,
505,
350
],
"lines": [
{
"bbox": [
105,
304,
505,
318
],
"spans": [
{
"bbox": [
105,
304,
505,
318
],
"score": 1.0,
"content": "For the pre-training algorithm, we finally chose GLM (Du et al., 2022) due to its high performance",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
317,
505,
329
],
"spans": [
{
"bbox": [
105,
317,
505,
329
],
"score": 1.0,
"content": "in practice. We eventually decided to train a GLM model of 130 billion parameters after several",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
328,
505,
341
],
"spans": [
{
"bbox": [
105,
328,
505,
341
],
"score": 1.0,
"content": "rounds of discussions and exploration, because such a size makes it possible to run the inference on",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
339,
235,
352
],
"spans": [
{
"bbox": [
105,
339,
167,
352
],
"score": 1.0,
"content": "a single A100",
"type": "text"
},
{
"bbox": [
167,
339,
203,
350
],
"score": 0.6,
"content": "( 4 0 G * 8 )",
"type": "inline_equation"
},
{
"bbox": [
203,
339,
235,
352
],
"score": 1.0,
"content": "server.",
"type": "text"
}
],
"index": 20
}
],
"index": 18.5
},
{
"type": "text",
"bbox": [
106,
355,
505,
455
],
"lines": [
{
"bbox": [
106,
356,
505,
368
],
"spans": [
{
"bbox": [
106,
356,
505,
368
],
"score": 1.0,
"content": "Our first attempt at training the model was in January 2022, shortly after we received a small sponsor",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
366,
505,
378
],
"spans": [
{
"bbox": [
106,
366,
505,
378
],
"score": 1.0,
"content": "of GPUs for test running. However, we soon realized that we had significantly underestimated the",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
104,
376,
507,
392
],
"spans": [
{
"bbox": [
104,
376,
353,
392
],
"score": 1.0,
"content": "technical difficulties of pre-training a model at such a scale",
"type": "text"
},
{
"bbox": [
353,
378,
387,
389
],
"score": 0.78,
"content": "\\left( > 1 0 0 \\mathbf { B } \\right)",
"type": "inline_equation"
},
{
"bbox": [
387,
376,
507,
392
],
"score": 1.0,
"content": ". It seems that pre-training a",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
388,
506,
402
],
"spans": [
{
"bbox": [
105,
388,
506,
402
],
"score": 1.0,
"content": "highly accurate 100B-scale model is quite different from training a 10B-scale one. Due to frequent",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
104,
399,
506,
413
],
"spans": [
{
"bbox": [
104,
399,
506,
413
],
"score": 1.0,
"content": "random hardware failures, model gradients exploding, unexpected excessive memory usage in the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
411,
505,
424
],
"spans": [
{
"bbox": [
106,
411,
505,
424
],
"score": 1.0,
"content": "algorithm, debug for the 3D pipeline in the new Megatron and DeepSpeed frameworks, inability to",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
421,
505,
434
],
"spans": [
{
"bbox": [
105,
421,
505,
434
],
"score": 1.0,
"content": "recover from optimizer states, blocked TCP responses between processes, and many many unex-",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
432,
506,
446
],
"spans": [
{
"bbox": [
105,
432,
506,
446
],
"score": 1.0,
"content": "pected “bugs”, the project was delayed for many times. The Tsinghua PACMAN team gave us a",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
442,
428,
457
],
"spans": [
{
"bbox": [
105,
442,
428,
457
],
"score": 1.0,
"content": "hand at this difficult time and together we successfully fixed most of the “bugs”.",
"type": "text"
}
],
"index": 29
}
],
"index": 25
},
{
"type": "text",
"bbox": [
107,
460,
505,
581
],
"lines": [
{
"bbox": [
105,
460,
506,
473
],
"spans": [
{
"bbox": [
105,
460,
506,
473
],
"score": 1.0,
"content": "By March, we were still short on computational resources, but fortunately got a chance to try test",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
471,
505,
484
],
"spans": [
{
"bbox": [
105,
471,
505,
484
],
"score": 1.0,
"content": "runs on several other platforms, including Ascend 910, Hygon DCU, NVIDIA, and Sunway. The",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
482,
505,
495
],
"spans": [
{
"bbox": [
105,
482,
505,
495
],
"score": 1.0,
"content": "immediate challenge was for us to adapt our training code to these different platforms, as the under-",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
493,
506,
506
],
"spans": [
{
"bbox": [
105,
493,
506,
506
],
"score": 1.0,
"content": "lying operators are quite different. Also, it introduced many new issues: the element-wise operators",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
504,
505,
516
],
"spans": [
{
"bbox": [
105,
504,
505,
516
],
"score": 1.0,
"content": "not supporting fast computation for large-dimension vectors, various issues that hindered conver-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
515,
505,
528
],
"spans": [
{
"bbox": [
105,
515,
505,
528
],
"score": 1.0,
"content": "gence—the large gradient norms of input embeddings, native Post-LN, Pre-LN, and Sandwich-LN,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
526,
505,
538
],
"spans": [
{
"bbox": [
105,
526,
505,
538
],
"score": 1.0,
"content": "dataloader state seeds, and computation precision choices in Softmax and Attention — as well as",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
536,
505,
550
],
"spans": [
{
"bbox": [
105,
536,
505,
550
],
"score": 1.0,
"content": "numerous mistakes we ourselves made. With tremendous help from all of our generous partners, we",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
547,
505,
561
],
"spans": [
{
"bbox": [
105,
547,
505,
561
],
"score": 1.0,
"content": "finally succeeded in making our pre-training algorithms runnable across all the platforms—frankly,",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
559,
506,
571
],
"spans": [
{
"bbox": [
104,
559,
506,
571
],
"score": 1.0,
"content": "a surprising achievement for this project. The timeline of GLM-130B in Figure 21 covers most of",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
569,
364,
583
],
"spans": [
{
"bbox": [
105,
569,
364,
583
],
"score": 1.0,
"content": "the issues we have encountered and addressed as of this writing.",
"type": "text"
}
],
"index": 40
}
],
"index": 35
},
{
"type": "text",
"bbox": [
107,
587,
505,
631
],
"lines": [
{
"bbox": [
105,
587,
506,
600
],
"spans": [
{
"bbox": [
105,
587,
506,
600
],
"score": 1.0,
"content": "On April 26th, we received a generous computing sponsorship from Zhipu.AI — an AI startup that",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
597,
506,
610
],
"spans": [
{
"bbox": [
105,
597,
506,
610
],
"score": 1.0,
"content": "aims to teach machines to think like humans. After another week of testing, we finally kicked off",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
608,
505,
622
],
"spans": [
{
"bbox": [
105,
608,
325,
622
],
"score": 1.0,
"content": "the training of the GLM-130B model on its 96 A100",
"type": "text"
},
{
"bbox": [
325,
609,
363,
620
],
"score": 0.33,
"content": "( 4 0 G * 8 )",
"type": "inline_equation"
},
{
"bbox": [
363,
608,
505,
622
],
"score": 1.0,
"content": "servers on May 6th. Additionally,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
620,
505,
632
],
"spans": [
{
"bbox": [
105,
620,
505,
632
],
"score": 1.0,
"content": "Zhipu.AI also sent a team to help evaluate the pre-trained model and build a demonstration website.",
"type": "text"
}
],
"index": 44
}
],
"index": 42.5
},
{
"type": "text",
"bbox": [
107,
636,
504,
703
],
"lines": [
{
"bbox": [
105,
636,
506,
649
],
"spans": [
{
"bbox": [
105,
636,
506,
649
],
"score": 1.0,
"content": "The training period spanned two months, during which we began developing a toolkit to allow GLM-",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
646,
506,
661
],
"spans": [
{
"bbox": [
105,
646,
506,
661
],
"score": 1.0,
"content": "130B’s inference in low-resource setting with swapping technique and quantization. Though it is",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
658,
506,
671
],
"spans": [
{
"bbox": [
105,
658,
506,
671
],
"score": 1.0,
"content": "already the most accessible model of its scale, together with our partner from Tsinghua NLP, we have",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
106,
670,
505,
681
],
"spans": [
{
"bbox": [
106,
670,
505,
681
],
"score": 1.0,
"content": "been exploring the limit of popularized hardware platforms, which would truly make the 100B-scale",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
680,
506,
694
],
"spans": [
{
"bbox": [
105,
680,
506,
694
],
"score": 1.0,
"content": "model accessible to as many people as possible. To date, we managed to reach the INT4 weight",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
690,
506,
705
],
"spans": [
{
"bbox": [
105,
690,
506,
705
],
"score": 1.0,
"content": "quantization for GLM-130B. Importantly, the INT4 version of GLM-130B without post training",
"type": "text"
}
],
"index": 50
}
],
"index": 47.5
}
],
"page_idx": 52,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
712,
502,
732
],
"lines": [
{
"bbox": [
115,
708,
504,
724
],
"spans": [
{
"bbox": [
115,
708,
504,
724
],
"score": 1.0,
"content": "16This section is largely extracted and updated from the blog introduction of GLM-130B at http://keg.",
"type": "text"
}
]
},
{
"bbox": [
105,
720,
372,
732
],
"spans": [
{
"bbox": [
105,
720,
372,
732
],
"score": 1.0,
"content": "cs.tsinghua.edu.cn/glm-130b/ (Posted date: August 4, 2022).",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
107,
27,
293,
37
],
"lines": [
{
"bbox": [
106,
25,
293,
38
],
"spans": [
{
"bbox": [
106,
25,
293,
38
],
"score": 1.0,
"content": "Published as a conference paper at ICLR 2023",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
298,
750,
313,
763
],
"spans": [
{
"bbox": [
298,
750,
313,
763
],
"score": 1.0,
"content": "53",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
107,
81,
301,
94
],
"lines": [
{
"bbox": [
106,
81,
301,
95
],
"spans": [
{
"bbox": [
106,
81,
301,
95
],
"score": 1.0,
"content": "F A BRIEF HISTORY OF GLM-130B",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
108,
505,
208
],
"lines": [
{
"bbox": [
105,
106,
505,
122
],
"spans": [
{
"bbox": [
105,
106,
505,
122
],
"score": 1.0,
"content": "The GLM-130B project16 was conceived in Dec. 2021 in a brainstorming meeting at Tsinghua KEG.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
119,
505,
131
],
"spans": [
{
"bbox": [
106,
119,
505,
131
],
"score": 1.0,
"content": "We firmly believe that it is of value to pre-train a highly accurate language model, in particular for",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
131,
505,
142
],
"spans": [
{
"bbox": [
106,
131,
505,
142
],
"score": 1.0,
"content": "both Chinese and English. Though GPT-3 (Brown et al., 2020) is the pioneer for this effort, it is not",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
142,
505,
154
],
"spans": [
{
"bbox": [
105,
142,
505,
154
],
"score": 1.0,
"content": "available to most people in the world. In addition, it supports English only. We therefore decide to",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
153,
506,
164
],
"spans": [
{
"bbox": [
105,
153,
506,
164
],
"score": 1.0,
"content": "initialize the project GLM-130B. Please note that the WuDao 1.75T model we built last year is a",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
163,
505,
176
],
"spans": [
{
"bbox": [
105,
163,
505,
176
],
"score": 1.0,
"content": "sparse model with 480 mixture-of-experts (MoE), rather than a dense one as GPT-3. Our goal then",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
175,
505,
186
],
"spans": [
{
"bbox": [
105,
175,
505,
186
],
"score": 1.0,
"content": "is to train a bilingual pre-trained dense model with high accuracy on downstream tasks, and to make",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
185,
505,
198
],
"spans": [
{
"bbox": [
105,
185,
505,
198
],
"score": 1.0,
"content": "it open to everyone in the world-anyone, anywhere can download it and use it on a single server with",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
197,
183,
208
],
"spans": [
{
"bbox": [
105,
197,
183,
208
],
"score": 1.0,
"content": "appropriate GPUs.",
"type": "text"
}
],
"index": 9
}
],
"index": 5,
"bbox_fs": [
105,
106,
506,
208
]
},
{
"type": "text",
"bbox": [
107,
213,
360,
225
],
"lines": [
{
"bbox": [
105,
211,
361,
227
],
"spans": [
{
"bbox": [
105,
211,
361,
227
],
"score": 1.0,
"content": "The ambitious project soon faced several important challenges:",
"type": "text"
}
],
"index": 10
}
],
"index": 10,
"bbox_fs": [
105,
211,
361,
227
]
},
{
"type": "text",
"bbox": [
106,
230,
506,
300
],
"lines": [
{
"bbox": [
105,
228,
505,
244
],
"spans": [
{
"bbox": [
105,
228,
505,
244
],
"score": 1.0,
"content": "• Lack of computational resources: No organization is willing to sponsor such a big project and",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
112,
240,
202,
254
],
"spans": [
{
"bbox": [
112,
240,
202,
254
],
"score": 1.0,
"content": "freely make it public.",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
253,
506,
267
],
"spans": [
{
"bbox": [
105,
253,
506,
267
],
"score": 1.0,
"content": "• Lack of a robust pre-training algorithm: Despite GPT-3’s success on English corpus, it is",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
113,
265,
447,
277
],
"spans": [
{
"bbox": [
113,
265,
447,
277
],
"score": 1.0,
"content": "unclear how to train a high-accurate bilingual model for both English and Chinese.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
277,
506,
291
],
"spans": [
{
"bbox": [
106,
277,
506,
291
],
"score": 1.0,
"content": "• Lack of fast inference solutions: Since the goal is to have the model public to everyone, we need",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
114,
289,
447,
302
],
"spans": [
{
"bbox": [
114,
289,
447,
302
],
"score": 1.0,
"content": "to design fast inference solutions with low resource requirements to run the model.",
"type": "text"
}
],
"index": 16
}
],
"index": 13.5,
"bbox_fs": [
105,
228,
506,
302
]
},
{
"type": "text",
"bbox": [
107,
306,
505,
350
],
"lines": [
{
"bbox": [
105,
304,
505,
318
],
"spans": [
{
"bbox": [
105,
304,
505,
318
],
"score": 1.0,
"content": "For the pre-training algorithm, we finally chose GLM (Du et al., 2022) due to its high performance",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
317,
505,
329
],
"spans": [
{
"bbox": [
105,
317,
505,
329
],
"score": 1.0,
"content": "in practice. We eventually decided to train a GLM model of 130 billion parameters after several",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
328,
505,
341
],
"spans": [
{
"bbox": [
105,
328,
505,
341
],
"score": 1.0,
"content": "rounds of discussions and exploration, because such a size makes it possible to run the inference on",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
339,
235,
352
],
"spans": [
{
"bbox": [
105,
339,
167,
352
],
"score": 1.0,
"content": "a single A100",
"type": "text"
},
{
"bbox": [
167,
339,
203,
350
],
"score": 0.6,
"content": "( 4 0 G * 8 )",
"type": "inline_equation"
},
{
"bbox": [
203,
339,
235,
352
],
"score": 1.0,
"content": "server.",
"type": "text"
}
],
"index": 20
}
],
"index": 18.5,
"bbox_fs": [
105,
304,
505,
352
]
},
{
"type": "text",
"bbox": [
106,
355,
505,
455
],
"lines": [
{
"bbox": [
106,
356,
505,
368
],
"spans": [
{
"bbox": [
106,
356,
505,
368
],
"score": 1.0,
"content": "Our first attempt at training the model was in January 2022, shortly after we received a small sponsor",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
366,
505,
378
],
"spans": [
{
"bbox": [
106,
366,
505,
378
],
"score": 1.0,
"content": "of GPUs for test running. However, we soon realized that we had significantly underestimated the",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
104,
376,
507,
392
],
"spans": [
{
"bbox": [
104,
376,
353,
392
],
"score": 1.0,
"content": "technical difficulties of pre-training a model at such a scale",
"type": "text"
},
{
"bbox": [
353,
378,
387,
389
],
"score": 0.78,
"content": "\\left( > 1 0 0 \\mathbf { B } \\right)",
"type": "inline_equation"
},
{
"bbox": [
387,
376,
507,
392
],
"score": 1.0,
"content": ". It seems that pre-training a",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
388,
506,
402
],
"spans": [
{
"bbox": [
105,
388,
506,
402
],
"score": 1.0,
"content": "highly accurate 100B-scale model is quite different from training a 10B-scale one. Due to frequent",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
104,
399,
506,
413
],
"spans": [
{
"bbox": [
104,
399,
506,
413
],
"score": 1.0,
"content": "random hardware failures, model gradients exploding, unexpected excessive memory usage in the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
411,
505,
424
],
"spans": [
{
"bbox": [
106,
411,
505,
424
],
"score": 1.0,
"content": "algorithm, debug for the 3D pipeline in the new Megatron and DeepSpeed frameworks, inability to",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
421,
505,
434
],
"spans": [
{
"bbox": [
105,
421,
505,
434
],
"score": 1.0,
"content": "recover from optimizer states, blocked TCP responses between processes, and many many unex-",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
432,
506,
446
],
"spans": [
{
"bbox": [
105,
432,
506,
446
],
"score": 1.0,
"content": "pected “bugs”, the project was delayed for many times. The Tsinghua PACMAN team gave us a",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
442,
428,
457
],
"spans": [
{
"bbox": [
105,
442,
428,
457
],
"score": 1.0,
"content": "hand at this difficult time and together we successfully fixed most of the “bugs”.",
"type": "text"
}
],
"index": 29
}
],
"index": 25,
"bbox_fs": [
104,
356,
507,
457
]
},
{
"type": "text",
"bbox": [
107,
460,
505,
581
],
"lines": [
{
"bbox": [
105,
460,
506,
473
],
"spans": [
{
"bbox": [
105,
460,
506,
473
],
"score": 1.0,
"content": "By March, we were still short on computational resources, but fortunately got a chance to try test",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
471,
505,
484
],
"spans": [
{
"bbox": [
105,
471,
505,
484
],
"score": 1.0,
"content": "runs on several other platforms, including Ascend 910, Hygon DCU, NVIDIA, and Sunway. The",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
482,
505,
495
],
"spans": [
{
"bbox": [
105,
482,
505,
495
],
"score": 1.0,
"content": "immediate challenge was for us to adapt our training code to these different platforms, as the under-",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
493,
506,
506
],
"spans": [
{
"bbox": [
105,
493,
506,
506
],
"score": 1.0,
"content": "lying operators are quite different. Also, it introduced many new issues: the element-wise operators",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
504,
505,
516
],
"spans": [
{
"bbox": [
105,
504,
505,
516
],
"score": 1.0,
"content": "not supporting fast computation for large-dimension vectors, various issues that hindered conver-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
515,
505,
528
],
"spans": [
{
"bbox": [
105,
515,
505,
528
],
"score": 1.0,
"content": "gence—the large gradient norms of input embeddings, native Post-LN, Pre-LN, and Sandwich-LN,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
526,
505,
538
],
"spans": [
{
"bbox": [
105,
526,
505,
538
],
"score": 1.0,
"content": "dataloader state seeds, and computation precision choices in Softmax and Attention — as well as",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
536,
505,
550
],
"spans": [
{
"bbox": [
105,
536,
505,
550
],
"score": 1.0,
"content": "numerous mistakes we ourselves made. With tremendous help from all of our generous partners, we",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
547,
505,
561
],
"spans": [
{
"bbox": [
105,
547,
505,
561
],
"score": 1.0,
"content": "finally succeeded in making our pre-training algorithms runnable across all the platforms—frankly,",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
559,
506,
571
],
"spans": [
{
"bbox": [
104,
559,
506,
571
],
"score": 1.0,
"content": "a surprising achievement for this project. The timeline of GLM-130B in Figure 21 covers most of",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
569,
364,
583
],
"spans": [
{
"bbox": [
105,
569,
364,
583
],
"score": 1.0,
"content": "the issues we have encountered and addressed as of this writing.",
"type": "text"
}
],
"index": 40
}
],
"index": 35,
"bbox_fs": [
104,
460,
506,
583
]
},
{
"type": "text",
"bbox": [
107,
587,
505,
631
],
"lines": [
{
"bbox": [
105,
587,
506,
600
],
"spans": [
{
"bbox": [
105,
587,
506,
600
],
"score": 1.0,
"content": "On April 26th, we received a generous computing sponsorship from Zhipu.AI — an AI startup that",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
597,
506,
610
],
"spans": [
{
"bbox": [
105,
597,
506,
610
],
"score": 1.0,
"content": "aims to teach machines to think like humans. After another week of testing, we finally kicked off",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
608,
505,
622
],
"spans": [
{
"bbox": [
105,
608,
325,
622
],
"score": 1.0,
"content": "the training of the GLM-130B model on its 96 A100",
"type": "text"
},
{
"bbox": [
325,
609,
363,
620
],
"score": 0.33,
"content": "( 4 0 G * 8 )",
"type": "inline_equation"
},
{
"bbox": [
363,
608,
505,
622
],
"score": 1.0,
"content": "servers on May 6th. Additionally,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
620,
505,
632
],
"spans": [
{
"bbox": [
105,
620,
505,
632
],
"score": 1.0,
"content": "Zhipu.AI also sent a team to help evaluate the pre-trained model and build a demonstration website.",
"type": "text"
}
],
"index": 44
}
],
"index": 42.5,
"bbox_fs": [
105,
587,
506,
632
]
},
{
"type": "text",
"bbox": [
107,
636,
504,
703
],
"lines": [
{
"bbox": [
105,
636,
506,
649
],
"spans": [
{
"bbox": [
105,
636,
506,
649
],
"score": 1.0,
"content": "The training period spanned two months, during which we began developing a toolkit to allow GLM-",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
646,
506,
661
],
"spans": [
{
"bbox": [
105,
646,
506,
661
],
"score": 1.0,
"content": "130B’s inference in low-resource setting with swapping technique and quantization. Though it is",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
658,
506,
671
],
"spans": [
{
"bbox": [
105,
658,
506,
671
],
"score": 1.0,
"content": "already the most accessible model of its scale, together with our partner from Tsinghua NLP, we have",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
106,
670,
505,
681
],
"spans": [
{
"bbox": [
106,
670,
505,
681
],
"score": 1.0,
"content": "been exploring the limit of popularized hardware platforms, which would truly make the 100B-scale",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
680,
506,
694
],
"spans": [
{
"bbox": [
105,
680,
506,
694
],
"score": 1.0,
"content": "model accessible to as many people as possible. To date, we managed to reach the INT4 weight",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
690,
506,
705
],
"spans": [
{
"bbox": [
105,
690,
506,
705
],
"score": 1.0,
"content": "quantization for GLM-130B. Importantly, the INT4 version of GLM-130B without post training",
"type": "text"
}
],
"index": 50
}
],
"index": 47.5,
"bbox_fs": [
105,
636,
506,
705
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
160,
71,
499,
90
],
"lines": [
{
"bbox": [
160,
71,
500,
91
],
"spans": [
{
"bbox": [
160,
71,
500,
91
],
"score": 1.0,
"content": "Major Issues Encountered for Training GLM-130B",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "title",
"bbox": [
142,
107,
171,
115
],
"lines": [
{
"bbox": [
142,
106,
171,
117
],
"spans": [
{
"bbox": [
142,
106,
171,
117
],
"score": 1.0,
"content": "2021.12",
"type": "text"
}
],
"index": 1
}
],
"index": 1
},
{
"type": "text",
"bbox": [
142,
117,
492,
154
],
"lines": [
{
"bbox": [
141,
115,
446,
127
],
"spans": [
{
"bbox": [
141,
115,
446,
127
],
"score": 1.0,
"content": "• The “千亿 ” (100B) project towards an open dense pre-trained GLM at 100B scale is conceived",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
140,
124,
493,
137
],
"spans": [
{
"bbox": [
140,
124,
422,
137
],
"score": 1.0,
"content": "• Survey pre-training strategies of existing models of similar scale, such as GPT-3, Gopher",
"type": "text"
},
{
"bbox": [
423,
127,
433,
134
],
"score": 0.69,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
433,
124,
493,
137
],
"score": 1.0,
"content": "Limited public info",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
147,
134,
305,
145
],
"spans": [
{
"bbox": [
147,
134,
305,
145
],
"score": 1.0,
"content": "about how they were trained and issues they met",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
141,
143,
289,
154
],
"spans": [
{
"bbox": [
141,
143,
289,
154
],
"score": 1.0,
"content": "• Search for possible GPU clusters & sponsors",
"type": "text"
}
],
"index": 5
}
],
"index": 3.5
},
{
"type": "title",
"bbox": [
142,
159,
166,
167
],
"lines": [
{
"bbox": [
141,
159,
168,
169
],
"spans": [
{
"bbox": [
141,
159,
168,
169
],
"score": 1.0,
"content": "2022.1",
"type": "text"
}
],
"index": 6
}
],
"index": 6
},
{
"type": "text",
"bbox": [
141,
168,
452,
205
],
"lines": [
{
"bbox": [
142,
168,
375,
178
],
"spans": [
{
"bbox": [
142,
168,
375,
178
],
"score": 1.0,
"content": "• Test the performance of FP16/FP32 at 100B scale on one testing cluster",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
140,
177,
453,
189
],
"spans": [
{
"bbox": [
140,
177,
290,
189
],
"score": 1.0,
"content": "• Unexpected excessive memory usage in GLM",
"type": "text"
},
{
"bbox": [
291,
179,
301,
186
],
"score": 0.8,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
301,
177,
453,
189
],
"score": 1.0,
"content": "Torch is better with fixed length input sequences",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
141,
187,
393,
196
],
"spans": [
{
"bbox": [
141,
187,
327,
196
],
"score": 1.0,
"content": "• Inability to converge and try tricks from CogView and ViT",
"type": "text"
},
{
"bbox": [
327,
188,
350,
195
],
"score": 0.71,
"content": "{ \\Rightarrow U s e }",
"type": "inline_equation"
},
{
"bbox": [
350,
187,
393,
196
],
"score": 1.0,
"content": "Sandwich-LN",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
141,
195,
393,
206
],
"spans": [
{
"bbox": [
141,
195,
261,
206
],
"score": 1.0,
"content": "• Frequent random hardware failures",
"type": "text"
},
{
"bbox": [
261,
197,
271,
204
],
"score": 0.76,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
272,
195,
393,
206
],
"score": 1.0,
"content": "Have to run HCPG test before each run",
"type": "text"
}
],
"index": 10
}
],
"index": 8.5
},
{
"type": "title",
"bbox": [
142,
212,
167,
220
],
"lines": [
{
"bbox": [
141,
211,
168,
221
],
"spans": [
{
"bbox": [
141,
211,
168,
221
],
"score": 1.0,
"content": "2022.2",
"type": "text"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "text",
"bbox": [
142,
221,
489,
259
],
"lines": [
{
"bbox": [
141,
219,
491,
232
],
"spans": [
{
"bbox": [
141,
219,
309,
232
],
"score": 1.0,
"content": "• Very slow training speed than previously calculated",
"type": "text"
},
{
"bbox": [
310,
222,
320,
229
],
"score": 0.74,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
320,
219,
434,
232
],
"score": 1.0,
"content": "Optimize kernels and fuse operators",
"type": "text"
},
{
"bbox": [
434,
222,
445,
229
],
"score": 0.61,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
445,
219,
491,
232
],
"score": 1.0,
"content": "Find the input",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
147,
230,
270,
241
],
"spans": [
{
"bbox": [
147,
230,
270,
241
],
"score": 1.0,
"content": "shape is critical to kernel performance",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
142,
239,
464,
249
],
"spans": [
{
"bbox": [
142,
239,
277,
249
],
"score": 1.0,
"content": "• Collect pre-training corpora and tokenize",
"type": "text"
},
{
"bbox": [
278,
240,
288,
248
],
"score": 0.69,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
289,
239,
464,
249
],
"score": 1.0,
"content": "Use icetk: the sentence piece is set to the unigram mode",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
247,
399,
259
],
"spans": [
{
"bbox": [
141,
247,
399,
259
],
"score": 1.0,
"content": "• Debug the 3D pipeline parallel in the newly-released Megatron and DeepSpeed",
"type": "text"
}
],
"index": 15
}
],
"index": 13.5
},
{
"type": "title",
"bbox": [
142,
264,
167,
272
],
"lines": [
{
"bbox": [
141,
263,
168,
273
],
"spans": [
{
"bbox": [
141,
263,
168,
273
],
"score": 1.0,
"content": "2022.3",
"type": "text"
}
],
"index": 16
}
],
"index": 16
},
{
"type": "text",
"bbox": [
142,
273,
486,
365
],
"lines": [
{
"bbox": [
141,
271,
460,
283
],
"spans": [
{
"bbox": [
141,
271,
282,
283
],
"score": 1.0,
"content": "• It can’t recover perfectly from checkpoints",
"type": "text"
},
{
"bbox": [
283,
274,
293,
281
],
"score": 0.64,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
293,
271,
460,
283
],
"score": 1.0,
"content": "Our customized dataloader do not save its state seed",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
147,
282,
246,
293
],
"spans": [
{
"bbox": [
147,
282,
246,
293
],
"score": 1.0,
"content": "properly in distributed training",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
289,
486,
302
],
"spans": [
{
"bbox": [
141,
289,
271,
302
],
"score": 1.0,
"content": "• The memory per processor is too small",
"type": "text"
},
{
"bbox": [
271,
292,
281,
299
],
"score": 0.79,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
281,
289,
385,
302
],
"score": 1.0,
"content": "Require too many pipeline stages",
"type": "text"
},
{
"bbox": [
385,
292,
395,
299
],
"score": 0.75,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
396,
289,
486,
302
],
"score": 1.0,
"content": "Batch size is too large (up to",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
148,
299,
283,
311
],
"spans": [
{
"bbox": [
148,
301,
203,
309
],
"score": 0.36,
"content": "1 2 , 0 0 0 ) \\Rightarrow H a r m",
"type": "inline_equation"
},
{
"bbox": [
203,
299,
283,
311
],
"score": 1.0,
"content": "the model’s convergency",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
141,
307,
473,
321
],
"spans": [
{
"bbox": [
141,
307,
303,
321
],
"score": 1.0,
"content": "• It can’t launch more than 2,000 computing nodes",
"type": "text"
},
{
"bbox": [
304,
311,
314,
317
],
"score": 0.76,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
314,
307,
473,
321
],
"score": 1.0,
"content": "Overcome this and support 6,000-node training by",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
147,
318,
261,
329
],
"spans": [
{
"bbox": [
147,
318,
261,
329
],
"score": 1.0,
"content": "tuning Linux kernel TCP parameters",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
141,
327,
307,
339
],
"spans": [
{
"bbox": [
141,
327,
307,
339
],
"score": 1.0,
"content": "• Collect data for multi-task instruction pre-training",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
141,
336,
349,
348
],
"spans": [
{
"bbox": [
141,
336,
349,
348
],
"score": 1.0,
"content": "• Receive opportunities to test trainings on several other clusters",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
141,
346,
465,
357
],
"spans": [
{
"bbox": [
141,
346,
273,
357
],
"score": 1.0,
"content": "• Very slow training speed than expected",
"type": "text"
},
{
"bbox": [
273,
348,
283,
355
],
"score": 0.76,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
284,
346,
465,
357
],
"score": 1.0,
"content": "The underlying element-wise operators don’t support fast",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
147,
356,
279,
367
],
"spans": [
{
"bbox": [
147,
356,
279,
367
],
"score": 1.0,
"content": "computation on large-dimension vectors.",
"type": "text"
}
],
"index": 26
}
],
"index": 21.5
},
{
"type": "title",
"bbox": [
126,
371,
167,
380
],
"lines": [
{
"bbox": [
141,
371,
168,
381
],
"spans": [
{
"bbox": [
141,
371,
168,
381
],
"score": 1.0,
"content": "2022.4",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "text",
"bbox": [
142,
381,
485,
501
],
"lines": [
{
"bbox": [
141,
380,
447,
392
],
"spans": [
{
"bbox": [
141,
380,
289,
392
],
"score": 1.0,
"content": "• Optimize A100 kernel’s computing efficiency",
"type": "text"
},
{
"bbox": [
289,
382,
299,
389
],
"score": 0.78,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
299,
380,
447,
392
],
"score": 1.0,
"content": "A100 kernels prefer square-shaped inputs, and",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
147,
391,
353,
401
],
"spans": [
{
"bbox": [
147,
391,
353,
401
],
"score": 1.0,
"content": "seq_len=2,048 is optimal for our hidden-state dimension (12,288)",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
142,
399,
482,
410
],
"spans": [
{
"bbox": [
142,
399,
303,
410
],
"score": 1.0,
"content": "• Inability to converge due to large gradient norms",
"type": "text"
},
{
"bbox": [
304,
400,
324,
409
],
"score": 0.6,
"content": "( 1 7 0 + )",
"type": "inline_equation"
},
{
"bbox": [
324,
399,
390,
410
],
"score": 1.0,
"content": "of input embeddings",
"type": "text"
},
{
"bbox": [
391,
400,
412,
408
],
"score": 0.41,
"content": "{ \\Rightarrow T r y }",
"type": "inline_equation"
},
{
"bbox": [
413,
399,
482,
410
],
"score": 1.0,
"content": "embedding norm and",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
146,
408,
322,
420
],
"spans": [
{
"bbox": [
146,
408,
322,
420
],
"score": 1.0,
"content": "gradient shrink, which turn out to be almost equivalent",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
142,
417,
477,
428
],
"spans": [
{
"bbox": [
142,
417,
368,
428
],
"score": 1.0,
"content": "• Naïve post-LN or pre-LN disconverges after several thousands of steps",
"type": "text"
},
{
"bbox": [
368,
419,
390,
427
],
"score": 0.35,
"content": "\\Rightarrow T r y",
"type": "inline_equation"
},
{
"bbox": [
390,
417,
477,
428
],
"score": 1.0,
"content": "Sandwich-LN with PB-Relax",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
141,
425,
487,
438
],
"spans": [
{
"bbox": [
141,
425,
277,
438
],
"score": 1.0,
"content": "• It still disconverges after one week’s trial",
"type": "text"
},
{
"bbox": [
277,
428,
288,
435
],
"score": 0.78,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
288,
425,
487,
438
],
"score": 1.0,
"content": "The dataloader state seeds are not unified for different pipeline",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
147,
437,
323,
446
],
"spans": [
{
"bbox": [
147,
437,
323,
446
],
"score": 1.0,
"content": "stages, resulting in a mismatch of input data and labels.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
141,
444,
449,
456
],
"spans": [
{
"bbox": [
141,
444,
292,
456
],
"score": 1.0,
"content": "• Test two positional encodings: RoPE and Alibi",
"type": "text"
},
{
"bbox": [
293,
446,
303,
454
],
"score": 0.8,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
303,
444,
449,
456
],
"score": 1.0,
"content": "Alibi can be slower as it requires element-wise",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
147,
454,
435,
466
],
"spans": [
{
"bbox": [
147,
454,
357,
466
],
"score": 1.0,
"content": "manipulation on attention matrices---changing num_heads *2,048",
"type": "text"
},
{
"bbox": [
358,
456,
363,
463
],
"score": 0.3,
"content": "^ *",
"type": "inline_equation"
},
{
"bbox": [
363,
454,
435,
466
],
"score": 1.0,
"content": "2,048 scalars per layer",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
141,
462,
482,
475
],
"spans": [
{
"bbox": [
141,
462,
202,
475
],
"score": 1.0,
"content": "• Test GeGLU and",
"type": "text"
},
{
"bbox": [
202,
465,
244,
472
],
"score": 0.41,
"content": "{ \\mathsf { G A U } } \\ { \\mathsf { \\Omega } } \\Rightarrow { \\mathsf { G A U } }",
"type": "inline_equation"
},
{
"bbox": [
244,
462,
482,
475
],
"score": 1.0,
"content": "converges faster with relatively poor performance on fine-tuned SuperGLUE",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
141,
472,
483,
484
],
"spans": [
{
"bbox": [
141,
472,
362,
484
],
"score": 1.0,
"content": "• Abnormal GPU memory usage of newly-added functions and classes",
"type": "text"
},
{
"bbox": [
362,
474,
372,
482
],
"score": 0.8,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
373,
472,
483,
484
],
"score": 1.0,
"content": "DeepSpeed hardcodes the function",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
147,
482,
251,
493
],
"spans": [
{
"bbox": [
147,
482,
251,
493
],
"score": 1.0,
"content": "names for checkpoint activation",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
142,
492,
441,
502
],
"spans": [
{
"bbox": [
142,
492,
300,
502
],
"score": 1.0,
"content": "• Decide to train GLM with 130 billion parameters",
"type": "text"
},
{
"bbox": [
300,
493,
310,
500
],
"score": 0.77,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
311,
492,
441,
502
],
"score": 1.0,
"content": "allow inference on a DGX-A100 40G node",
"type": "text"
}
],
"index": 40
}
],
"index": 34
},
{
"type": "title",
"bbox": [
142,
507,
173,
515
],
"lines": [
{
"bbox": [
141,
506,
174,
516
],
"spans": [
{
"bbox": [
141,
506,
174,
516
],
"score": 1.0,
"content": "2022.5-6",
"type": "text"
}
],
"index": 41
}
],
"index": 41
},
{
"type": "text",
"bbox": [
142,
516,
495,
600
],
"lines": [
{
"bbox": [
142,
516,
479,
525
],
"spans": [
{
"bbox": [
142,
516,
261,
525
],
"score": 1.0,
"content": "• Implement a RoPE cuda operator in",
"type": "text"
},
{
"bbox": [
261,
517,
284,
525
],
"score": 0.56,
"content": "{ \\mathsf { C } } + + = >",
"type": "inline_equation"
},
{
"bbox": [
285,
516,
479,
525
],
"score": 1.0,
"content": "See unexpected precision errors and finally have it abandoned",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
142,
523,
473,
537
],
"spans": [
{
"bbox": [
142,
523,
245,
537
],
"score": 1.0,
"content": "• Sandwich-LN still disconverges",
"type": "text"
},
{
"bbox": [
245,
526,
259,
534
],
"score": 0.77,
"content": "\\Longrightarrow 1",
"type": "inline_equation"
},
{
"bbox": [
260,
523,
473,
537
],
"score": 1.0,
"content": ") Reducing learning rate does not help; 2) Using Hinge cross-entropy",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
146,
534,
411,
545
],
"spans": [
{
"bbox": [
146,
534,
411,
545
],
"score": 1.0,
"content": "becomes slower and harms performance; 3) Shifting to DeepNorm still disconverges",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
142,
544,
288,
554
],
"spans": [
{
"bbox": [
142,
544,
252,
554
],
"score": 1.0,
"content": "• Use FP32 in softmax of attention",
"type": "text"
},
{
"bbox": [
252,
545,
262,
552
],
"score": 0.81,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
262,
544,
288,
554
],
"score": 1.0,
"content": "Success",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
141,
552,
496,
563
],
"spans": [
{
"bbox": [
141,
552,
288,
563
],
"score": 1.0,
"content": "• Find PB-Relax unnecessary for FP32 softmax",
"type": "text"
},
{
"bbox": [
288,
554,
297,
561
],
"score": 0.41,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
298,
552,
496,
563
],
"score": 1.0,
"content": "It also slows down training as it needs to manipulate the whole",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
147,
563,
226,
572
],
"spans": [
{
"bbox": [
147,
563,
226,
572
],
"score": 1.0,
"content": "attention score matrices",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
143,
572,
491,
582
],
"spans": [
{
"bbox": [
143,
572,
268,
582
],
"score": 1.0,
"content": "• Experience few spikes in later training",
"type": "text"
},
{
"bbox": [
268,
573,
279,
580
],
"score": 0.67,
"content": "\\mathbf { \\Phi } = > \\mathbf { \\Phi } .",
"type": "inline_equation"
},
{
"bbox": [
280,
572,
491,
582
],
"score": 1.0,
"content": "1) Reduce gradient shrink factor from 1 to 0.1: useful; 2) Reduce the",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
147,
580,
407,
591
],
"spans": [
{
"bbox": [
147,
580,
407,
591
],
"score": 1.0,
"content": "learning rate: sometimes useful; 3) Jump the noisy data batches: sometimes useful",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
141,
588,
490,
601
],
"spans": [
{
"bbox": [
141,
588,
346,
601
],
"score": 1.0,
"content": "• Find a mistake in multi-task data after training for 20,000 steps",
"type": "text"
},
{
"bbox": [
346,
591,
356,
599
],
"score": 0.66,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
357,
588,
490,
601
],
"score": 1.0,
"content": "Use the correct data but it does not forget",
"type": "text"
}
],
"index": 50
}
],
"index": 46
},
{
"type": "title",
"bbox": [
142,
605,
173,
613
],
"lines": [
{
"bbox": [
141,
605,
174,
615
],
"spans": [
{
"bbox": [
141,
605,
174,
615
],
"score": 1.0,
"content": "2022.6-7",
"type": "text"
}
],
"index": 51
}
],
"index": 51
},
{
"type": "text",
"bbox": [
142,
615,
498,
700
],
"lines": [
{
"bbox": [
141,
614,
496,
625
],
"spans": [
{
"bbox": [
141,
614,
496,
625
],
"score": 1.0,
"content": "• Adapt the pipeline parallel checkpoints to ordinary parallel checkpoints for efficient inference on a single A100",
"type": "text"
}
],
"index": 52
},
{
"bbox": [
141,
623,
405,
635
],
"spans": [
{
"bbox": [
141,
623,
405,
635
],
"score": 1.0,
"content": "• Work on evaluation scripts on datasets: MMLU, Big-bench, CLUE, SuperCLUE, etc.",
"type": "text"
}
],
"index": 53
},
{
"bbox": [
141,
633,
486,
643
],
"spans": [
{
"bbox": [
141,
633,
486,
643
],
"score": 1.0,
"content": "• Implement P-Tuning and P-Tuning v2 for parameter-efficient tuning on GLM-130B for tuning on SuperGLUE",
"type": "text"
}
],
"index": 54
},
{
"bbox": [
141,
641,
491,
653
],
"spans": [
{
"bbox": [
141,
641,
423,
653
],
"score": 1.0,
"content": "• Work with BMInf on adapting GLM-130B to perform inference on a single V100 or 3090",
"type": "text"
},
{
"bbox": [
423,
643,
433,
650
],
"score": 0.51,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
434,
641,
491,
653
],
"score": 1.0,
"content": "Use pipeline-style",
"type": "text"
}
],
"index": 55
},
{
"bbox": [
147,
651,
354,
662
],
"spans": [
{
"bbox": [
147,
651,
354,
662
],
"score": 1.0,
"content": "asynchronous swapping between main memory and GPU memory",
"type": "text"
}
],
"index": 56
},
{
"bbox": [
141,
659,
488,
672
],
"spans": [
{
"bbox": [
141,
659,
365,
672
],
"score": 1.0,
"content": "• Try to fine-tune GLM-130B with fewer A100 nodes (i.e., 12-16 nodes)",
"type": "text"
},
{
"bbox": [
366,
662,
376,
669
],
"score": 0.79,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
376,
659,
488,
672
],
"score": 1.0,
"content": "Pipeline-style fails due to too many",
"type": "text"
}
],
"index": 57
},
{
"bbox": [
147,
670,
498,
682
],
"spans": [
{
"bbox": [
147,
670,
195,
682
],
"score": 1.0,
"content": "pipeline stages",
"type": "text"
},
{
"bbox": [
196,
671,
206,
679
],
"score": 0.6,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
206,
670,
392,
682
],
"score": 1.0,
"content": "Find that data parallel can not be introduced for fine-tuning",
"type": "text"
},
{
"bbox": [
393,
671,
403,
679
],
"score": 0.64,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
403,
670,
498,
682
],
"score": 1.0,
"content": "Use 32-way model parallel for",
"type": "text"
}
],
"index": 58
},
{
"bbox": [
147,
679,
279,
690
],
"spans": [
{
"bbox": [
147,
679,
279,
690
],
"score": 1.0,
"content": "fine-tuning with reasonable performance",
"type": "text"
}
],
"index": 59
},
{
"bbox": [
349,
690,
488,
702
],
"spans": [
{
"bbox": [
349,
690,
488,
702
],
"score": 1.0,
"content": "https://github.com/THUDM/GLM-130B",
"type": "text"
}
],
"index": 60
}
],
"index": 56
},
{
"type": "text",
"bbox": [
106,
712,
506,
734
],
"lines": [
{
"bbox": [
105,
711,
506,
725
],
"spans": [
{
"bbox": [
105,
711,
506,
725
],
"score": 1.0,
"content": "Figure 21: The timeline of major issues that training GLM-130B encountered and addressed, as of",
"type": "text"
}
],
"index": 61
},
{
"bbox": [
105,
721,
172,
734
],
"spans": [
{
"bbox": [
105,
721,
172,
734
],
"score": 1.0,
"content": "July 31st, 2022.",
"type": "text"
}
],
"index": 62
}
],
"index": 61.5
}
],
"page_idx": 53,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
293,
37
],
"lines": [
{
"bbox": [
106,
26,
293,
38
],
"spans": [
{
"bbox": [
106,
26,
293,
38
],
"score": 1.0,
"content": "Published as a conference paper at ICLR 2023",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
298,
749,
313,
764
],
"spans": [
{
"bbox": [
298,
749,
313,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 15,
"width": 15
}
]
}
]
},
{
"type": "discarded",
"bbox": [
118,
61,
159,
96
],
"lines": []
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
160,
71,
499,
90
],
"lines": [
{
"bbox": [
160,
71,
500,
91
],
"spans": [
{
"bbox": [
160,
71,
500,
91
],
"score": 1.0,
"content": "Major Issues Encountered for Training GLM-130B",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "title",
"bbox": [
142,
107,
171,
115
],
"lines": [
{
"bbox": [
142,
106,
171,
117
],
"spans": [
{
"bbox": [
142,
106,
171,
117
],
"score": 1.0,
"content": "2021.12",
"type": "text"
}
],
"index": 1
}
],
"index": 1
},
{
"type": "list",
"bbox": [
142,
117,
492,
154
],
"lines": [
{
"bbox": [
141,
115,
446,
127
],
"spans": [
{
"bbox": [
141,
115,
446,
127
],
"score": 1.0,
"content": "• The “千亿 ” (100B) project towards an open dense pre-trained GLM at 100B scale is conceived",
"type": "text"
}
],
"index": 2,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
140,
124,
493,
137
],
"spans": [
{
"bbox": [
140,
124,
422,
137
],
"score": 1.0,
"content": "• Survey pre-training strategies of existing models of similar scale, such as GPT-3, Gopher",
"type": "text"
},
{
"bbox": [
423,
127,
433,
134
],
"score": 0.69,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
433,
124,
493,
137
],
"score": 1.0,
"content": "Limited public info",
"type": "text"
}
],
"index": 3,
"is_list_start_line": true
},
{
"bbox": [
147,
134,
305,
145
],
"spans": [
{
"bbox": [
147,
134,
305,
145
],
"score": 1.0,
"content": "about how they were trained and issues they met",
"type": "text"
}
],
"index": 4,
"is_list_end_line": true
},
{
"bbox": [
141,
143,
289,
154
],
"spans": [
{
"bbox": [
141,
143,
289,
154
],
"score": 1.0,
"content": "• Search for possible GPU clusters & sponsors",
"type": "text"
}
],
"index": 5,
"is_list_start_line": true,
"is_list_end_line": true
}
],
"index": 3.5,
"bbox_fs": [
140,
115,
493,
154
]
},
{
"type": "title",
"bbox": [
142,
159,
166,
167
],
"lines": [
{
"bbox": [
141,
159,
168,
169
],
"spans": [
{
"bbox": [
141,
159,
168,
169
],
"score": 1.0,
"content": "2022.1",
"type": "text"
}
],
"index": 6
}
],
"index": 6
},
{
"type": "text",
"bbox": [
141,
168,
452,
205
],
"lines": [
{
"bbox": [
142,
168,
375,
178
],
"spans": [
{
"bbox": [
142,
168,
375,
178
],
"score": 1.0,
"content": "• Test the performance of FP16/FP32 at 100B scale on one testing cluster",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
140,
177,
453,
189
],
"spans": [
{
"bbox": [
140,
177,
290,
189
],
"score": 1.0,
"content": "• Unexpected excessive memory usage in GLM",
"type": "text"
},
{
"bbox": [
291,
179,
301,
186
],
"score": 0.8,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
301,
177,
453,
189
],
"score": 1.0,
"content": "Torch is better with fixed length input sequences",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
141,
187,
393,
196
],
"spans": [
{
"bbox": [
141,
187,
327,
196
],
"score": 1.0,
"content": "• Inability to converge and try tricks from CogView and ViT",
"type": "text"
},
{
"bbox": [
327,
188,
350,
195
],
"score": 0.71,
"content": "{ \\Rightarrow U s e }",
"type": "inline_equation"
},
{
"bbox": [
350,
187,
393,
196
],
"score": 1.0,
"content": "Sandwich-LN",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
141,
195,
393,
206
],
"spans": [
{
"bbox": [
141,
195,
261,
206
],
"score": 1.0,
"content": "• Frequent random hardware failures",
"type": "text"
},
{
"bbox": [
261,
197,
271,
204
],
"score": 0.76,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
272,
195,
393,
206
],
"score": 1.0,
"content": "Have to run HCPG test before each run",
"type": "text"
}
],
"index": 10
}
],
"index": 8.5,
"bbox_fs": [
140,
168,
453,
206
]
},
{
"type": "title",
"bbox": [
142,
212,
167,
220
],
"lines": [
{
"bbox": [
141,
211,
168,
221
],
"spans": [
{
"bbox": [
141,
211,
168,
221
],
"score": 1.0,
"content": "2022.2",
"type": "text"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "list",
"bbox": [
142,
221,
489,
259
],
"lines": [
{
"bbox": [
141,
219,
491,
232
],
"spans": [
{
"bbox": [
141,
219,
309,
232
],
"score": 1.0,
"content": "• Very slow training speed than previously calculated",
"type": "text"
},
{
"bbox": [
310,
222,
320,
229
],
"score": 0.74,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
320,
219,
434,
232
],
"score": 1.0,
"content": "Optimize kernels and fuse operators",
"type": "text"
},
{
"bbox": [
434,
222,
445,
229
],
"score": 0.61,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
445,
219,
491,
232
],
"score": 1.0,
"content": "Find the input",
"type": "text"
}
],
"index": 12,
"is_list_start_line": true
},
{
"bbox": [
147,
230,
270,
241
],
"spans": [
{
"bbox": [
147,
230,
270,
241
],
"score": 1.0,
"content": "shape is critical to kernel performance",
"type": "text"
}
],
"index": 13,
"is_list_start_line": true
},
{
"bbox": [
142,
239,
464,
249
],
"spans": [
{
"bbox": [
142,
239,
277,
249
],
"score": 1.0,
"content": "• Collect pre-training corpora and tokenize",
"type": "text"
},
{
"bbox": [
278,
240,
288,
248
],
"score": 0.69,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
289,
239,
464,
249
],
"score": 1.0,
"content": "Use icetk: the sentence piece is set to the unigram mode",
"type": "text"
}
],
"index": 14,
"is_list_start_line": true
},
{
"bbox": [
141,
247,
399,
259
],
"spans": [
{
"bbox": [
141,
247,
399,
259
],
"score": 1.0,
"content": "• Debug the 3D pipeline parallel in the newly-released Megatron and DeepSpeed",
"type": "text"
}
],
"index": 15,
"is_list_start_line": true
}
],
"index": 13.5,
"bbox_fs": [
141,
219,
491,
259
]
},
{
"type": "title",
"bbox": [
142,
264,
167,
272
],
"lines": [
{
"bbox": [
141,
263,
168,
273
],
"spans": [
{
"bbox": [
141,
263,
168,
273
],
"score": 1.0,
"content": "2022.3",
"type": "text"
}
],
"index": 16
}
],
"index": 16
},
{
"type": "list",
"bbox": [
142,
273,
486,
365
],
"lines": [
{
"bbox": [
141,
271,
460,
283
],
"spans": [
{
"bbox": [
141,
271,
282,
283
],
"score": 1.0,
"content": "• It can’t recover perfectly from checkpoints",
"type": "text"
},
{
"bbox": [
283,
274,
293,
281
],
"score": 0.64,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
293,
271,
460,
283
],
"score": 1.0,
"content": "Our customized dataloader do not save its state seed",
"type": "text"
}
],
"index": 17,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
147,
282,
246,
293
],
"spans": [
{
"bbox": [
147,
282,
246,
293
],
"score": 1.0,
"content": "properly in distributed training",
"type": "text"
}
],
"index": 18,
"is_list_end_line": true
},
{
"bbox": [
141,
289,
486,
302
],
"spans": [
{
"bbox": [
141,
289,
271,
302
],
"score": 1.0,
"content": "• The memory per processor is too small",
"type": "text"
},
{
"bbox": [
271,
292,
281,
299
],
"score": 0.79,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
281,
289,
385,
302
],
"score": 1.0,
"content": "Require too many pipeline stages",
"type": "text"
},
{
"bbox": [
385,
292,
395,
299
],
"score": 0.75,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
396,
289,
486,
302
],
"score": 1.0,
"content": "Batch size is too large (up to",
"type": "text"
}
],
"index": 19,
"is_list_start_line": true
},
{
"bbox": [
148,
299,
283,
311
],
"spans": [
{
"bbox": [
148,
301,
203,
309
],
"score": 0.36,
"content": "1 2 , 0 0 0 ) \\Rightarrow H a r m",
"type": "inline_equation"
},
{
"bbox": [
203,
299,
283,
311
],
"score": 1.0,
"content": "the model’s convergency",
"type": "text"
}
],
"index": 20,
"is_list_end_line": true
},
{
"bbox": [
141,
307,
473,
321
],
"spans": [
{
"bbox": [
141,
307,
303,
321
],
"score": 1.0,
"content": "• It can’t launch more than 2,000 computing nodes",
"type": "text"
},
{
"bbox": [
304,
311,
314,
317
],
"score": 0.76,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
314,
307,
473,
321
],
"score": 1.0,
"content": "Overcome this and support 6,000-node training by",
"type": "text"
}
],
"index": 21,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
147,
318,
261,
329
],
"spans": [
{
"bbox": [
147,
318,
261,
329
],
"score": 1.0,
"content": "tuning Linux kernel TCP parameters",
"type": "text"
}
],
"index": 22,
"is_list_end_line": true
},
{
"bbox": [
141,
327,
307,
339
],
"spans": [
{
"bbox": [
141,
327,
307,
339
],
"score": 1.0,
"content": "• Collect data for multi-task instruction pre-training",
"type": "text"
}
],
"index": 23,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
141,
336,
349,
348
],
"spans": [
{
"bbox": [
141,
336,
349,
348
],
"score": 1.0,
"content": "• Receive opportunities to test trainings on several other clusters",
"type": "text"
}
],
"index": 24,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
141,
346,
465,
357
],
"spans": [
{
"bbox": [
141,
346,
273,
357
],
"score": 1.0,
"content": "• Very slow training speed than expected",
"type": "text"
},
{
"bbox": [
273,
348,
283,
355
],
"score": 0.76,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
284,
346,
465,
357
],
"score": 1.0,
"content": "The underlying element-wise operators don’t support fast",
"type": "text"
}
],
"index": 25,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
147,
356,
279,
367
],
"spans": [
{
"bbox": [
147,
356,
279,
367
],
"score": 1.0,
"content": "computation on large-dimension vectors.",
"type": "text"
}
],
"index": 26,
"is_list_end_line": true
}
],
"index": 21.5,
"bbox_fs": [
141,
271,
486,
367
]
},
{
"type": "title",
"bbox": [
126,
371,
167,
380
],
"lines": [
{
"bbox": [
141,
371,
168,
381
],
"spans": [
{
"bbox": [
141,
371,
168,
381
],
"score": 1.0,
"content": "2022.4",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "list",
"bbox": [
142,
381,
485,
501
],
"lines": [
{
"bbox": [
141,
380,
447,
392
],
"spans": [
{
"bbox": [
141,
380,
289,
392
],
"score": 1.0,
"content": "• Optimize A100 kernel’s computing efficiency",
"type": "text"
},
{
"bbox": [
289,
382,
299,
389
],
"score": 0.78,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
299,
380,
447,
392
],
"score": 1.0,
"content": "A100 kernels prefer square-shaped inputs, and",
"type": "text"
}
],
"index": 28,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
147,
391,
353,
401
],
"spans": [
{
"bbox": [
147,
391,
353,
401
],
"score": 1.0,
"content": "seq_len=2,048 is optimal for our hidden-state dimension (12,288)",
"type": "text"
}
],
"index": 29,
"is_list_end_line": true
},
{
"bbox": [
142,
399,
482,
410
],
"spans": [
{
"bbox": [
142,
399,
303,
410
],
"score": 1.0,
"content": "• Inability to converge due to large gradient norms",
"type": "text"
},
{
"bbox": [
304,
400,
324,
409
],
"score": 0.6,
"content": "( 1 7 0 + )",
"type": "inline_equation"
},
{
"bbox": [
324,
399,
390,
410
],
"score": 1.0,
"content": "of input embeddings",
"type": "text"
},
{
"bbox": [
391,
400,
412,
408
],
"score": 0.41,
"content": "{ \\Rightarrow T r y }",
"type": "inline_equation"
},
{
"bbox": [
413,
399,
482,
410
],
"score": 1.0,
"content": "embedding norm and",
"type": "text"
}
],
"index": 30,
"is_list_start_line": true
},
{
"bbox": [
146,
408,
322,
420
],
"spans": [
{
"bbox": [
146,
408,
322,
420
],
"score": 1.0,
"content": "gradient shrink, which turn out to be almost equivalent",
"type": "text"
}
],
"index": 31,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
142,
417,
477,
428
],
"spans": [
{
"bbox": [
142,
417,
368,
428
],
"score": 1.0,
"content": "• Naïve post-LN or pre-LN disconverges after several thousands of steps",
"type": "text"
},
{
"bbox": [
368,
419,
390,
427
],
"score": 0.35,
"content": "\\Rightarrow T r y",
"type": "inline_equation"
},
{
"bbox": [
390,
417,
477,
428
],
"score": 1.0,
"content": "Sandwich-LN with PB-Relax",
"type": "text"
}
],
"index": 32,
"is_list_start_line": true
},
{
"bbox": [
141,
425,
487,
438
],
"spans": [
{
"bbox": [
141,
425,
277,
438
],
"score": 1.0,
"content": "• It still disconverges after one week’s trial",
"type": "text"
},
{
"bbox": [
277,
428,
288,
435
],
"score": 0.78,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
288,
425,
487,
438
],
"score": 1.0,
"content": "The dataloader state seeds are not unified for different pipeline",
"type": "text"
}
],
"index": 33,
"is_list_start_line": true
},
{
"bbox": [
147,
437,
323,
446
],
"spans": [
{
"bbox": [
147,
437,
323,
446
],
"score": 1.0,
"content": "stages, resulting in a mismatch of input data and labels.",
"type": "text"
}
],
"index": 34,
"is_list_end_line": true
},
{
"bbox": [
141,
444,
449,
456
],
"spans": [
{
"bbox": [
141,
444,
292,
456
],
"score": 1.0,
"content": "• Test two positional encodings: RoPE and Alibi",
"type": "text"
},
{
"bbox": [
293,
446,
303,
454
],
"score": 0.8,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
303,
444,
449,
456
],
"score": 1.0,
"content": "Alibi can be slower as it requires element-wise",
"type": "text"
}
],
"index": 35,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
147,
454,
435,
466
],
"spans": [
{
"bbox": [
147,
454,
357,
466
],
"score": 1.0,
"content": "manipulation on attention matrices---changing num_heads *2,048",
"type": "text"
},
{
"bbox": [
358,
456,
363,
463
],
"score": 0.3,
"content": "^ *",
"type": "inline_equation"
},
{
"bbox": [
363,
454,
435,
466
],
"score": 1.0,
"content": "2,048 scalars per layer",
"type": "text"
}
],
"index": 36,
"is_list_end_line": true
},
{
"bbox": [
141,
462,
482,
475
],
"spans": [
{
"bbox": [
141,
462,
202,
475
],
"score": 1.0,
"content": "• Test GeGLU and",
"type": "text"
},
{
"bbox": [
202,
465,
244,
472
],
"score": 0.41,
"content": "{ \\mathsf { G A U } } \\ { \\mathsf { \\Omega } } \\Rightarrow { \\mathsf { G A U } }",
"type": "inline_equation"
},
{
"bbox": [
244,
462,
482,
475
],
"score": 1.0,
"content": "converges faster with relatively poor performance on fine-tuned SuperGLUE",
"type": "text"
}
],
"index": 37,
"is_list_start_line": true
},
{
"bbox": [
141,
472,
483,
484
],
"spans": [
{
"bbox": [
141,
472,
362,
484
],
"score": 1.0,
"content": "• Abnormal GPU memory usage of newly-added functions and classes",
"type": "text"
},
{
"bbox": [
362,
474,
372,
482
],
"score": 0.8,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
373,
472,
483,
484
],
"score": 1.0,
"content": "DeepSpeed hardcodes the function",
"type": "text"
}
],
"index": 38,
"is_list_start_line": true
},
{
"bbox": [
147,
482,
251,
493
],
"spans": [
{
"bbox": [
147,
482,
251,
493
],
"score": 1.0,
"content": "names for checkpoint activation",
"type": "text"
}
],
"index": 39,
"is_list_end_line": true
},
{
"bbox": [
142,
492,
441,
502
],
"spans": [
{
"bbox": [
142,
492,
300,
502
],
"score": 1.0,
"content": "• Decide to train GLM with 130 billion parameters",
"type": "text"
},
{
"bbox": [
300,
493,
310,
500
],
"score": 0.77,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
311,
492,
441,
502
],
"score": 1.0,
"content": "allow inference on a DGX-A100 40G node",
"type": "text"
}
],
"index": 40,
"is_list_start_line": true,
"is_list_end_line": true
}
],
"index": 34,
"bbox_fs": [
141,
380,
487,
502
]
},
{
"type": "title",
"bbox": [
142,
507,
173,
515
],
"lines": [
{
"bbox": [
141,
506,
174,
516
],
"spans": [
{
"bbox": [
141,
506,
174,
516
],
"score": 1.0,
"content": "2022.5-6",
"type": "text"
}
],
"index": 41
}
],
"index": 41
},
{
"type": "list",
"bbox": [
142,
516,
495,
600
],
"lines": [
{
"bbox": [
142,
516,
479,
525
],
"spans": [
{
"bbox": [
142,
516,
261,
525
],
"score": 1.0,
"content": "• Implement a RoPE cuda operator in",
"type": "text"
},
{
"bbox": [
261,
517,
284,
525
],
"score": 0.56,
"content": "{ \\mathsf { C } } + + = >",
"type": "inline_equation"
},
{
"bbox": [
285,
516,
479,
525
],
"score": 1.0,
"content": "See unexpected precision errors and finally have it abandoned",
"type": "text"
}
],
"index": 42,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
142,
523,
473,
537
],
"spans": [
{
"bbox": [
142,
523,
245,
537
],
"score": 1.0,
"content": "• Sandwich-LN still disconverges",
"type": "text"
},
{
"bbox": [
245,
526,
259,
534
],
"score": 0.77,
"content": "\\Longrightarrow 1",
"type": "inline_equation"
},
{
"bbox": [
260,
523,
473,
537
],
"score": 1.0,
"content": ") Reducing learning rate does not help; 2) Using Hinge cross-entropy",
"type": "text"
}
],
"index": 43,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
146,
534,
411,
545
],
"spans": [
{
"bbox": [
146,
534,
411,
545
],
"score": 1.0,
"content": "becomes slower and harms performance; 3) Shifting to DeepNorm still disconverges",
"type": "text"
}
],
"index": 44,
"is_list_end_line": true
},
{
"bbox": [
142,
544,
288,
554
],
"spans": [
{
"bbox": [
142,
544,
252,
554
],
"score": 1.0,
"content": "• Use FP32 in softmax of attention",
"type": "text"
},
{
"bbox": [
252,
545,
262,
552
],
"score": 0.81,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
262,
544,
288,
554
],
"score": 1.0,
"content": "Success",
"type": "text"
}
],
"index": 45,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
141,
552,
496,
563
],
"spans": [
{
"bbox": [
141,
552,
288,
563
],
"score": 1.0,
"content": "• Find PB-Relax unnecessary for FP32 softmax",
"type": "text"
},
{
"bbox": [
288,
554,
297,
561
],
"score": 0.41,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
298,
552,
496,
563
],
"score": 1.0,
"content": "It also slows down training as it needs to manipulate the whole",
"type": "text"
}
],
"index": 46,
"is_list_start_line": true
},
{
"bbox": [
147,
563,
226,
572
],
"spans": [
{
"bbox": [
147,
563,
226,
572
],
"score": 1.0,
"content": "attention score matrices",
"type": "text"
}
],
"index": 47,
"is_list_end_line": true
},
{
"bbox": [
143,
572,
491,
582
],
"spans": [
{
"bbox": [
143,
572,
268,
582
],
"score": 1.0,
"content": "• Experience few spikes in later training",
"type": "text"
},
{
"bbox": [
268,
573,
279,
580
],
"score": 0.67,
"content": "\\mathbf { \\Phi } = > \\mathbf { \\Phi } .",
"type": "inline_equation"
},
{
"bbox": [
280,
572,
491,
582
],
"score": 1.0,
"content": "1) Reduce gradient shrink factor from 1 to 0.1: useful; 2) Reduce the",
"type": "text"
}
],
"index": 48,
"is_list_start_line": true
},
{
"bbox": [
147,
580,
407,
591
],
"spans": [
{
"bbox": [
147,
580,
407,
591
],
"score": 1.0,
"content": "learning rate: sometimes useful; 3) Jump the noisy data batches: sometimes useful",
"type": "text"
}
],
"index": 49,
"is_list_end_line": true
},
{
"bbox": [
141,
588,
490,
601
],
"spans": [
{
"bbox": [
141,
588,
346,
601
],
"score": 1.0,
"content": "• Find a mistake in multi-task data after training for 20,000 steps",
"type": "text"
},
{
"bbox": [
346,
591,
356,
599
],
"score": 0.66,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
357,
588,
490,
601
],
"score": 1.0,
"content": "Use the correct data but it does not forget",
"type": "text"
}
],
"index": 50,
"is_list_start_line": true
}
],
"index": 46,
"bbox_fs": [
141,
516,
496,
601
]
},
{
"type": "title",
"bbox": [
142,
605,
173,
613
],
"lines": [
{
"bbox": [
141,
605,
174,
615
],
"spans": [
{
"bbox": [
141,
605,
174,
615
],
"score": 1.0,
"content": "2022.6-7",
"type": "text"
}
],
"index": 51
}
],
"index": 51
},
{
"type": "list",
"bbox": [
142,
615,
498,
700
],
"lines": [
{
"bbox": [
141,
614,
496,
625
],
"spans": [
{
"bbox": [
141,
614,
496,
625
],
"score": 1.0,
"content": "• Adapt the pipeline parallel checkpoints to ordinary parallel checkpoints for efficient inference on a single A100",
"type": "text"
}
],
"index": 52,
"is_list_start_line": true
},
{
"bbox": [
141,
623,
405,
635
],
"spans": [
{
"bbox": [
141,
623,
405,
635
],
"score": 1.0,
"content": "• Work on evaluation scripts on datasets: MMLU, Big-bench, CLUE, SuperCLUE, etc.",
"type": "text"
}
],
"index": 53,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
141,
633,
486,
643
],
"spans": [
{
"bbox": [
141,
633,
486,
643
],
"score": 1.0,
"content": "• Implement P-Tuning and P-Tuning v2 for parameter-efficient tuning on GLM-130B for tuning on SuperGLUE",
"type": "text"
}
],
"index": 54,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
141,
641,
491,
653
],
"spans": [
{
"bbox": [
141,
641,
423,
653
],
"score": 1.0,
"content": "• Work with BMInf on adapting GLM-130B to perform inference on a single V100 or 3090",
"type": "text"
},
{
"bbox": [
423,
643,
433,
650
],
"score": 0.51,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
434,
641,
491,
653
],
"score": 1.0,
"content": "Use pipeline-style",
"type": "text"
}
],
"index": 55,
"is_list_start_line": true
},
{
"bbox": [
147,
651,
354,
662
],
"spans": [
{
"bbox": [
147,
651,
354,
662
],
"score": 1.0,
"content": "asynchronous swapping between main memory and GPU memory",
"type": "text"
}
],
"index": 56,
"is_list_end_line": true
},
{
"bbox": [
141,
659,
488,
672
],
"spans": [
{
"bbox": [
141,
659,
365,
672
],
"score": 1.0,
"content": "• Try to fine-tune GLM-130B with fewer A100 nodes (i.e., 12-16 nodes)",
"type": "text"
},
{
"bbox": [
366,
662,
376,
669
],
"score": 0.79,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
376,
659,
488,
672
],
"score": 1.0,
"content": "Pipeline-style fails due to too many",
"type": "text"
}
],
"index": 57,
"is_list_start_line": true
},
{
"bbox": [
147,
670,
498,
682
],
"spans": [
{
"bbox": [
147,
670,
195,
682
],
"score": 1.0,
"content": "pipeline stages",
"type": "text"
},
{
"bbox": [
196,
671,
206,
679
],
"score": 0.6,
"content": "= >",
"type": "inline_equation"
},
{
"bbox": [
206,
670,
392,
682
],
"score": 1.0,
"content": "Find that data parallel can not be introduced for fine-tuning",
"type": "text"
},
{
"bbox": [
393,
671,
403,
679
],
"score": 0.64,
"content": "\\Rightarrow",
"type": "inline_equation"
},
{
"bbox": [
403,
670,
498,
682
],
"score": 1.0,
"content": "Use 32-way model parallel for",
"type": "text"
}
],
"index": 58
},
{
"bbox": [
147,
679,
279,
690
],
"spans": [
{
"bbox": [
147,
679,
279,
690
],
"score": 1.0,
"content": "fine-tuning with reasonable performance",
"type": "text"
}
],
"index": 59,
"is_list_end_line": true
},
{
"bbox": [
349,
690,
488,
702
],
"spans": [
{
"bbox": [
349,
690,
488,
702
],
"score": 1.0,
"content": "https://github.com/THUDM/GLM-130B",
"type": "text"
}
],
"index": 60
}
],
"index": 56,
"bbox_fs": [
141,
614,
498,
702
]
},
{
"type": "text",
"bbox": [
106,
712,
506,
734
],
"lines": [
{
"bbox": [
105,
711,
506,
725
],
"spans": [
{
"bbox": [
105,
711,
506,
725
],
"score": 1.0,
"content": "Figure 21: The timeline of major issues that training GLM-130B encountered and addressed, as of",
"type": "text"
}
],
"index": 61
},
{
"bbox": [
105,
721,
172,
734
],
"spans": [
{
"bbox": [
105,
721,
172,
734
],
"score": 1.0,
"content": "July 31st, 2022.",
"type": "text"
}
],
"index": 62
}
],
"index": 61.5,
"bbox_fs": [
105,
711,
506,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
107,
82,
504,
127
],
"lines": [
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "faces negligible performance degradation compared to its uncompressed original, while it consumes",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
93,
506,
107
],
"spans": [
{
"bbox": [
105,
93,
127,
107
],
"score": 1.0,
"content": "only",
"type": "text"
},
{
"bbox": [
127,
94,
147,
104
],
"score": 0.87,
"content": "2 5 \\%",
"type": "inline_equation"
},
{
"bbox": [
147,
93,
506,
107
],
"score": 1.0,
"content": "of the GPU memory required by the uncompressed version, thus supporting its effective",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
104,
506,
117
],
"spans": [
{
"bbox": [
105,
104,
158,
117
],
"score": 1.0,
"content": "inference on",
"type": "text"
},
{
"bbox": [
159,
105,
225,
115
],
"score": 0.62,
"content": "4 \\times \\mathrm { R T X } 3 0 9 0 1",
"type": "inline_equation"
},
{
"bbox": [
226,
104,
268,
117
],
"score": 1.0,
"content": "i (24G) or",
"type": "text"
},
{
"bbox": [
268,
105,
341,
115
],
"score": 0.56,
"content": "8 \\times \\mathrm { R T X } 2 0 8 0 \\mathrm { T i }",
"type": "inline_equation"
},
{
"bbox": [
342,
104,
506,
117
],
"score": 1.0,
"content": "(11G). We will attempt to further reduce",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
116,
474,
128
],
"spans": [
{
"bbox": [
106,
116,
474,
128
],
"score": 1.0,
"content": "the resource requirements and keep the community updated on this important working item.",
"type": "text"
}
],
"index": 3
}
],
"index": 1.5
},
{
"type": "title",
"bbox": [
108,
143,
221,
156
],
"lines": [
{
"bbox": [
106,
143,
223,
157
],
"spans": [
{
"bbox": [
106,
143,
223,
157
],
"score": 1.0,
"content": "G BROADER IMPACT",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
107,
168,
505,
257
],
"lines": [
{
"bbox": [
106,
168,
505,
181
],
"spans": [
{
"bbox": [
106,
168,
505,
181
],
"score": 1.0,
"content": "This paper introduces an open bilingual pre-trained language model with 130 billion parameters.",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
179,
505,
191
],
"spans": [
{
"bbox": [
105,
179,
505,
191
],
"score": 1.0,
"content": "Currently most pre-trained language models with over 100 billion parameters are privately owned",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
191,
505,
203
],
"spans": [
{
"bbox": [
106,
191,
505,
203
],
"score": 1.0,
"content": "by governments and large corporations (Brown et al., 2020; Thoppilan et al., 2022; Rae et al., 2021;",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
201,
505,
213
],
"spans": [
{
"bbox": [
106,
201,
505,
213
],
"score": 1.0,
"content": "Chowdhery et al., 2022; Wang et al., 2021). A few of them (Brown et al., 2020; Lieber et al., 2021)",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
212,
505,
225
],
"spans": [
{
"bbox": [
105,
212,
505,
225
],
"score": 1.0,
"content": "provide limited inference APIs with fees. In contrast, the weights and code of GLM-130B are open",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
223,
505,
236
],
"spans": [
{
"bbox": [
106,
223,
505,
236
],
"score": 1.0,
"content": "to anyone who is interested in LLMs. Moreover, we significantly lower the hardware requirements",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
234,
505,
248
],
"spans": [
{
"bbox": [
105,
234,
505,
248
],
"score": 1.0,
"content": "for inference by speed-up implementation and INT4 quantization. The paper can have a broader",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
245,
474,
258
],
"spans": [
{
"bbox": [
106,
245,
474,
258
],
"score": 1.0,
"content": "impact on the research community, individual developers and small companies, and society.",
"type": "text"
}
],
"index": 12
}
],
"index": 8.5
},
{
"type": "title",
"bbox": [
108,
271,
245,
282
],
"lines": [
{
"bbox": [
106,
270,
246,
283
],
"spans": [
{
"bbox": [
106,
270,
246,
283
],
"score": 1.0,
"content": "G.1 IMPACT ON AI RESEARCH",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
107,
292,
505,
380
],
"lines": [
{
"bbox": [
105,
292,
504,
304
],
"spans": [
{
"bbox": [
105,
292,
504,
304
],
"score": 1.0,
"content": "Most research institutions cannot afford the substantial cost of pretraining large language models.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
303,
505,
315
],
"spans": [
{
"bbox": [
106,
303,
505,
315
],
"score": 1.0,
"content": "As a result, most researchers, except employees of governments and large corporations, only have",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
314,
505,
326
],
"spans": [
{
"bbox": [
105,
314,
505,
326
],
"score": 1.0,
"content": "access to the limited inference APIs with fees. With the inference APIs, researchers can only analyze",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
325,
505,
336
],
"spans": [
{
"bbox": [
106,
325,
505,
336
],
"score": 1.0,
"content": "the outputs of models as black boxes, which limits the scope of potential work. With GLM-130B,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
335,
505,
349
],
"spans": [
{
"bbox": [
105,
335,
505,
349
],
"score": 1.0,
"content": "researchers can analyze the model parameters and internal states corresponding to specific inputs,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
347,
505,
358
],
"spans": [
{
"bbox": [
105,
347,
505,
358
],
"score": 1.0,
"content": "leading to in-depth studies of LLMs’ theory, capacity, and flaws. Researchers can also modify the",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
358,
505,
369
],
"spans": [
{
"bbox": [
105,
358,
505,
369
],
"score": 1.0,
"content": "model architecture and weights, to validate the proposed algorithms to improve LLMs Zhu et al.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
369,
375,
380
],
"spans": [
{
"bbox": [
106,
369,
375,
380
],
"score": 1.0,
"content": "(2020); Cao et al. (2021); Hase et al. (2021); Mitchell et al. (2022).",
"type": "text"
}
],
"index": 21
}
],
"index": 17.5
},
{
"type": "text",
"bbox": [
108,
385,
505,
419
],
"lines": [
{
"bbox": [
106,
385,
504,
398
],
"spans": [
{
"bbox": [
106,
385,
465,
398
],
"score": 1.0,
"content": "With INT4 quantization, GLM-130B can perform inference on popularized GPUs such as",
"type": "text"
},
{
"bbox": [
465,
385,
504,
396
],
"score": 0.86,
"content": "4 \\times \\mathrm { R T X }",
"type": "inline_equation"
}
],
"index": 22
},
{
"bbox": [
106,
396,
505,
408
],
"spans": [
{
"bbox": [
106,
396,
139,
408
],
"score": 1.0,
"content": "3090 or",
"type": "text"
},
{
"bbox": [
140,
397,
210,
407
],
"score": 0.33,
"content": "8 \\times \\mathrm { R T X } 2 0 8 0 \\mathrm { T i }",
"type": "inline_equation"
},
{
"bbox": [
210,
396,
505,
408
],
"score": 1.0,
"content": ", which can be easily accessed from cloud service. As a result, researchers",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
408,
502,
420
],
"spans": [
{
"bbox": [
106,
408,
502,
420
],
"score": 1.0,
"content": "who cannot afford powerful data-center GPU servers like DGX-A100 can also utilize GLM-130B.",
"type": "text"
}
],
"index": 24
}
],
"index": 23
},
{
"type": "title",
"bbox": [
108,
433,
404,
444
],
"lines": [
{
"bbox": [
106,
432,
405,
446
],
"spans": [
{
"bbox": [
106,
432,
405,
446
],
"score": 1.0,
"content": "G.2 IMPACT ON INDIVIDUAL DEVELOPERS AND SMALL COMPANIES",
"type": "text"
}
],
"index": 25
}
],
"index": 25
},
{
"type": "text",
"bbox": [
107,
453,
505,
542
],
"lines": [
{
"bbox": [
106,
454,
505,
466
],
"spans": [
{
"bbox": [
106,
454,
505,
466
],
"score": 1.0,
"content": "Currently, individual developers and small companies who want to integrate LLMs into their busi-",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
465,
505,
477
],
"spans": [
{
"bbox": [
105,
465,
505,
477
],
"score": 1.0,
"content": "ness can only choose paid inference APIs. The increased cost can hinder their attempts. Instead,",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
475,
505,
488
],
"spans": [
{
"bbox": [
105,
475,
505,
488
],
"score": 1.0,
"content": "GLM-130B can be deployed on popularized hardware that they own or can access via cloud service",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
487,
505,
499
],
"spans": [
{
"bbox": [
105,
487,
505,
499
],
"score": 1.0,
"content": "to reduce the cost. Furthermore, they can utilize distillation techniques Sanh et al. (2019); Jiao et al.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
497,
505,
510
],
"spans": [
{
"bbox": [
105,
497,
505,
510
],
"score": 1.0,
"content": "(2020) to obtain smaller models that preserve comparable performance on their specific tasks. While",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
509,
505,
521
],
"spans": [
{
"bbox": [
105,
509,
505,
521
],
"score": 1.0,
"content": "some developers may lack the ability to complete deployment and distillation on their own, we be-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
519,
505,
532
],
"spans": [
{
"bbox": [
105,
519,
505,
532
],
"score": 1.0,
"content": "lieve with GLM-130B and more open LLMs in the future, the corresponding toolkits and service",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
531,
262,
542
],
"spans": [
{
"bbox": [
105,
531,
262,
542
],
"score": 1.0,
"content": "providers will become more available.",
"type": "text"
}
],
"index": 33
}
],
"index": 29.5
},
{
"type": "text",
"bbox": [
108,
547,
505,
603
],
"lines": [
{
"bbox": [
106,
547,
505,
560
],
"spans": [
{
"bbox": [
106,
547,
505,
560
],
"score": 1.0,
"content": "We also note that currently most applications of LLMs are based on prompt engineering, partly",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
558,
505,
571
],
"spans": [
{
"bbox": [
106,
558,
505,
571
],
"score": 1.0,
"content": "due to the limitation of inference APIs. In downstream scenarios such as online customer service,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
568,
505,
583
],
"spans": [
{
"bbox": [
105,
568,
505,
583
],
"score": 1.0,
"content": "the companies accumulate huge amounts of human-generated data that contain domain knowledge.",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
580,
505,
593
],
"spans": [
{
"bbox": [
106,
580,
505,
593
],
"score": 1.0,
"content": "With the open-source weights and code, developers can finetune GLM-130B on their own data to",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
591,
264,
605
],
"spans": [
{
"bbox": [
106,
591,
264,
605
],
"score": 1.0,
"content": "mitigate the gap of domain knowledge.",
"type": "text"
}
],
"index": 38
}
],
"index": 36
},
{
"type": "title",
"bbox": [
108,
617,
202,
628
],
"lines": [
{
"bbox": [
105,
615,
204,
630
],
"spans": [
{
"bbox": [
105,
615,
204,
630
],
"score": 1.0,
"content": "G.3 SOCIAL IMPACT",
"type": "text"
}
],
"index": 39
}
],
"index": 39
},
{
"type": "text",
"bbox": [
107,
638,
504,
704
],
"lines": [
{
"bbox": [
105,
636,
506,
651
],
"spans": [
{
"bbox": [
105,
636,
506,
651
],
"score": 1.0,
"content": "Large language models, together with other machine learning models in different modalities (e.g.,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
648,
506,
660
],
"spans": [
{
"bbox": [
105,
648,
506,
660
],
"score": 1.0,
"content": "Image (Ramesh et al., 2021; Ding et al., 2021; Saharia et al.) and Video (Hong et al., 2022)), could",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
660,
505,
672
],
"spans": [
{
"bbox": [
106,
660,
505,
672
],
"score": 1.0,
"content": "be used to generate synthetic text for harmful applications, such as telemarketing fraud, political",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
671,
505,
682
],
"spans": [
{
"bbox": [
105,
671,
505,
682
],
"score": 1.0,
"content": "propaganda, and personal harassment as is discussed in (Weidinger et al., 2021; Sheng et al., 2021;",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
681,
506,
694
],
"spans": [
{
"bbox": [
106,
681,
506,
694
],
"score": 1.0,
"content": "Dev et al., 2021). We do not anticipate any hazardous outputs, especially towards vulnerable and",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
692,
374,
705
],
"spans": [
{
"bbox": [
106,
692,
374,
705
],
"score": 1.0,
"content": "historically disadvantaged groups of people, after using the model.",
"type": "text"
}
],
"index": 45
}
],
"index": 42.5
},
{
"type": "text",
"bbox": [
107,
709,
503,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "While some people think that restricting access to LLMs can prevent such harmful applications, we",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
720,
505,
734
],
"spans": [
{
"bbox": [
105,
720,
505,
734
],
"score": 1.0,
"content": "argue that promoting LLM inclusivity can lead to better defense against potential harm caused by",
"type": "text"
}
],
"index": 47
}
],
"index": 46.5
}
],
"page_idx": 54,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
108,
27,
293,
37
],
"lines": [
{
"bbox": [
106,
26,
293,
38
],
"spans": [
{
"bbox": [
106,
26,
293,
38
],
"score": 1.0,
"content": "Published as a conference paper at ICLR 2023",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
298,
750,
313,
763
],
"spans": [
{
"bbox": [
298,
750,
313,
763
],
"score": 1.0,
"content": "55",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
107,
82,
504,
127
],
"lines": [
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "faces negligible performance degradation compared to its uncompressed original, while it consumes",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
93,
506,
107
],
"spans": [
{
"bbox": [
105,
93,
127,
107
],
"score": 1.0,
"content": "only",
"type": "text"
},
{
"bbox": [
127,
94,
147,
104
],
"score": 0.87,
"content": "2 5 \\%",
"type": "inline_equation"
},
{
"bbox": [
147,
93,
506,
107
],
"score": 1.0,
"content": "of the GPU memory required by the uncompressed version, thus supporting its effective",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
104,
506,
117
],
"spans": [
{
"bbox": [
105,
104,
158,
117
],
"score": 1.0,
"content": "inference on",
"type": "text"
},
{
"bbox": [
159,
105,
225,
115
],
"score": 0.62,
"content": "4 \\times \\mathrm { R T X } 3 0 9 0 1",
"type": "inline_equation"
},
{
"bbox": [
226,
104,
268,
117
],
"score": 1.0,
"content": "i (24G) or",
"type": "text"
},
{
"bbox": [
268,
105,
341,
115
],
"score": 0.56,
"content": "8 \\times \\mathrm { R T X } 2 0 8 0 \\mathrm { T i }",
"type": "inline_equation"
},
{
"bbox": [
342,
104,
506,
117
],
"score": 1.0,
"content": "(11G). We will attempt to further reduce",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
116,
474,
128
],
"spans": [
{
"bbox": [
106,
116,
474,
128
],
"score": 1.0,
"content": "the resource requirements and keep the community updated on this important working item.",
"type": "text"
}
],
"index": 3
}
],
"index": 1.5,
"bbox_fs": [
105,
83,
506,
128
]
},
{
"type": "title",
"bbox": [
108,
143,
221,
156
],
"lines": [
{
"bbox": [
106,
143,
223,
157
],
"spans": [
{
"bbox": [
106,
143,
223,
157
],
"score": 1.0,
"content": "G BROADER IMPACT",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
107,
168,
505,
257
],
"lines": [
{
"bbox": [
106,
168,
505,
181
],
"spans": [
{
"bbox": [
106,
168,
505,
181
],
"score": 1.0,
"content": "This paper introduces an open bilingual pre-trained language model with 130 billion parameters.",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
179,
505,
191
],
"spans": [
{
"bbox": [
105,
179,
505,
191
],
"score": 1.0,
"content": "Currently most pre-trained language models with over 100 billion parameters are privately owned",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
191,
505,
203
],
"spans": [
{
"bbox": [
106,
191,
505,
203
],
"score": 1.0,
"content": "by governments and large corporations (Brown et al., 2020; Thoppilan et al., 2022; Rae et al., 2021;",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
201,
505,
213
],
"spans": [
{
"bbox": [
106,
201,
505,
213
],
"score": 1.0,
"content": "Chowdhery et al., 2022; Wang et al., 2021). A few of them (Brown et al., 2020; Lieber et al., 2021)",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
212,
505,
225
],
"spans": [
{
"bbox": [
105,
212,
505,
225
],
"score": 1.0,
"content": "provide limited inference APIs with fees. In contrast, the weights and code of GLM-130B are open",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
223,
505,
236
],
"spans": [
{
"bbox": [
106,
223,
505,
236
],
"score": 1.0,
"content": "to anyone who is interested in LLMs. Moreover, we significantly lower the hardware requirements",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
234,
505,
248
],
"spans": [
{
"bbox": [
105,
234,
505,
248
],
"score": 1.0,
"content": "for inference by speed-up implementation and INT4 quantization. The paper can have a broader",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
245,
474,
258
],
"spans": [
{
"bbox": [
106,
245,
474,
258
],
"score": 1.0,
"content": "impact on the research community, individual developers and small companies, and society.",
"type": "text"
}
],
"index": 12
}
],
"index": 8.5,
"bbox_fs": [
105,
168,
505,
258
]
},
{
"type": "title",
"bbox": [
108,
271,
245,
282
],
"lines": [
{
"bbox": [
106,
270,
246,
283
],
"spans": [
{
"bbox": [
106,
270,
246,
283
],
"score": 1.0,
"content": "G.1 IMPACT ON AI RESEARCH",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
107,
292,
505,
380
],
"lines": [
{
"bbox": [
105,
292,
504,
304
],
"spans": [
{
"bbox": [
105,
292,
504,
304
],
"score": 1.0,
"content": "Most research institutions cannot afford the substantial cost of pretraining large language models.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
303,
505,
315
],
"spans": [
{
"bbox": [
106,
303,
505,
315
],
"score": 1.0,
"content": "As a result, most researchers, except employees of governments and large corporations, only have",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
314,
505,
326
],
"spans": [
{
"bbox": [
105,
314,
505,
326
],
"score": 1.0,
"content": "access to the limited inference APIs with fees. With the inference APIs, researchers can only analyze",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
325,
505,
336
],
"spans": [
{
"bbox": [
106,
325,
505,
336
],
"score": 1.0,
"content": "the outputs of models as black boxes, which limits the scope of potential work. With GLM-130B,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
335,
505,
349
],
"spans": [
{
"bbox": [
105,
335,
505,
349
],
"score": 1.0,
"content": "researchers can analyze the model parameters and internal states corresponding to specific inputs,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
347,
505,
358
],
"spans": [
{
"bbox": [
105,
347,
505,
358
],
"score": 1.0,
"content": "leading to in-depth studies of LLMs’ theory, capacity, and flaws. Researchers can also modify the",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
358,
505,
369
],
"spans": [
{
"bbox": [
105,
358,
505,
369
],
"score": 1.0,
"content": "model architecture and weights, to validate the proposed algorithms to improve LLMs Zhu et al.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
369,
375,
380
],
"spans": [
{
"bbox": [
106,
369,
375,
380
],
"score": 1.0,
"content": "(2020); Cao et al. (2021); Hase et al. (2021); Mitchell et al. (2022).",
"type": "text"
}
],
"index": 21
}
],
"index": 17.5,
"bbox_fs": [
105,
292,
505,
380
]
},
{
"type": "text",
"bbox": [
108,
385,
505,
419
],
"lines": [
{
"bbox": [
106,
385,
504,
398
],
"spans": [
{
"bbox": [
106,
385,
465,
398
],
"score": 1.0,
"content": "With INT4 quantization, GLM-130B can perform inference on popularized GPUs such as",
"type": "text"
},
{
"bbox": [
465,
385,
504,
396
],
"score": 0.86,
"content": "4 \\times \\mathrm { R T X }",
"type": "inline_equation"
}
],
"index": 22
},
{
"bbox": [
106,
396,
505,
408
],
"spans": [
{
"bbox": [
106,
396,
139,
408
],
"score": 1.0,
"content": "3090 or",
"type": "text"
},
{
"bbox": [
140,
397,
210,
407
],
"score": 0.33,
"content": "8 \\times \\mathrm { R T X } 2 0 8 0 \\mathrm { T i }",
"type": "inline_equation"
},
{
"bbox": [
210,
396,
505,
408
],
"score": 1.0,
"content": ", which can be easily accessed from cloud service. As a result, researchers",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
408,
502,
420
],
"spans": [
{
"bbox": [
106,
408,
502,
420
],
"score": 1.0,
"content": "who cannot afford powerful data-center GPU servers like DGX-A100 can also utilize GLM-130B.",
"type": "text"
}
],
"index": 24
}
],
"index": 23,
"bbox_fs": [
106,
385,
505,
420
]
},
{
"type": "title",
"bbox": [
108,
433,
404,
444
],
"lines": [
{
"bbox": [
106,
432,
405,
446
],
"spans": [
{
"bbox": [
106,
432,
405,
446
],
"score": 1.0,
"content": "G.2 IMPACT ON INDIVIDUAL DEVELOPERS AND SMALL COMPANIES",
"type": "text"
}
],
"index": 25
}
],
"index": 25
},
{
"type": "text",
"bbox": [
107,
453,
505,
542
],
"lines": [
{
"bbox": [
106,
454,
505,
466
],
"spans": [
{
"bbox": [
106,
454,
505,
466
],
"score": 1.0,
"content": "Currently, individual developers and small companies who want to integrate LLMs into their busi-",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
465,
505,
477
],
"spans": [
{
"bbox": [
105,
465,
505,
477
],
"score": 1.0,
"content": "ness can only choose paid inference APIs. The increased cost can hinder their attempts. Instead,",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
475,
505,
488
],
"spans": [
{
"bbox": [
105,
475,
505,
488
],
"score": 1.0,
"content": "GLM-130B can be deployed on popularized hardware that they own or can access via cloud service",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
487,
505,
499
],
"spans": [
{
"bbox": [
105,
487,
505,
499
],
"score": 1.0,
"content": "to reduce the cost. Furthermore, they can utilize distillation techniques Sanh et al. (2019); Jiao et al.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
497,
505,
510
],
"spans": [
{
"bbox": [
105,
497,
505,
510
],
"score": 1.0,
"content": "(2020) to obtain smaller models that preserve comparable performance on their specific tasks. While",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
509,
505,
521
],
"spans": [
{
"bbox": [
105,
509,
505,
521
],
"score": 1.0,
"content": "some developers may lack the ability to complete deployment and distillation on their own, we be-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
519,
505,
532
],
"spans": [
{
"bbox": [
105,
519,
505,
532
],
"score": 1.0,
"content": "lieve with GLM-130B and more open LLMs in the future, the corresponding toolkits and service",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
531,
262,
542
],
"spans": [
{
"bbox": [
105,
531,
262,
542
],
"score": 1.0,
"content": "providers will become more available.",
"type": "text"
}
],
"index": 33
}
],
"index": 29.5,
"bbox_fs": [
105,
454,
505,
542
]
},
{
"type": "text",
"bbox": [
108,
547,
505,
603
],
"lines": [
{
"bbox": [
106,
547,
505,
560
],
"spans": [
{
"bbox": [
106,
547,
505,
560
],
"score": 1.0,
"content": "We also note that currently most applications of LLMs are based on prompt engineering, partly",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
558,
505,
571
],
"spans": [
{
"bbox": [
106,
558,
505,
571
],
"score": 1.0,
"content": "due to the limitation of inference APIs. In downstream scenarios such as online customer service,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
568,
505,
583
],
"spans": [
{
"bbox": [
105,
568,
505,
583
],
"score": 1.0,
"content": "the companies accumulate huge amounts of human-generated data that contain domain knowledge.",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
580,
505,
593
],
"spans": [
{
"bbox": [
106,
580,
505,
593
],
"score": 1.0,
"content": "With the open-source weights and code, developers can finetune GLM-130B on their own data to",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
591,
264,
605
],
"spans": [
{
"bbox": [
106,
591,
264,
605
],
"score": 1.0,
"content": "mitigate the gap of domain knowledge.",
"type": "text"
}
],
"index": 38
}
],
"index": 36,
"bbox_fs": [
105,
547,
505,
605
]
},
{
"type": "title",
"bbox": [
108,
617,
202,
628
],
"lines": [
{
"bbox": [
105,
615,
204,
630
],
"spans": [
{
"bbox": [
105,
615,
204,
630
],
"score": 1.0,
"content": "G.3 SOCIAL IMPACT",
"type": "text"
}
],
"index": 39
}
],
"index": 39
},
{
"type": "text",
"bbox": [
107,
638,
504,
704
],
"lines": [
{
"bbox": [
105,
636,
506,
651
],
"spans": [
{
"bbox": [
105,
636,
506,
651
],
"score": 1.0,
"content": "Large language models, together with other machine learning models in different modalities (e.g.,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
648,
506,
660
],
"spans": [
{
"bbox": [
105,
648,
506,
660
],
"score": 1.0,
"content": "Image (Ramesh et al., 2021; Ding et al., 2021; Saharia et al.) and Video (Hong et al., 2022)), could",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
660,
505,
672
],
"spans": [
{
"bbox": [
106,
660,
505,
672
],
"score": 1.0,
"content": "be used to generate synthetic text for harmful applications, such as telemarketing fraud, political",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
671,
505,
682
],
"spans": [
{
"bbox": [
105,
671,
505,
682
],
"score": 1.0,
"content": "propaganda, and personal harassment as is discussed in (Weidinger et al., 2021; Sheng et al., 2021;",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
681,
506,
694
],
"spans": [
{
"bbox": [
106,
681,
506,
694
],
"score": 1.0,
"content": "Dev et al., 2021). We do not anticipate any hazardous outputs, especially towards vulnerable and",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
692,
374,
705
],
"spans": [
{
"bbox": [
106,
692,
374,
705
],
"score": 1.0,
"content": "historically disadvantaged groups of people, after using the model.",
"type": "text"
}
],
"index": 45
}
],
"index": 42.5,
"bbox_fs": [
105,
636,
506,
705
]
},
{
"type": "text",
"bbox": [
107,
709,
503,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "While some people think that restricting access to LLMs can prevent such harmful applications, we",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
720,
505,
734
],
"spans": [
{
"bbox": [
105,
720,
505,
734
],
"score": 1.0,
"content": "argue that promoting LLM inclusivity can lead to better defense against potential harm caused by",
"type": "text"
}
],
"index": 47
}
],
"index": 46.5,
"bbox_fs": [
105,
709,
505,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
82,
505,
159
],
"lines": [
{
"bbox": [
105,
81,
506,
95
],
"spans": [
{
"bbox": [
105,
81,
506,
95
],
"score": 1.0,
"content": "LLMs. Currently, only governments and large corporations can afford the considerable costs of pre-",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
93,
505,
106
],
"spans": [
{
"bbox": [
105,
93,
505,
106
],
"score": 1.0,
"content": "training LLMs. There is no guarantee that organizations having the substantial financial resources",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
105,
505,
117
],
"spans": [
{
"bbox": [
105,
105,
505,
117
],
"score": 1.0,
"content": "to pretrain an LLM will not do harm with it. Without access to such LLMs, individuals cannot",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
115,
505,
128
],
"spans": [
{
"bbox": [
105,
115,
505,
128
],
"score": 1.0,
"content": "even realize the role of LLMs in harm. Conversely, releasing an open LLM can provide access and",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
127,
505,
138
],
"spans": [
{
"bbox": [
106,
127,
505,
138
],
"score": 1.0,
"content": "transparency to all the researchers and promote the research to reduce the potential harm of LLMs,",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
137,
505,
150
],
"spans": [
{
"bbox": [
105,
137,
505,
150
],
"score": 1.0,
"content": "like algorithms to identify the synthetic text Gehrmann et al. (2019) or detect fake news Li et al.",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
148,
138,
160
],
"spans": [
{
"bbox": [
106,
148,
138,
160
],
"score": 1.0,
"content": "(2021).",
"type": "text"
}
],
"index": 6
}
],
"index": 3
},
{
"type": "text",
"bbox": [
107,
165,
504,
221
],
"lines": [
{
"bbox": [
106,
165,
505,
177
],
"spans": [
{
"bbox": [
106,
165,
505,
177
],
"score": 1.0,
"content": "Also, it is known that LLMs can suffer from problems in fairness, bias, privacy, and truthful-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
175,
505,
189
],
"spans": [
{
"bbox": [
105,
175,
505,
189
],
"score": 1.0,
"content": "ness Zhang et al. (2021); Lin et al. (2022); Liang et al. (2021); Bender et al. (2021). An open",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
187,
505,
199
],
"spans": [
{
"bbox": [
105,
187,
505,
199
],
"score": 1.0,
"content": "LLM can reveal the model parameters and internal states corresponding to specific inputs instead",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
198,
504,
210
],
"spans": [
{
"bbox": [
106,
198,
504,
210
],
"score": 1.0,
"content": "of providing APIs to black-box models. In conclusion, researchers can conduct analysis of LLMs’",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
208,
392,
222
],
"spans": [
{
"bbox": [
105,
208,
392,
222
],
"score": 1.0,
"content": "flaws in depth and propose improved algorithms to solve the problems.",
"type": "text"
}
],
"index": 11
}
],
"index": 9
},
{
"type": "title",
"bbox": [
108,
236,
261,
249
],
"lines": [
{
"bbox": [
106,
237,
263,
250
],
"spans": [
{
"bbox": [
106,
237,
263,
250
],
"score": 1.0,
"content": "H ENVIRONMENTAL IMPACT",
"type": "text"
}
],
"index": 12
}
],
"index": 12
},
{
"type": "text",
"bbox": [
107,
261,
505,
360
],
"lines": [
{
"bbox": [
106,
261,
505,
273
],
"spans": [
{
"bbox": [
106,
261,
505,
273
],
"score": 1.0,
"content": "One of the major concerns about large language models is their huge energy usage and associated",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
271,
504,
283
],
"spans": [
{
"bbox": [
106,
271,
504,
283
],
"score": 1.0,
"content": "carbon emissions Strubell et al. (2019); Lacoste et al. (2019); Patterson et al. (2021); Bender et al.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
283,
504,
295
],
"spans": [
{
"bbox": [
106,
283,
504,
295
],
"score": 1.0,
"content": "(2021). GPT-3 was estimated to use 500 tons of carbon emissions footprint (CO2eq) Patterson et al.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
293,
505,
306
],
"spans": [
{
"bbox": [
106,
293,
505,
306
],
"score": 1.0,
"content": "(2021). We consumed a total of 442.4MWh of electricity over the 60-day course of training. Given",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
304,
505,
317
],
"spans": [
{
"bbox": [
105,
304,
121,
317
],
"score": 1.0,
"content": "the",
"type": "text"
},
{
"bbox": [
122,
305,
186,
316
],
"score": 0.77,
"content": "0 . 5 8 1 0 \\mathrm { ~ k g / k W h }",
"type": "inline_equation"
},
{
"bbox": [
186,
304,
505,
317
],
"score": 1.0,
"content": "carbon efficiency of local power grid, the pre-training released 257.01 metric",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
315,
505,
328
],
"spans": [
{
"bbox": [
105,
315,
136,
328
],
"score": 1.0,
"content": "tons of",
"type": "text"
},
{
"bbox": [
137,
316,
156,
327
],
"score": 0.88,
"content": "\\mathrm { C O _ { 2 } }",
"type": "inline_equation"
},
{
"bbox": [
156,
315,
505,
328
],
"score": 1.0,
"content": ". This is around half of GPT-3’s carbon footprint, probably due to the efficient parallel",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
325,
506,
339
],
"spans": [
{
"bbox": [
105,
325,
506,
339
],
"score": 1.0,
"content": "strategies and NVIDIA’s hardware improvements. The carbon emission is roughly the equivalent of",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
337,
505,
350
],
"spans": [
{
"bbox": [
105,
337,
505,
350
],
"score": 1.0,
"content": "the yearly emissions of 18 average Americans. However, we believe that with GLM-130B released,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
349,
396,
361
],
"spans": [
{
"bbox": [
106,
349,
396,
361
],
"score": 1.0,
"content": "more carbon emissions for reproducing 100B-scale LLMs can be saved.",
"type": "text"
}
],
"index": 21
}
],
"index": 17
}
],
"page_idx": 55,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
108,
27,
293,
37
],
"lines": [
{
"bbox": [
106,
26,
293,
38
],
"spans": [
{
"bbox": [
106,
26,
293,
38
],
"score": 1.0,
"content": "Published as a conference paper at ICLR 2023",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
752,
311,
760
],
"lines": [
{
"bbox": [
298,
750,
313,
763
],
"spans": [
{
"bbox": [
298,
750,
313,
763
],
"score": 1.0,
"content": "",
"type": "text",
"height": 13,
"width": 15
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
106,
82,
505,
159
],
"lines": [
{
"bbox": [
105,
81,
506,
95
],
"spans": [
{
"bbox": [
105,
81,
506,
95
],
"score": 1.0,
"content": "LLMs. Currently, only governments and large corporations can afford the considerable costs of pre-",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
93,
505,
106
],
"spans": [
{
"bbox": [
105,
93,
505,
106
],
"score": 1.0,
"content": "training LLMs. There is no guarantee that organizations having the substantial financial resources",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
105,
505,
117
],
"spans": [
{
"bbox": [
105,
105,
505,
117
],
"score": 1.0,
"content": "to pretrain an LLM will not do harm with it. Without access to such LLMs, individuals cannot",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
115,
505,
128
],
"spans": [
{
"bbox": [
105,
115,
505,
128
],
"score": 1.0,
"content": "even realize the role of LLMs in harm. Conversely, releasing an open LLM can provide access and",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
127,
505,
138
],
"spans": [
{
"bbox": [
106,
127,
505,
138
],
"score": 1.0,
"content": "transparency to all the researchers and promote the research to reduce the potential harm of LLMs,",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
137,
505,
150
],
"spans": [
{
"bbox": [
105,
137,
505,
150
],
"score": 1.0,
"content": "like algorithms to identify the synthetic text Gehrmann et al. (2019) or detect fake news Li et al.",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
148,
138,
160
],
"spans": [
{
"bbox": [
106,
148,
138,
160
],
"score": 1.0,
"content": "(2021).",
"type": "text"
}
],
"index": 6
}
],
"index": 3,
"bbox_fs": [
105,
81,
506,
160
]
},
{
"type": "text",
"bbox": [
107,
165,
504,
221
],
"lines": [
{
"bbox": [
106,
165,
505,
177
],
"spans": [
{
"bbox": [
106,
165,
505,
177
],
"score": 1.0,
"content": "Also, it is known that LLMs can suffer from problems in fairness, bias, privacy, and truthful-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
175,
505,
189
],
"spans": [
{
"bbox": [
105,
175,
505,
189
],
"score": 1.0,
"content": "ness Zhang et al. (2021); Lin et al. (2022); Liang et al. (2021); Bender et al. (2021). An open",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
187,
505,
199
],
"spans": [
{
"bbox": [
105,
187,
505,
199
],
"score": 1.0,
"content": "LLM can reveal the model parameters and internal states corresponding to specific inputs instead",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
198,
504,
210
],
"spans": [
{
"bbox": [
106,
198,
504,
210
],
"score": 1.0,
"content": "of providing APIs to black-box models. In conclusion, researchers can conduct analysis of LLMs’",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
208,
392,
222
],
"spans": [
{
"bbox": [
105,
208,
392,
222
],
"score": 1.0,
"content": "flaws in depth and propose improved algorithms to solve the problems.",
"type": "text"
}
],
"index": 11
}
],
"index": 9,
"bbox_fs": [
105,
165,
505,
222
]
},
{
"type": "title",
"bbox": [
108,
236,
261,
249
],
"lines": [
{
"bbox": [
106,
237,
263,
250
],
"spans": [
{
"bbox": [
106,
237,
263,
250
],
"score": 1.0,
"content": "H ENVIRONMENTAL IMPACT",
"type": "text"
}
],
"index": 12
}
],
"index": 12
},
{
"type": "text",
"bbox": [
107,
261,
505,
360
],
"lines": [
{
"bbox": [
106,
261,
505,
273
],
"spans": [
{
"bbox": [
106,
261,
505,
273
],
"score": 1.0,
"content": "One of the major concerns about large language models is their huge energy usage and associated",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
271,
504,
283
],
"spans": [
{
"bbox": [
106,
271,
504,
283
],
"score": 1.0,
"content": "carbon emissions Strubell et al. (2019); Lacoste et al. (2019); Patterson et al. (2021); Bender et al.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
283,
504,
295
],
"spans": [
{
"bbox": [
106,
283,
504,
295
],
"score": 1.0,
"content": "(2021). GPT-3 was estimated to use 500 tons of carbon emissions footprint (CO2eq) Patterson et al.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
293,
505,
306
],
"spans": [
{
"bbox": [
106,
293,
505,
306
],
"score": 1.0,
"content": "(2021). We consumed a total of 442.4MWh of electricity over the 60-day course of training. Given",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
304,
505,
317
],
"spans": [
{
"bbox": [
105,
304,
121,
317
],
"score": 1.0,
"content": "the",
"type": "text"
},
{
"bbox": [
122,
305,
186,
316
],
"score": 0.77,
"content": "0 . 5 8 1 0 \\mathrm { ~ k g / k W h }",
"type": "inline_equation"
},
{
"bbox": [
186,
304,
505,
317
],
"score": 1.0,
"content": "carbon efficiency of local power grid, the pre-training released 257.01 metric",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
315,
505,
328
],
"spans": [
{
"bbox": [
105,
315,
136,
328
],
"score": 1.0,
"content": "tons of",
"type": "text"
},
{
"bbox": [
137,
316,
156,
327
],
"score": 0.88,
"content": "\\mathrm { C O _ { 2 } }",
"type": "inline_equation"
},
{
"bbox": [
156,
315,
505,
328
],
"score": 1.0,
"content": ". This is around half of GPT-3’s carbon footprint, probably due to the efficient parallel",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
325,
506,
339
],
"spans": [
{
"bbox": [
105,
325,
506,
339
],
"score": 1.0,
"content": "strategies and NVIDIA’s hardware improvements. The carbon emission is roughly the equivalent of",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
337,
505,
350
],
"spans": [
{
"bbox": [
105,
337,
505,
350
],
"score": 1.0,
"content": "the yearly emissions of 18 average Americans. However, we believe that with GLM-130B released,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
349,
396,
361
],
"spans": [
{
"bbox": [
106,
349,
396,
361
],
"score": 1.0,
"content": "more carbon emissions for reproducing 100B-scale LLMs can be saved.",
"type": "text"
}
],
"index": 21
}
],
"index": 17,
"bbox_fs": [
105,
261,
506,
361
]
}
]
}
],
"_backend": "pipeline",
"_version_name": "2.2.2"
}