{"id":"https://openalex.org/W4330337916","doi":"https://doi.org/10.48550/arxiv.2303.10464","title":"SPDF: Sparse Pre-training and Dense Fine-tuning for Large Language Models","display_name":"SPDF: Sparse Pre-training and Dense Fine-tuning for Large Language Models","publication_year":2023,"publication_date":"2023-01-01","ids":{"openalex":"https://openalex.org/W4330337916","doi":"https://doi.org/10.48550/arxiv.2303.10464"},"language":"en","primary_location":{"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2303.10464","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":["Cornell University"],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false},"type":"preprint","type_crossref":"posted-content","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/abs/2303.10464","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5022019488","display_name":"Vithursan Thangarasa","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Thangarasa, Vithursan","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5015243339","display_name":"Abhay Gupta","orcid":"https://orcid.org/0000-0003-4767-4865"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Gupta, Abhay","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5015216742","display_name":"William Marshall","orcid":"https://orcid.org/0000-0002-3520-4500"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Marshall, William","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5032073184","display_name":"Tianda Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Tianda","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5102722010","display_name":"K.M.K.H. Leong","orcid":"https://orcid.org/0000-0002-5968-4342"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Leong, Kevin","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5108368276","display_name":"Dennis DeCoste","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"DeCoste, Dennis","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5001174692","display_name":"Sean Lie","orcid":"https://orcid.org/0000-0001-6074-4640"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lie, Sean","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5025685148","display_name":"Shreyas Saxena","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Saxena, Shreyas","raw_affiliation_strings":[],"affiliations":[]}],"institution_assertions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":4,"citation_normalized_percentile":{"value":0.787004,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":{"min":88,"max":91},"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.9994,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.9994,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.999,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10201","display_name":"Speech Recognition and Synthesis","score":0.9589,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/flops","display_name":"FLOPS","score":0.849084},{"id":"https://openalex.org/keywords/language-modeling","display_name":"Language Modeling","score":0.584446},{"id":"https://openalex.org/keywords/topic-modeling","display_name":"Topic Modeling","score":0.546858},{"id":"https://openalex.org/keywords/statistical-language-modeling","display_name":"Statistical Language Modeling","score":0.542096},{"id":"https://openalex.org/keywords/neural-machine-translation","display_name":"Neural Machine Translation","score":0.534699},{"id":"https://openalex.org/keywords/pretrained-models","display_name":"Pretrained Models","score":0.527398},{"id":"https://openalex.org/keywords/natural-language-generation","display_name":"Natural language generation","score":0.4650928}],"concepts":[{"id":"https://openalex.org/C3826847","wikidata":"https://www.wikidata.org/wiki/Q188768","display_name":"FLOPS","level":2,"score":0.849084},{"id":"https://openalex.org/C170858558","wikidata":"https://www.wikidata.org/wiki/Q1394144","display_name":"Automatic summarization","level":2,"score":0.788881},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.76362664},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.68072236},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.5644434},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.48859316},{"id":"https://openalex.org/C2776187449","wikidata":"https://www.wikidata.org/wiki/Q1513879","display_name":"Natural language generation","level":3,"score":0.4650928},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.46271464},{"id":"https://openalex.org/C36503486","wikidata":"https://www.wikidata.org/wiki/Q11235244","display_name":"Domain (mathematical analysis)","level":2,"score":0.4362526},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.42689174},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4023286},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.221845},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.1670267},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.080504805},{"id":"https://openalex.org/C134306372","wikidata":"https://www.wikidata.org/wiki/Q7754","display_name":"Mathematical analysis","level":1,"score":0.0},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.0},{"id":"https://openalex.org/C2524010","wikidata":"https://www.wikidata.org/wiki/Q8087","display_name":"Geometry","level":1,"score":0.0},{"id":"https://openalex.org/C187736073","wikidata":"https://www.wikidata.org/wiki/Q2920921","display_name":"Management","level":1,"score":0.0},{"id":"https://openalex.org/C153294291","wikidata":"https://www.wikidata.org/wiki/Q25261","display_name":"Meteorology","level":1,"score":0.0},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.0}],"mesh":[],"locations_count":3,"locations":[{"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2303.10464","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":["Cornell University"],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false},{"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2303.10464","pdf_url":"http://arxiv.org/pdf/2303.10464","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":["Cornell University"],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false},{"is_oa":false,"landing_page_url":"https://api.datacite.org/dois/10.48550/arxiv.2303.10464","pdf_url":null,"source":{"id":"https://openalex.org/S4393179698","display_name":"DataCite API","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I4210145204","host_organization_name":"DataCite","host_organization_lineage":["https://openalex.org/I4210145204"],"host_organization_lineage_names":["DataCite"],"type":"metadata"},"license":null,"license_id":null,"version":null}],"best_oa_location":{"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2303.10464","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":["Cornell University"],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false},"sustainable_development_goals":[{"score":0.83,"id":"https://metadata.un.org/sdg/4","display_name":"Quality education"}],"grants":[],"datasets":[],"versions":[],"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4323520239","https://openalex.org/W4242223894","https://openalex.org/W3148229873","https://openalex.org/W3034878914","https://openalex.org/W3019688513","https://openalex.org/W2366403280","https://openalex.org/W2150160875","https://openalex.org/W2091301346","https://openalex.org/W1517524280","https://openalex.org/W1495108544"],"abstract_inverted_index":{"The":[0],"pre-training":[1,145,186],"and":[2,40,56,87,117,121,148,218],"fine-tuning":[3,86],"paradigm":[4],"has":[5,59],"contributed":[6],"to":[7,71,108,137,159,170,199,227],"a":[8,22,140,174,182,189,212,224,233],"number":[9],"of":[10,18,64,81,132,142,235,246],"breakthroughs":[11],"in":[12,181,185,192],"Natural":[13],"Language":[14],"Processing":[15],"(NLP).":[16],"Instead":[17],"directly":[19],"training":[20,101,104,237],"on":[21,30,43,194],"downstream":[23,196,207,251],"task,":[24],"language":[25,48],"models":[26,231],"are":[27],"first":[28],"pre-trained":[29,247],"large":[31,229],"datasets":[32],"with":[33],"cross-domain":[34],"knowledge":[35],"(e.g.,":[36,46],"Pile,":[37],"MassiveText,":[38],"etc.)":[39],"then":[41,149],"fine-tuned":[42],"task-specific":[44],"data":[45],"natural":[47],"generation,":[49],"text":[50],"summarization,":[51],"etc.).":[52],"Scaling":[53],"the":[54,62,88,93,96,110,114,130,151,156,195,200,236,244],"model":[55,89,111,179],"dataset":[57,219],"size":[58],"helped":[60],"improve":[61],"performance":[63],"LLMs,":[65],"but":[66],"unfortunately,":[67],"this":[68,126],"also":[69,210],"lead":[70],"highly":[72],"prohibitive":[73],"computational":[74],"costs.":[75],"Pre-training":[76,120],"LLMs":[77],"often":[78,91],"require":[79],"orders":[80],"magnitude":[82],"more":[83],"FLOPs":[84,238],"than":[85],"capacity":[90,112,153],"remains":[92],"same":[94],"between":[95,113,214],"two":[97,115],"phases.":[98],"To":[99],"achieve":[100],"efficiency":[102],"w.r.t":[103],"FLOPs,":[105,187],"we":[106,128,166,209],"propose":[107],"decouple":[109],"phases":[116],"introduce":[118],"Sparse":[119],"Dense":[122],"Fine-tuning":[123],"(SPDF).":[124],"In":[125],"work,":[127],"show":[129],"benefits":[131,245],"using":[133,239],"unstructured":[134],"weight":[135,240],"sparsity":[136,172],"train":[138,228],"only":[139],"subset":[141],"weights":[143,158],"during":[144],"(Sparse":[146],"Pre-training)":[147],"recover":[150],"representational":[152],"by":[154],"allowing":[155],"zeroed":[157],"learn":[160],"(Dense":[161],"Fine-tuning).":[162],"We":[163],"demonstrate":[164],"that":[165],"can":[167],"induce":[168],"up":[169],"75%":[171],"into":[173],"1.3B":[175],"parameter":[176],"GPT-3":[177],"XL":[178],"resulting":[180],"2.5x":[183],"reduction":[184],"without":[188],"significant":[190],"loss":[191],"accuracy":[193],"tasks":[197],"relative":[198],"dense":[201],"baseline.":[202],"By":[203],"rigorously":[204],"evaluating":[205],"multiple":[206],"tasks,":[208],"establish":[211],"relationship":[213],"sparsity,":[215,241],"task":[216],"complexity":[217],"size.":[220],"Our":[221],"work":[222],"presents":[223],"promising":[225],"direction":[226],"GPT":[230],"at":[232],"fraction":[234],"while":[242],"retaining":[243],"textual":[248],"representations":[249],"for":[250],"tasks.":[252]},"cited_by_api_url":"https://api.openalex.org/works?filter=cites:W4330337916","counts_by_year":[{"year":2024,"cited_by_count":4}],"updated_date":"2024-12-05T07:44:08.614783","created_date":"2023-03-22"}