{"id":"https://openalex.org/W4392019816","doi":"https://doi.org/10.48550/arxiv.2402.13228","title":"Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive","display_name":"Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive","publication_year":2024,"publication_date":"2024-02-20","ids":{"openalex":"https://openalex.org/W4392019816","doi":"https://doi.org/10.48550/arxiv.2402.13228"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2402.13228","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2402.13228","pdf_url":"https://arxiv.org/pdf/2402.13228","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by-nc","license_id":"https://openalex.org/licenses/cc-by-nc","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2402.13228","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5102528281","display_name":"Arka Pal","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pal, Arka","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5045970638","display_name":"Deep Karkhanis","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Karkhanis, Deep","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5063355813","display_name":"Samuel W. Dooley","orcid":"https://orcid.org/0000-0001-8598-7681","observed_orcids":["https://orcid.org/0000-0001-8598-7681"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dooley, Samuel","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5041830853","display_name":"Manley Roberts","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Roberts, Manley","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5018487743","display_name":"Siddartha Naidu","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Naidu, Siddartha","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":null,"display_name":"White, Colin","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"White, Colin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":6,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.7559000253677368,"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":0.7559000253677368,"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":0.12770000100135803,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.045099999755620956,"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/loss-function-design","display_name":"loss function design","score":0.7710000276565552},{"id":"https://openalex.org/keywords/llm-fine-tuning","display_name":"LLM fine-tuning","score":0.8209999799728394},{"id":"https://openalex.org/keywords/smaug","display_name":"Smaug","score":0.6100000143051147},{"id":"https://openalex.org/keywords/open-source-large-language-models","display_name":"open-source large language models","score":0.8659999966621399},{"id":"https://openalex.org/keywords/mt-bench","display_name":"MT-Bench","score":0.5830000042915344}],"concepts":[{"id":"https://openalex.org/C2781249084","wikidata":"https://www.wikidata.org/wiki/Q908656","display_name":"Preference","level":2,"score":0.6091527342796326},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.3409738540649414},{"id":"https://openalex.org/C175444787","wikidata":"https://www.wikidata.org/wiki/Q39072","display_name":"Microeconomics","level":1,"score":0.22600647807121277}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2402.13228","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2402.13228","pdf_url":"https://arxiv.org/pdf/2402.13228","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by-nc","license_id":"https://openalex.org/licenses/cc-by-nc","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2402.13228","doi":"https://doi.org/10.48550/arxiv.2402.13228","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2402.13228","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2402.13228","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2402.13228","pdf_url":"https://arxiv.org/pdf/2402.13228","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by-nc","license_id":"https://openalex.org/licenses/cc-by-nc","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4392019816.pdf","grobid_xml":"https://content.openalex.org/works/W4392019816.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4402670121","https://openalex.org/W4412889550","https://openalex.org/W4416073047","https://openalex.org/W4416342251","https://openalex.org/W7164498433","https://openalex.org/W4414586725","https://openalex.org/W4417348360","https://openalex.org/W4406491496","https://openalex.org/W4412888609","https://openalex.org/W4403555955"],"abstract_inverted_index":{"Direct":[0],"Preference":[1],"Optimisation":[2],"(DPO)":[3],"is":[4,106],"effective":[5],"at":[6],"significantly":[7],"improving":[8],"the":[9,34,51,61,65,71,75,99,159,163,173,190,193,205],"performance":[10],"of":[11,27,37,60,64,104,142,172,202],"large":[12],"language":[13],"models":[14,33],"(LLMs)":[15],"on":[16,92,169,204],"downstream":[17,145],"tasks":[18],"such":[19,176],"as":[20,68,70,177],"reasoning,":[21],"summarisation,":[22],"and":[23,29,77,119,134,144,184,187],"alignment.":[24],"Using":[25,108],"pairs":[26,103],"preferred":[28,66,76],"dispreferred":[30,78],"data,":[31,175],"DPO":[32,53,133],"relative":[35,72],"probability":[36,73],"picking":[38],"one":[39],"response":[40],"over":[41],"another.":[42],"In":[43],"this":[44,86,124],"work,":[45],"first":[46,194],"we":[47,111,128,156,182],"show":[48,83],"theoretically":[49],"that":[50,85,130,158],"standard":[52],"loss":[54,117],"can":[55],"lead":[56],"to":[57,197],"a":[58,115,139],"reduction":[59],"model's":[62],"likelihood":[63],"examples,":[67],"long":[69],"between":[74,102,153],"classes":[79],"increases.":[80],"We":[81],"then":[82],"empirically":[84],"phenomenon":[87],"occurs":[88],"when":[89],"fine-tuning":[90,136,174],"LLMs":[91],"common":[93],"datasets,":[94],"especially":[95],"datasets":[96,143,148],"in":[97],"which":[98,122],"edit":[100,151],"distance":[101],"completions":[105],"low.":[107],"these":[109],"insights,":[110],"design":[112],"DPO-Positive":[113],"(DPOP),":[114],"new":[116],"function":[118],"training":[120],"procedure":[121],"avoids":[123],"failure":[125],"mode.":[126],"Surprisingly,":[127],"find":[129,157],"DPOP":[131],"outperforms":[132,162],"other":[135],"procedures":[137],"across":[138],"wide":[140],"variety":[141],"tasks,":[146],"including":[147],"with":[149,189],"high":[150],"distances":[152],"completions.":[154],"Furthermore,":[155],"DPOP-tuned":[160],"model":[161,165],"DPO-tuned":[164],"(all":[166],"else":[167],"equal)":[168],"benchmarks":[170],"independent":[171],"MT-Bench.":[178],"Finally,":[179],"using":[180],"DPOP,":[181],"create":[183],"open-source":[185,195],"Smaug-34B":[186],"Smaug-72B,":[188],"latter":[191],"becoming":[192],"LLM":[196,208],"surpass":[198],"an":[199],"average":[200],"accuracy":[201],"80%":[203],"HuggingFace":[206],"Open":[207],"Leaderboard.":[209]},"counts_by_year":[{"year":2026,"cited_by_count":1},{"year":2025,"cited_by_count":5}],"updated_date":"2026-09-27T08:07:27.782135","created_date":"2024-02-22T00:00:00"}