{"id":"https://openalex.org/W4396822008","doi":"https://doi.org/10.48550/arxiv.2405.05417","title":"Fishing for Magikarp: Automatically Detecting Under-trained Tokens in\n Large Language Models","display_name":"Fishing for Magikarp: Automatically Detecting Under-trained Tokens in\n Large Language Models","publication_year":2024,"publication_date":"2024-05-08","ids":{"openalex":"https://openalex.org/W4396822008","doi":"https://doi.org/10.48550/arxiv.2405.05417"},"language":"en","primary_location":{"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2405.05417","pdf_url":"http://arxiv.org/pdf/2405.05417","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_indexed_in_scopus":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},"type":"preprint","type_crossref":"posted-content","indexed_in":["arxiv"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"http://arxiv.org/pdf/2405.05417","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5085200850","display_name":"Sander Land","orcid":"https://orcid.org/0000-0001-8572-251X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Land, Sander","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5018864006","display_name":"Max Bartolo","orcid":"https://orcid.org/0009-0007-3301-7895"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bartolo, Max","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":0,"citation_normalized_percentile":{"value":0.0,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":{"min":0,"max":82},"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.9969,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.9969,"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/T10028","display_name":"Topic Modeling","score":0.98,"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":[],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.68239903},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.5596835},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.49037957},{"id":"https://openalex.org/C514101110","wikidata":"https://www.wikidata.org/wiki/Q14373","display_name":"Fishing","level":2,"score":0.45313305},{"id":"https://openalex.org/C505870484","wikidata":"https://www.wikidata.org/wiki/Q180538","display_name":"Fishery","level":1,"score":0.11288047},{"id":"https://openalex.org/C86803240","wikidata":"https://www.wikidata.org/wiki/Q420","display_name":"Biology","level":0,"score":0.086387426}],"mesh":[],"locations_count":1,"locations":[{"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2405.05417","pdf_url":"http://arxiv.org/pdf/2405.05417","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_indexed_in_scopus":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}],"best_oa_location":{"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2405.05417","pdf_url":"http://arxiv.org/pdf/2405.05417","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_indexed_in_scopus":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},"sustainable_development_goals":[],"grants":[],"datasets":[],"versions":[],"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4396701345","https://openalex.org/W4391375266","https://openalex.org/W3204019825","https://openalex.org/W2886339410","https://openalex.org/W2748952813","https://openalex.org/W2390279801","https://openalex.org/W2376932109","https://openalex.org/W2358668433","https://openalex.org/W2056369879","https://openalex.org/W2001405890"],"abstract_inverted_index":{"The":[0],"disconnect":[1],"between":[2],"tokenizer":[3,38,91],"creation":[4],"and":[5,84,96,120,127],"model":[6,93],"training":[7],"in":[8,36,46],"language":[9,130],"models":[10,119],"has":[11,63],"been":[12,49,64],"known":[13],"to":[14,25],"allow":[15],"for":[16,103],"certain":[17],"inputs,":[18],"such":[19,30,115],"as":[20],"the":[21,37,112,125],"infamous":[22],"SolidGoldMagikarp":[23],"token,":[24],"induce":[26],"unwanted":[27],"behaviour.":[28],"Although":[29],"`glitch":[31],"tokens'":[32],"that":[33],"are":[34,41],"present":[35,67],"vocabulary,":[39],"but":[40],"nearly":[42],"or":[43],"fully":[44],"absent":[45],"training,":[47],"have":[48],"observed":[50],"across":[51,117],"a":[52,57,68,88],"variety":[53],"of":[54,60,71,81,90,114,129],"different":[55],"models,":[56],"consistent":[58],"way":[59],"identifying":[61],"them":[62],"missing.":[65],"We":[66],"comprehensive":[69],"analysis":[70],"Large":[72],"Language":[73],"Model":[74],"(LLM)":[75],"tokenizers,":[76],"specifically":[77],"targeting":[78],"this":[79],"issue":[80],"detecting":[82,105],"untrained":[83],"under-trained":[85],"tokens.":[86,108],"Through":[87],"combination":[89],"analysis,":[92],"weight-based":[94],"indicators,":[95],"prompting":[97],"techniques,":[98],"we":[99],"develop":[100],"effective":[101],"methods":[102],"automatically":[104],"these":[106],"problematic":[107],"Our":[109],"findings":[110],"demonstrate":[111],"prevalence":[113],"tokens":[116],"various":[118],"provide":[121],"insights":[122],"into":[123],"improving":[124],"efficiency":[126],"safety":[128],"models.":[131]},"cited_by_api_url":"https://api.openalex.org/works?filter=cites:W4396822008","counts_by_year":[],"updated_date":"2025-01-22T14:44:14.561623","created_date":"2024-05-11"}