YaRN na Context Uburebure
YaRN (Nyamara ubundi RoPE extensioN) nubuhanga bukora neza bwo kurambura imiterere yimiterere yimiterere yimiterere irenze kure ibyo yatojwe.
Incamake
It cleverly rescales rotary position embeddings so a model trained on, say, 4K tokens can handle 32K or more with minimal fine-tuning.
Kwibira cyane
Ibyinshi muri LLM bigezweho byerekana ibimenyetso hamwe na RoPE (Rotary Position Embeddings), izenguruka ikibazo hamwe nurufunguzo rwibanze kumpande zifatanije numwanya. Iyo ugaburira urutonde rurerure kuruta uburebure bwamahugurwa, ibyo bizunguruka byinjira mubitagaragara kandi icyitegererezo kiracika. YaRN, yatangijwe mu 2023 na Bowen Peng nabafatanyabikorwa, ikosora ibi hamwe na NTK-izi interpolation ikoreshwa kuri frequence: isiga ibipimo byinshyi nyinshi (bifata umubano waho, intera ngufi) ahanini idakorwa mugihe ihuza interuro ntoya (ikurikirana umwanya muremure). YaRN kandi yongeraho ubushyuhe bwo guhindura kugirango yitondere guhangana na entropiya ihinduka kuva murwego rurerure. Igisubizo ni imbaraga ndende-mikorere nyuma yo guhuza neza agace gato kamakuru hamwe nintambwe uburyo bworoshye busaba.
Ubushishozi
RoPE igenera buri gushyiramo ibipimo byo kuzenguruka inshuro. Naive umurongo interpolation ikomatanya imirongo yose uko ingana, yangiza ibipimo byinshi-bigizwe nibisobanuro byiza byaho. YaRN ikoresha imikorere ya ramp kugirango ihuze gusa ibipimo bito-birebire (birebire-birebire) mugihe irinze izifite umurongo mwinshi, hiyongereyeho igipimo cya 1 / sqrt (t) cyerekana ubushyuhe butuma ubukana bwa softmax butajegajega uko uburebure bukurikirana. Ubu buryo bwa NTK-by-ibice byagura imiterere hamwe no gutesha agaciro cyane.
Ingaruka z'Ingamba
Igiciro na bije
Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka.
Ibyemezo bisobanutse
Ubuhanga bwa tekinike bufasha amakipe guhitamo umurongo ukwiye, ntabwo ari shyashya gusa.
Kugenzura ubuziranenge
Guhitamo neza bya injeniyeri bigabanya ibintu byizewe mubikorwa.
Ejo hazaza ha YaRN no Kwagura Uburebure
Kwagura imvugo ni imyitozo isanzwe: gufungura imiterere isanzwe yohereza YaRN-yaguwe ihinduka igera kuri 128K ibimenyetso cyangwa birenze. Ubushakashatsi bugenda bwerekeza kuburyo bwaguka imiterere hamwe na zeru cyangwa hafi ya zeru neza-guhuza, guhuza RoPE ikuraho amayeri-yerekana amayeri, kandi igakomeza ubuziranenge hejuru yidirishya ryuzuye aho kuba impera gusa. Tegereza guhuza ubu buhanga muburyo bwo kwerekana imiterere ndende ni kavukire aho guhindurwa.
Gushyira mu bikorwa Isi
Kwagura moderi ifunguye ya 4K kuri 32K cyangwa 128K kubibazo byinyandiko ndende isubiza hamwe neza-neza
Gushoboza kugarura-kongera sisitemu yo kwinjiza ibice byinshi bihujwe nta guca
Imbaraga za kode zifasha zikeneye dosiye nini yububiko cyangwa dosiye nyinshi mugihe kimwe
Guhuza icyitegererezo cyibanze kubiganiro byinshi birebire bikusanya amateka manini yo kuganira
Ingaruka & Kurinda
Gutezimbere igipimo kimwe gishobora guhisha intege nke za sisitemu.
Ibikorwa Remezo no kubungabunga akenshi usanga bidahabwa agaciro.
Icyuho cyumutekano no kwitegereza birashobora kwiyongera uko sisitemu igenda igorana.
Igishushanyo mbonera
Sobanura ubukererwe, ubuziranenge, nigiciro cyibiciro mbere yo kubishyira mubikorwa.
Ibipimo byerekana umutwaro ufatika hamwe namakuru yimiterere.
Gukurikirana ibikoresho kubikosa, drift, ningaruka zabakoresha.
Tegura inzira yo gusubiza ibyabaye mbere yo gupima.
Komeza Ubushakashatsi
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the YaRN and Context Length Extension quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ubuyobozi bukurikira
Umwanya Interpolation yo Kwagura Imirongo
Ibibazo bikunze kubazwa
What is YaRN and Context Length Extension?
YaRN (Nyamara ubundi RoPE extensioN) nubuhanga bukora neza bwo kurambura imiterere yimiterere yimiterere yimiterere irenze kure ibyo yatojwe. Ihinduranya ubushishozi ikizunguruka cyumwanya kugirango moderi yatojwe, tuvuge, ibimenyetso 4K bishobora gukora 32K cyangwa birenga hamwe na tuning nziza.
Ni ubuhe buryo-kodegisi ya YaRN ihindura kugirango yongere imiterere?
YaRN, izina ryayo risobanura Ubundi RoPE extensioN, ikuraho imyanya izenguruka ikoreshwa muri LLM nyinshi zigezweho.
Niki kitagenda neza mugihe moderi ya RoPE ibonye urukurikirane rurerure kuruta imyitozo yarwo?
Imyanya irenze imyitozo itanga inguni yo kuzenguruka icyitegererezo kitigeze kibona, bityo imyitwarire yo kwitondera igabanuka cyane.
Nigute YaRN ifata ibipimo bitandukanye bya RoPE?
YaRN ikoresha NTK-by-ibice byambukiranya interpolate intera ndende-yumurambararo muto kandi igasiga intera ngufi-nini cyane cyane idahwitse.
Usibye gukuraho imirongo, ni ubuhe buryo bwinyongera YaRN ikoresha?
YaRN yongeramo ubushyuhe bwo gupima ibitekerezo kugirango uhangane na entropie ihinduka ibaho uko imiterere ikura.
Ni izihe nyungu zingenzi za YaRN kurenza na interpolation?
YaRN igera ku ntera ndende-ndende nyuma yo guhuza neza agace gato k'amakuru yuburyo bworoshye busaba.