Optimisation Ratio Odds
Odds Ratio Preference Optimization (ORPO) xeetu tànneef la buy jàngal xeetu làkk bi jikko ju baax ak tànneefi doomi aadama yi ci benn tàggat.
Résumé
It matters because it skips the usual separate reward model and reference model, making alignment cheaper and simpler.
Plongeur bu xóot
ORPO, bi Hong, Lee, ak Thorne dugal ci 2024, dafa boole ajustement buñ saytu ak tànneef ci benn jéego. Tuyo yi gëna bari dañu njëkka def SFT ci misaal yu baax, ba noppi ñu def ñaareelu njuréef bu melni RLHF wala DPO buy laaj kopi buñ jël ci model bi (royuwaay) boole ci ñaari tànneef yuñ denc. ORPO dindi na xeetu royuwaay bi yépp. Perteem dafay yokk benn penalti ci mébetu token bi ci topp: dafay yokk chance yi model bi jox ci tontu biñ tànn (ñu taamu) ci noonu muy wàññi chance yi ñu bañ. Ndax dafay jëfandikoo ratio odd bi moo gën gap log-probabilite bu dëgër, penalti bi yomb na, moo tax model bi dafay jàng bëgg tontu yu baax te baña fàtte generation bu neex.
Gis-gis xarala
Perte ORPO mooy SFT cross-entropy perte boole ci log-sigmoid buñ pondéree ci ratio odd log bi ci digganté tontu yiñ tann ak yiñ bañ. Chance yi ñooy p/(1-p), kon ratio bi dafay méngale ni model bi di gëna am wërsëg am tontu bu baax ak bu baaxul bi. Jëfandikoo coût ci barabu probabilite bu ñor dafay tax contrast bi gëna woyof, loolu mooy tere ñu baña bañ jeton yu bari yi mën yàq model buñu joxewul royuwaay.
njeextalu pexe
Gaawaay ak yaatuwaay
Liggéeyukaay yi ci làkk yi mën nañu gëna gaaw te duñu yàq deggoo gi.
Dugg ak yegg
Dafay yaatal jëfandikoo gi ci làkk yi ak ci anam yi ñuy jokkoo.
dogal yu gëna leer
Ekip yi mën nañu gëna yàgg ci àtte ci jamono ji otomatisation di liggéey ci baamtu.
Ëlëgu Odds Ratio Tanneef
ORPO mingi gëna am doole ndax dafay dagg mémoire bi ak calcul bi ci daaneel modelu royuwaay bi, te loolu luy xëcc ekip yi ñuy fine-tuning ci hardware bu néew. Xaarandil mu gëna bari lumuy feeñ ci rëset yu ubbeeku yi ak muy tànneef buñ jagleel ci bibliotek yu melni Hugging Face TRL. Liggéey yi ci kanam dina ñu defar pondération lambda ci saasi, jaxase ORPO ak yeneen mébet yu amul royuwaay, ba noppi yokk ko ci model yu bari te yaatu lool, fu tëye ñaari kopi ci mémoire lu seer la.
Doxal ci àdduna dëgg
Defar bu baax benn xeetu waxtaan 7B bu ubbeeku ci ñaari tànneef te doo sarse ñaareelu kopi royuwaay, di xaaj mémoire GPU
Ab startup buy boole ab assistant buy jàppale kiliyaan yi ngir mu taamu tontu yu rafet te lalu ci politik ci benn tàggat ci palaasu SFT-then-DPO
Gëstukat yi dañuy méngale ORPO ak DPO ci benn done ngir wane ni dañu méngoo ak xayma yu gëna ndaw
Defar ab xeetu basis ci benn domen buñ jagleel (lu melni, bind yoon) fu misaal yu baax ak yu baaxul am waaye budget xeetu neexal amul
Risk yi ak balustrade yi
Lépp lu jaarul yoon mën na dugg ci rapoor yi, jàppale ci liggéey bi, wala ci njariñu gëstu bi.
Sensibilite bu gaaw mën na jur njariñ yu wuute ci laajte yu noonu mel.
Done yu am solo mën nañu feeñ sudee seytu jëfandikoo gi néew doole.
Roadmap ngir samp gi
Mandargal formaa génne gi, melokaan bi, ak standard kalite yi laata ngay dugal ko.
Tontu yu am solo ak balluwaay yu wóor saa yu dëggu bi di am solo.
Fexeel am barabu xool nit ñi ngir am njariñ yu am solo.
Toppal anami gacce yi ak di faral di tàggataat ay laaj wala def-liggéey.
Weyal di banneexu
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 Odds Ratio Preference Optimization 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
Gis bi ci topp
Optimisation taamu direct
Laaj yi ñuy faral di laaj
What is Odds Ratio Preference Optimization?
Odds Ratio Preference Optimization (ORPO) xeetu tànneef la buy jàngal xeetu làkk bi jikko ju baax ak tànneefi doomi aadama yi ci benn tàggat. Dafa am solo ndax dafay salte xeetu neexal bi ñuy faral di tàqale ak xeetu royuwaay, gëna yombal liggéey bi.
Lan mooy njariñu ORPO ci anam yu melni DPO?
ORPO dafay boole jàng yiñ taamu ci ñàkkum SFT te soxlawul ab kopi royuwaay buñ gelé ci model bi, di sakkanal mémoire ak calcul.
Lan mooy 'odds ratio' bi ci ORPO?
ORPO dafay jëfandikoo ratio odd (p/(1-p)) bi model bi jox tontu biñ taamu ak biñu taamuwul.
Ban ñaari liggéey la ORPO di def ci benn pas tàggat?
ORPO dafa boole mébetu SFT bi ci topp ak penalti buñ taamu ci benn ñàkk buñ boole.
Lan moo tax ORPO di jëfandikoo coût yi moo gën différence log-probabilite bu ñor ngir penalti bi?
Penalty bi lalu ci chance yi moo gëna woyof, di jàppale model biñu joxul royuwaay bi mu wéy di génne ay tontu yu baax te ba leegi mingi taamu tontu yu baax.
Ban boole lañu gëna mëna wax ni ORPO ñàkk?
ORPO dafay yokk benn log-sigmoid buñ pondéree ci kaw entropi buñ saytu biy ñàkk.