Réjection de échantillonnage de rejection
Rejection Sampling Fine-Tuning (RFT) dafay defar tontu yu bari ci kàndidaa yi, bàyyi ci ñi gëna am poñ, ba noppi tàggataat xeetu tontu yooyu ci ñi jël raw gàddu gi.
Résumé
It matters because it offers much of RLHF's benefit using straightforward supervised learning instead of complex reinforcement learning.
Plongeur bu xóot
Bañ jël misaal Reglage bu baax, yenn saa yi ñu koy woowe reglage bu gëna baax ci N, bokk na ci li gëna am solo ci ni ñu defaree model yu melni Meta's Llama 2 ak Llama 3. Reset bi yomb na: ci laaj bu nekk, jëlal misaalu tontu yu bari (maanaam 4 ba 64) ci xeetu neexal bi fi nekk, nga jox poñ bu nekk ak xeetu neexal wala xoolkat bu otomatik, ba noppi sànni ('bañ') lépp lu moy génne yi gëna kawe. Misali kalite yu kawe yi mucc nekk nañu ay done yu bees yuñ saytu, ba noppi ñu tàggat model bi ci ñoom ak ñàkka am token bu ci topp. Baamtu loop bii dafay tax model bi génne tontu yu gëna baax ci boppam. Ndax model bi dafay jàngee ci ay génnam yuñ segg, RFT moytu instabilite ak tuning boppu buy metti ci politik-gradient RL ci di jëfandikoo siñaal neexal.
Gis-gis xarala
RFT dafay jëfandikoo li sampling yoon yu bari ak tëye tontub neexal bu gëna mag biy jege tànneef ci distribution bu ñaw, bu gëna baax. Taggat ci ñi jël raw-gàddu gi jaaraleko ci cross-entropy buñ miin, dafay distill bu baax jikko bu gëna baax ci N ci génnug benn misaalu model bi. Ci wàll yiñ mëna saytu lu ci melni math wala kode, 'neexal' bi mën na nekk tontu bu mujj bi wala test unité bi jàll, dindi soxla xeetu neexal buñ jàng lëmm.
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 Rejection Sampling Fine-Tuning
RFT lu am solo la ci tàggat-yaram bu bees bi, ñu koy faral di jëfandikoo balaa wala ci wetu pexe RL yu melni PPO ak DPO. Appel bi dafay màgg ndax inference yu yomb ak verifier otomatik yu dëgër: ginaaw bi model yi gënee xarañ ci defar seen bopp ak saytu seen bopp, sampling rejection buy baamtu dafay jàppale done synthetik ak bouclage yuy gëna suqali seen bopp. Xaarandil lëkkaloo bu gëna dëgër ak xeetu xalaat yuy defar ay xalaat yuñ mëna saytu, ak jàngat buy wéy ci ni ñuy moytu piratage neexal ak daanu bu wuute suñuy tàggat lu bari ci génnug xeetu boppam.
Doxal ci àdduna dëgg
Liggéeyukaay xeetu Llama ci sampling tontu yu bari ci laaj bu nekk, tëye poñ yi gëna rëy ci xeetu neexal, ba noppi SFT ci ñooñu
Yokkateg ab solver math ci defar ay solution yu bari te bàyyi ci tontu bu jaar yoon, buñu mëna saytu
Defar kodu fu ñuy denc kàndidaa yi sudee jàll nañu examen yi, ñu jëfandikoo leen ni done tàggat
Tabax ay done yuñ defaree ay tegtal ci segg tontu yu gëna baax yu model bi defar boppam ngir tàggat yaram bi ci topp
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 Rejection Sampling Fine-Tuning 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
QLoRA ak 4-bit
Laaj yi ñuy faral di laaj
What is Rejection Sampling Fine-Tuning?
Rejection Sampling Fine-Tuning (RFT) dafay defar tontu yu bari ci kàndidaa yi, bàyyi ci ñi gëna am poñ, ba noppi tàggataat xeetu tontu yooyu ci ñi jël raw gàddu gi. Dafa am solo ndax dafay joxe njariñ yu bari ci RLHF ci jëfandikoo njàng mu jub te ñu yor ko ci palaasu njàng mu jafee dooleel.
Lan mooy xalaat bu mag bi ci Rejection bu baax ci sampling?
RFT dafay jël misaali tontu yu bari, segg ci ñi gëna am poñ, ba noppi defar model bi ci ñi jël raw-gàddu gi.
Ci wàll yiñ mëna saytu lu melni math wala kode, lan moo mëna nekk neexal bi?
Ngir liggéey yiñ mëna saytu, RFT mën na jëfandikoo njub wala jàll ci test yu benn, moytu xeetu neexal buñ jàng.
Ban xeetu famiy moo gëna jëfandikoo échantillonnage rejection ci diiru alignment bi?
Meta dafa leer ni dañuy jëfandikoo misaalu bañ bi bokk ci rëset bi ñuy jëfandikoo ginaaw tàggat yaram ngir xeetu Llama 2 ak Llama 3.
Lan moo waral ekip yi taamu RFT ci kaw RL bu am politik bu melni PPO?
RFT dafay tàggataat ak ñàkka token bu ci topp ci misaal yuñ segg, di moytu jafe-jafe tuning ak ñàkka dal gi ci njuréefi politik-gradient.
Taggat ci misaali neexal bu gëna mag, lan lay def?
Suñu jàngee ci tontu yiñ segg ci kaw, xeetu model bi dafay dugal ci boppam jeffin ju gëna baax ci benn jikko.