Baamtu Penalty ak Dekodage Kontrol
Kontrolu dekodaas yi ñooy butoŋ yiy wane ni modelu làkk bi di tànnee baat bu ci topp ci distribusioŋ probabilite bi.
Résumé
Settings like temperature, top-p, and repetition penalty shape whether output feels creative, focused, or stuck in loops.
Plongeur bu xóot
Modèlu làkk du génne mbind ci saasi; Dafay genne ab probabilite ngir bépp token bu ci topp. Decodage mooy pexem soppi probabilite yooyu ñu nekk kàddu yu dëggu. Tàngoor dafay soppi jëmmu séddale bi: valeur yu woyof yi dañu koy gëna ñaw ci token bi gëna am wërsëg (focuse, deterministe), valeur yu kawe yi dañu koy platre (divers, riskant). Top-k du denc k token yi gëna am wërsëg; top-p (échantillonnage nucleus) dafay tëye ensemble bi gëna ndaw bi nga xamni probabilité yi dañuy tollu ci 0.9. Baamtu penalti dafay xaaj poñ yi jetons yi jota jëfandikoo, di tere model bi baamtu boppam. Kontrol yi ci laale bokkuna ci penalti frequency (ñu koy xayma ci ni token bi feeñee) ak penalti presence (penalty bu plat bu token bi feeñee). Tuning yii dafay tere boucles robot yi ak rambling bu jaxasoo.
Gis-gis xarala
Baamtu penalti dafay dox ci niveau logit. Laata ñuy soppi poñ yi ci probabilite jaaraleko ci softmax, logit bu token bu nekk buñu njëkka defar dañu koy xaaj ak facteur penalite (1.1 ba 1.3) sudee positif la, wala ñu yokk ko sudee negatif la. Loolu dina wàññi saas ngir tannaat token yooyu. Penalty frequency lu moy loolu dafay dindi benn lim bu méngoo ak limu token bi, fekk penalty presence dafay dindi benn tolluwaayu benn yoon bu token bi feeñee, frequency bi du ci dara.
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 Baamtu Penalty ak Kontrolu Dekodaas
Decodage barabu gëstu la bu am solo. Pexe yu bees yu melni seetlu bu wuute, échantillonnage tipik, échantillonnage eta, ak echantillonnage min-p dañu bëgga yemale lëkkaloo ak wuute ci anam wu gëna am xel, moo gën ay threshold yu fiks. Decodage speculatif dafay jëfandikoo model projet bu ndaw ngir gaawlu defar. Xaarandi sistem yu ëlëg ñu méngale ay parametru dekodaas ci anam wu dinaamik ci muy tekki, ak wane ay seytu yu gëna yomba suko defee jëfandikukat yi mëna laaj 'gëna sos' wala 'gëna leer' te duñu def benn jongante ak loxo ci tàngoor ak ay daan.
Doxal ci àdduna dëgg
Aplikaasioŋ buy bind ci anam wu xarañ dafay yokk tàngoor wi ak top-p ngir génne ay jaar-jaar yu bari te yéeme.
Jàppalekatu kodage dafay wàññi tàngoor wi ci wetu nul suko defee mu delloosi kode bi gëna am wërsëg, matt deterministik.
Chatbot dafay jëfandikoo penalti buy baamtu lu tollu ci 1.2 ngir tere ko baamtu benn mbind bi lu bari yoon.
Jëfandikukat API dafay tëral ay daan ngir tere ab resumé jëfandikoo lu ëpp benn baat bu siiw ci benn këyit bu gudd.
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 Repetition Penalty and Decoding Controls 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
Modèlu projet de decodage speculatif
Laaj yi ñuy faral di laaj
What is Repetition Penalty and Decoding Controls?
Kontrôlu dekodaas yi ñooy butoŋ yiy wane ni modelu làkk bi di tànnee baat bu ci topp ci distribusioŋ probabilite bi. Jekkal yu melni tàngoor, top-p, ak baamtu penalti dañuy wane ndax génne gi dafay nuru lu xarañ, fokalise, wala tëju ci ay boucles.
Lan mooy yokk parametru 'temperatiir' bi ci génnug model bi?
Tàngoor wu gëna rëy dafay dalal seddaleb probabilite, tax token yu néew probabilite gëna am taxawaayu joŋante ak génne gi gëna bari te kenn mënu la wax lumuy jur.
naka la top-p (nucleus) di jëlee dogal ci ban jeton lañu wara bàyyi xel?
Top-p dafay tànn mbooloo mi gëna ndaw ci token yu mag yi nga xamni probabilite cumulatif bi yegg na ci buntu bi (lu melni, 0.9), suko defee poolu kàndidaa bi méngoo ak muy tekki.
Ban etap la penalti buy baamtu di dox?
Baamtu penalti dafay soppi logits (poñ yu bees) ci jeton yuñ jota jëfandikoo balaa ñu leen di soppi def probabilite, wàññi seen tànneef.
Lan mooy wuute bu mag bi am ci digganté daan nekk ak daan bariwaay?
Penalty frequency dafay màgg ak limu yoon yi ñu jëfandikoo benn token, fekk penalty presence dafay jëfandikoo benn wàññi bu takku saa yu token bi feeñee.
Ngir assistant codage bi wara delloo benn mbir bi gëna wóor, ban parametre moo gëna baax?
Tàngoor wu jege nul dafay tax dekodaas bi nekk luñuy xam, daanaka saa yu nekk dañuy tànn token bi gëna am wërsëg, te loolu mooy gëna baax ci kod buñ mëna seentu.