HejuruUbuyobozi bukurikira
Kwiga Gushyira mu Bikorwa (Reinforcement Learning) mu Buhahirane
Tekiniki
Ubuyobozi bwa tekiniki
Reinforcement learning from human feedback uses human judgments to construct a training signal for a model’s behavior.
Umva iki gitaboAI-byavuzwe · ~2 min
It can improve agreement with collected preferences, but does not guarantee truthful, safe, or universally preferred answers.
A pretrained language model predicts plausible text, but plausible is not the same as helpful, honest, or safe. A common RLHF pipeline addresses these goals in stages. First, supervised fine-tuning teaches the model to follow instructions using human-written example answers. Next, humans compare pairs of model responses to the same prompt and pick the better one; these comparisons train a separate reward model that scores any response. Finally, the language model is optimized with reinforcement learning to produce responses the reward model rates highly. A penalty regularizes deviation from a reference policy; it does not eliminate exploitation of reward-model errors. RLHF was central to making ChatGPT-style assistants usable.
Guhuza umufasha wibiganiro kugirango yange ibyifuzo byangiza kandi bitange ibisubizo byingirakamaro, byubatswe neza aho kuba inyandiko yemewe.
Gutondekanya ibice bibiri byincamake kubantu bakunda gutoza icyitegererezo cyandika incamake abantu basanga ari ingirakamaro.
Kugabanya umusaruro wuburozi cyangwa kubogama mugushimira ibisubizo abatanga amanota yabantu bubaha kandi bafite umutekano.
Gukoresha DPO kuri dataset yikundwa na vanze ibisubizo kugirango uhuze moderi ifunguye isoko idakoresheje PPO yuzuye.
RLHF irimo kugenda neza kandi igice cyikora. DPO hamwe nuburyo bujyanye nuburyo bukurikira busimbuza umuyoboro uremereye wa PPO kumakipe menshi, kandi RLAIF ikoresha ibitekerezo byatanzwe na AI (nko muri AI Itegekonshinga) kugirango igabanye ibiciro. Ubushakashatsi burimo guhangana nigihembo cyibihembo, kubogama kwa annotator, hamwe ningorabahizi zo gusuzuma ibisubizo birebire cyangwa byinzobere, hamwe nubuhanga nko kugenzura inzira no kujya impaka. Witege guhuza guhuza ibitekerezo byabantu na AI, ibimenyetso byinshi byigihembo birenze igikumwe kimwe, hamwe no kugenzura neza uwatanze ibyo akunda nindangagaciro bashiraho.
Ibaruwa y'amakuru y'ubuntu
Inkuru eshatu zizewe za AI buri gitondo cy'iminsi y'akazi, zanditse mu rurimi rworoshye. Ubuntu iteka ryose, nta kwamamaza.
Imeri imwe buri munsi y'akazi. Kwiyimura mu buryo bumwe. Ntitwigera tugurisha cyangwa ngo dusangire aderesi yawe.
Isuzume
Ibisubizo ako kanya kuri buri kibazo, n'icyemezo gishobora gusangirwa gifite ID yagenzurwa umaze gutsinda isomo.
Fasha uburezi bwa AI bw'ubuntu. AI Understanding ni umuryango utegamiye kuri leta wa 501(c)(3) — nta kwamamaza, nta kwishyura, iteka ryose. Tanga impano
Reinforcement learning from human feedback uses human judgments to construct a training signal for a model’s behavior. It can improve agreement with collected preferences, but does not guarantee truthful, safe, or universally preferred answers.
RLHF iyobora icyitegererezo kubisubizo abantu bakunda, bikarushaho gufasha, kuba inyangamugayo, n'umutekano aho kuba byemewe gusa.
Icyitegererezo cyigihembo cyahuguwe kubigereranyo byabantu kugereranya amanota kugirango batange ibisubizo, bitanga ikimenyetso politiki nziza.
The penalty discourages large policy shifts. It is not proof that the reward model cannot be exploited or that generated answers are correct.
Annotator batoranya igisubizo cyatoranijwe mugereranya, itoza icyitegererezo cyo guhemba binyuze mu gihombo cya Bradley-Terry.
DPO ikura intego iturutse kubyo ukunda, irinda icyitegererezo cyigihembo cyihariye hamwe nintambwe yo kwiga ishimangira.
Komeza wige
Abandi bayobozi batoranijwe kuriyi ngingo
HejuruUbuyobozi bukurikira
Kwiga Gushyira mu Bikorwa (Reinforcement Learning) mu Buhahirane
Tekiniki