Iphrojekthi ye-OpenAI, ethuthukisa amaphrojekthi omphakathi emkhakheni wobuhlakani bokwenziwa, ishicilele i-Transformer Debugger, eklanyelwe ukuhlaziya ukwenziwa kusebenze kwezakhiwo kumamodeli olimi lokufunda ngomshini lapho kucutshungulwa idatha ethile. Njengezixazululi zendabuko, i-Transformer Debugger isekela ukuzulazula kwesinyathelo ngesinyathelo ngokukhishwa kwemodeli, ukulandelela, nokunqamula umsebenzi othile. Ngokuvamile, i-Transformer Debugger ikuvumela ukuthi uqonde ukuthi kungani imodeli yolimi ibonisa ithokheni eyodwa esikhundleni senye ekuphenduleni isicelo esithile, noma kungani imodeli inaka kakhulu amathokheni athile esicelweni. Ikhodi ibhalwe ngePython futhi isatshalaliswa ngaphansi kwelayisense ye-MIT.
Ukwakheka kuhlanganisa izingxenye ezilandelayo:
- Isibukeli se-Neuron siyitilosi ngokusebenzisa izingxenye zemodeli ngayinye, njengama-neuron e-MLP, i-foci yokunaka, nezethulo ze-autoencoder ezifihliwe.
- Iseva yokwenza kusebenze ingemuva leseva elisebenzisana nemodeli ecutshungulwayo futhi libuyisela idatha yesilungisi sephutha.
- Amamodeli yilabhulali yokusebenzelana namamodeli olimi e-GPT-2 kanye nama-autoencoder asetshenziswa kuwo, ahlinzeka ngokushintshwa kwezibambi ukuze kuvinjwe ukwenziwa kusebenze.
- Izibonelo zokuvula amasethi edatha we-MLP neurons, ifoci yokunaka, nokumelwa okucashile kwama-autoencoder.
Source: opennet.ru
