Ikibazo gihuza ibibazo · Hagati Urwego

Ikibazo cya RoBERTa

Ongera usubiremo masike ya RoBERTa, intego yo kwitegura, igipimo cyamakuru, tokenizer, indimi nyinshi, hamwe namasomo yerekeranye no kugereranya icyitegererezo.

Inzira zijyanye no kuyoboraImyitozo ya Roberta
Ikibazo 1 Bya 8

Ni ubuhe butumwa nyamukuru RoBERTa yerekeye BERT y'umwimerere?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Ni ubuhe butumwa nyamukuru RoBERTa yerekeye BERT y'umwimerere?
    • Ubwubatsi bwa BERT bwari bufite amakosa
    • BERT yakoresheje ibipimo byinshi
    • BERT yariyemeje cyane
    • BERT yari ikeneye decoder yongeyeho
  2. Niyihe ntego ya BERT RoBERTa yakuyeho?
    • Guhanura interuro ikurikira
    • Kwerekana imiterere y'ururimi
    • Itondekanya
    • Gusimbuza ibimenyetso
  3. Masking ifite imbaraga muri RoBERTa ni iki?
    • Guhisha gusa amagambo adakunze kuboneka muri corpus
    • Guhisha interuro zose icyarimwe
    • Ntuzigere uhisha ikimenyetso kimwe kabiri muri corpus
    • Kubyara mask nshya ishusho burigihe burigihe ikurikiranye
  4. Nigute amakuru ya RoBERTa yamahugurwa yagereranije na BERT?
    • Yakoresheje inyandiko imwe ya 16GB
    • Yakoresheje hafi 160GB, hafi 10x
    • Yakoresheje amakuru make ariko ibirango bisukuye
    • Yakoresheje Wikipedia gusa
  5. Nuwuhe muryango wasohoye RoBERTa?
    • Facebook AI (Meta)
    • Google ubwenge bw'ikoranabuhanga
    • OpenAI
    • Microsoft Ubushakashatsi
  6. Indimi nyinshi za RoBERTa yitwa iki?
    • MBERT
    • RoBERTa-XL
    • XLM-R (XLM-RoBERTa)
    • PolyRoBERTa
  7. Which text encoding did the RoBERTa recipe use according to the guide?
    • A fixed vocabulary of complete words only
    • Byte-level byte-pair encoding
    • One image patch per token
    • A phoneme-only vocabulary
  8. A researcher credits a new layer for beating BERT but also changes the data volume and training duration. Which lesson from RoBERTa matters?
    • Architecture changes always explain every improvement
    • Training duration can be ignored if the benchmark is the same
    • Data volume cannot affect an encoder
    • Training choices can account for gains, so they must be considered when comparing models