Kontynuacja quizu dotyczącego szkolenia wstępnego i dostrajania

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. What is the main data difference between continued pretraining and supervised fine-tuning?
   • Both use only labeled question and answer pairs
   • Continued pretraining uses raw domain text; fine-tuning uses curated labeled examples
   • Fine-tuning uses more data than continued pretraining
   • Continued pretraining uses only images
2. In supervised fine-tuning, where is the loss usually computed?
   • On every token including the prompt
   • Only on the prompt tokens
   • On randomly chosen tokens
   • Only on the response tokens
3. What is replay in continued pretraining?
   • Mixing some general data into training to limit forgetting
   • Running the same batch twice
   • Restarting training from scratch
   • Replaying user conversations at inference time
4. How was Code Llama built?
   • From scratch on code only
   • By prompt engineering Llama 2
   • By continuing to train Llama 2 on a large code corpus
   • By fine-tuning on a few hundred coding questions
5. For facts that change often, which approach does the guide suggest is often better than fine-tuning?
   • Continued pretraining every day
   • Retrieval of up-to-date documents
   • Raising the sampling temperature
   • Training a larger tokenizer
6. Which statement about BloombergGPT is correct according to the guide?
   • It was continued pretraining of Llama 2
   • It was created only with LoRA
   • It was a prompt-engineered GPT-3
   • It was trained from scratch on financial and general data
7. What did the 2020 paper Don't Stop Pretraining show?
   • A second phase of pretraining on in-domain text improved RoBERTa on domain tasks
   • Pretraining should always stop after one epoch
   • Fine-tuning is useless for classification
   • Domain text harms every model
8. Why re-warm the learning rate at the start of continued pretraining?
   • To reset the tokenizer
   • To freeze the base weights
   • Resuming at a high rate destabilizes training while too low a rate learns little
   • To skip the need for domain data