w 10l1 large language models

Question 1

what is the pretrain fine tune problem and solution

Answer 1

traditional machine learning invovled training models from scratch for every new tasks

so now we train on simple tasks then fine tune for specific tasks

Question 2

what is the pretrain fine tune paradigm

Answer 2

● This idea of using a different (and simple) training objective
to learn linguistic priors, before then training (fine-tuning)
on the downstream task is the pre-train fine-tune paradigm.
● Notice that the pre-training task is self-supervised
○ No manual annotation is required for this

Question 3

will language modeling lead to bias?

Answer 3

yes

because it will only store the most freuqeny asspects of the data which are limited and biased

leads to hallucination

Question 4

how do we evaluate our llm

Answer 4

train and then test on test set with evlautaiton metric

Question 5

in a transformer, gpt and bert which one is encorder and decorder

Answer 5

transformer encoder, decorder

gpt decorder only

bert encoder only

Question 6

Why can’t you train an Encoder-Decoder model and
“break” it?
● Encoder only architectures are used for classification:
○ BERT (Bidirectional Encoder Representations from
Transformers), RoBERTa, …
● Decoder only models are used for generation:

Question 7

how should i phrase my input for llm

Answer 7

BAD:
● Where in London should I buy a house?
Better:
● This is the drop in interest rates, this is the increase in availability in
…
Based on the content provided, where should I buy a house?
Provide examples from the lecture content. [chain of thought]
Best:
● Same as above, but give an example of the kind of reasoning required
if what you want is more complex.