Tradeoffs in LLM performance. The following list compiles some modifications we can make to a given LLM in hopes

of achieving superior performance. For each item, state one potential drawback in training or deployment resulting from

this change. You may not repeat the same drawback.

a$)$ In an LLM tailored to help high school students learn machine learning, train the LLM on a dataset consisting of all

machine learning papers published at NeurIPS in the past $5$ years.

b$)$ Train an LLM on a dataset generated using only outputs from a different state$-$of$-$the$-$art LLM$.$

c$)$ Increase the number of parameters in the LLM by a factor of $100$ by incorporating additional attention layers.

d$)$ Increase the vocabulary size by a factor of $10.$