AI Engineering: Building Applications with Foundation Models

it’s important to be aware of how they work under the hood to avoid unnecessary costs and headaches.

Language models are generally better with text than with numbers.

the low quality of training

it’s common to use a weaker model for intent classification and a stronger model to generate user responses.

According to their paper, models trained to be more aligned “are much more likely to express specific political views (pro-gun rights and immigration) and religious views (Buddhist), self-reported conscious experience and moral self-worth, and a desire to not be shut down.”

you’re doing supervised finetuning, your data is most likely in the format (instruction, response). Instructions can be further decomposed into (system prompt, user prompt). If you’ve graduated to finetuning from prompt engineering, the instructions used for finetuning might be different from the instructions used during prompt engineering. During

you don’t want preambles.

The “use what we have, not what we want” approach may lead to models that perform well on tasks present in the training data but not necessarily on the tasks you care about. To address this issue, it’s crucial to curate datasets that align with your specific needs. This section focuses on curating data for specific languages and domains, providing

examples