A developer wants to improve the output quality of a generative AI app they just built. They want to use the easiest-to-try customization technique that improves output by refining how instructions and examples are written, without changing the model's weights. Which technique is the MOST suitable?
Select the easiest customization technique that does not change the weights.
Fine-tuning
Fine-tuning is a technique that additionally trains the model's weights with labeled data, requiring data preparation and training.
It is not the easiest-to-try technique without changing weights, so it is incorrect.
The differences among pre-training, continued pre-training, and fine-tuning are summarized in the key point diagram below.
Continued pre-training
Continued pre-training is a technique that additionally trains a model with a large amount of unlabeled data to add domain knowledge, at significant cost.
The difference from pre-training is that pre-training is the first process of building a model from nothing, whereas continued pre-training continues pre-training on an already pretrained model with data from a specific domain to layer on domain knowledge. Both update the weights, so it is not the easiest-to-try technique without changing weights, making it incorrect.
The differences among the three techniques are summarized in the key point diagram below.
Pre-training
Pre-training is the most costly technique that trains a model at large scale from scratch.
The difference from continued pre-training is that continued pre-training adds data from a specific domain to an already pretrained model to layer on domain knowledge, whereas pre-training is the first process of building, from a large amount of data, a model that has no knowledge yet. Because it is the first model-building process, it is not the easiest-to-try technique without changing weights, making it incorrect.
The differences among the three techniques are summarized in the key point diagram below.
Prompt engineering
This is correct. Prompt engineering is the easiest-to-try technique that improves output by refining how instructions, examples, output formats, and constraints are written, without changing the model's weights.
Customization techniques, in order of increasing cost and effort: 'prompt engineering < RAG < fine-tuning < continued pre-training < pre-training.' Prompt engineering is the easiest, without changing the weights, and is the means to try first. Fine-tuning and beyond involve training the weights and increase cost.