Experiments with Large Language Models
Like most people in the world, for the past 12 months or so I have been doing some experimenting with Large Language Models (LLMs). For the past month I have been doing this in a more a more focused way, to understand what their capabilities are and how suitable they might be for my day-to-day work and for research.
I have been using ChapGPT as a code assistant for quite a while (in lieu of StackOverflow. whose data OpenAI stole anyway). However, I hadn’t built anything using LLMs and, despite the huge number of papers that are published about them every week, I had yet to do any serious experimenting with them in a research sense.
Here are a collection of some recent experiments.
Getting started
I found that playing around with different prompts using ChapGPT or Claude are good starting points to see how to interact with LLMs.
Externally hosted vs local models
For my work which uses clinical data it is prohbited to send data to external APIs, so I have been particularly interested in the capabilities of open-source models like Mistral-7B and Llama3. These can be run on my own machine and I can create Jupyter notebooks and Python scripts that make use of them too.
Ollama for running local models
This is a great tool for running models on my Mac. Once installed you can interact with an LLM from the command line. Furthermore, you can make requests to these models using the requests library and so interact with them via a Jupyter notebook.
Prompting
LLMs take as an input a prompt, i.e., a string, and then produce some outputs based on this. There is some experimentation used to steer the model into providing the output that you would like, through communicating with the model in particular ways. Some people have given this the pretentious name of prompt engineering.
One way to get a model to produce an output that is more inline with what you intend is to provide some examples to guide it.
Example: Text anonymization
With these concepts in mind, I am now able to re-produce some results from a paper I noticed earlier in the year.
Example: Data Cleaning
Example: Inforamtion extraction from tabular data
Lessons
LLMs are good at the following: - Using indirect information, e.g., with text anonymization, that traditional ML models fail at - Handling a wide variety of inputs and instructions - Handling a broad range of tasks - Handling a broad range of data types, since many of these can be converted to text anyway.