Enhancing LLMs’ capabilities with tools
What is a tool? Let’s continue asking an LLM to solve a mathematical problem. An obvious way to improve the quality is to collect some examples and add them into training (or a fine-tuning dataset). But are there any easier ways? Now, let’s ask it a different question:
math_problem2 = "How much is 23*2**2+156/4-18?"
As we’ve mentioned earlier, these days, even smaller LLMs (such as Gemma 2 2B or Llama 3.1 8B) do a surprisingly amazing job on such mathematical examples. But still, for the sake of an argument, let’s discuss what else we can do besides improving our datasets.
What if instead of teaching LLMs how to solve mathematical problems, we give them access to a calculator that was designed exactly to solve such problems? This logic can be applied to any specific instruments (we will call them tools) that are developed for certain tasks (such as creating weather forecasts, using CRM systems...
