toolformer-pytorch

Why we included this project

This is a working reference implementation of Meta's Toolformer, the training method that teaches a language model when to insert API calls instead of guessing. Rather than shipping a finished agent, it walks the whole loop: generate candidate tool calls from a prompt, score each one by how much it lowers perplexity, keep the ones that help, and fine-tune on those. The pieces are separated cleanly, so you can study the fitness filter on its own or plug in your own model and any Python function as the tool. For researchers and grad students who want to see how tool-use fine-tuning actually works, or to adapt the technique to their own experiments, it is a useful starting point, though not something you would drop into production as-is.