Meta's language model 'Toolformer' calls and uses search engines, calculators, calendars, etc. with API



ChatGTP can be said to have revolutionized natural language processing, but on the other hand, it can have problems with basic tasks such as fact checking and math. Meta's large-scale language model ' Toolformer ' learns how to use external tools via API without sacrificing the language model part, so you can use search engines, calculators, calendars, etc. .

[2302.04761] Toolformer: Language Models Can Teach Themselves to Use Tools

https://doi.org/10.48550/arXiv.2302.04761



Meta develops an AI language bot that can use external software tools | Ars Technica
https://arstechnica.com/information-technology/2023/02/meta-develops-an-ai-language-bot-that-can-use-external-software-tools/

Toolformer is a model trained to determine which APIs to call, when to call them, what arguments to pass, and how to best incorporate the results into future token predictions.

During training, researchers provided a small human-generated sample set showing how each API would be used, and annotated the large language model dataset with potential API calls. allow. As a result of making it possible to learn without explicit human guidance, Toolformer learned to predict text-based API calls as if they were text in other formats.

Historically, large language models have not been very good at math, but Toolformer gets around this by calling calculators in its API.

In addition, the news site Ars Technica has a large language model that can now call APIs, which may damage user data in the app or cause problems externally via the browser etc. I point out that there is.

in Software, Posted by logc_nt