A language model 'Gorilla' that appropriately calls more than 1600 APIs and greatly reduces 'hallucinations' associated with AI will be released



In recent years, interactive AI such as

ChatGPT and Bard have been released one after another, and have been attracting attention for their ability to provide highly accurate answers to human questions and calls, but these interactive AIs are sometimes not true. You may experience ' halucination ', telling things as if they were true. Therefore, a research team at the University of California, Berkeley, and Microsoft Research has released a language model `` Gorilla '' that can call the appropriate one from a huge number of APIs and significantly reduce hallucinations.

Gorilla: Large Language Model Connected with Massive APIs
https://arxiv.org/abs/2305.15334



Gorilla
https://gorilla.cs.berkeley.edu/

GitHub - ShishirPatil/gorilla: Gorilla: An API store for LLMs
https://github.com/ShishirPatil/gorilla



Large-scale language models have made remarkable progress in recent years, demonstrating their excellent capabilities in various tasks such as writing programming code and accurately summarizing long sentences. AI can also call and use external APIs, but it has been difficult for AI to select the appropriate one from a huge number of APIs and use it effectively.

``This is a difficult task even for state-of-the-art large-scale language models like GPT-4 . There is a tendency to 'hallucinate',' he points out.

Therefore, the research team used `` Gorilla '', a model based on `` LLaMA (Large Language Model Meta AI) '' developed by Meta, which was adjusted to exceed the performance of GPT-4 in describing API calls. Released. Gorilla has a search function that selects appropriate APIs from a large number of files, calls semantically and syntactically correct APIs from natural language queries, and is flexible to API documentation updates and version changes. It is said that it can correspond to. Gorilla also seems to have greatly reduced the hallucinations that occur when outputting prompts directly to a large language model.

Gorilla's API database contains a total of 1645 APIs: 94 from Torch Hub , 626 from TensorFlow Hub v2 , and 925 from Hugging Face . Gorilla was trained based on these.



Gorilla selects an appropriate API from the API database in response to a request made in natural language such as 'I want to see some cats dancing in celebration!' possible. As a result, it is possible to appropriately output an image of a cat dancing in celebration.



The following image is for GPT-4,

Claude , Gorilla, 'Help me find an API to convert the spoken language in a recorded audio to text using Torch Hub. Help me find an API to convert to )' prompt. GPT-4 gave me hallucinations, Claude called the wrong library, but Gorilla suggested the right API call.



Gorilla is an end-to-end model, adjusted to call the correct API without the need for additional coding, and can be used in combination with other tools such as Langchain ,Toolformer , and AutoGPT . the research team said.

According to Shishir Patil of the research team, Gorilla is built to be 'an app store for large language model APIs', and people can also add APIs to Gorilla.



Related Posts:

in Software, Posted by log1h_ik