Gorilla: Remodeling LLMs to Accurately Invoke Massive API Integration and Reduce Hallucination

A team of researchers successfully integrated over 1600 APIs into the LLaMa-7B model, surpassing the performance of even GPT-4 and Claude AI in terms of the quality of API calls. Open-source projects have been at the forefront of this advancement, paving the way for greater accessibility and collaboration.

The primary focus lies in the seamless integration of neural networks with popular platforms such as Torch Hub, HuggingFace, and Tensorflow Hub. The objective is simple: given a natural language query, Gorilla determines the most appropriate API call to make. This capability opens up a realm of possibilities, allowing for the development of smart pipelines where different models are invoked based on user requests.

LLMs have demonstrated exceptional progress in various tasks, including mathematical reasoning and program synthesis. Their ability to effectively utilize tools through API calls has remained underutilized. Gorilla aims to bridge this gap, offering a finely-tuned LLaMA-based model that outperforms GPT-4 in generating accurate API calls.

The significant advantage of Gorilla is its integration with a document retriever, enabling the model to adapt to real-time document changes. This flexibility allows for seamless updates or version changes, ensuring reliable and up-to-date outputs. Gorilla reduces the issue of hallucination, a common challenge faced when directly prompting LLMs.

To evaluate Gorilla’s capabilities, the researchers introduce APIBench, a comprehensive dataset comprising HuggingFace, TorchHub, and TensorHub APIs. The successful integration of the retrieval system with Gorilla showcases the potential for LLMs to utilize tools more accurately, keeping pace with frequently updated documentation and enhancing the reliability and applicability of their outputs.

Gorilla marks a significant milestone in the evolution of LLMs, empowering them to invoke APIs with precision. With the ability to accurately determine the appropriate API call based on natural language queries, Gorilla sets a new standard for reducing hallucination and improving overall performance. The researchers have also released APIBench, the largest curated collection of APIs, providing a valuable resource for training and development.

The team behind Gorilla invites collaboration, aiming to expand the largest API store and continue teaching LLMs how to effectively interact with APIs. Whether through joining their Discord community, opening a PR, or reaching out via email, developers and API creators are encouraged to participate in this endeavour.

Read more about AI: