Mcgill-nlp

Llama-3-8B-Web

McGill-NLP

178

Llama-3-8B-Web is a finetuned Meta-Llama-3-8B-Instruct model developed by McGill-NLP. It uses the recently released Meta Llama 3 model as a base and finetunes it on the WebLINX dataset, a collection of over 100K web navigation and dialogue instances. This allows the model to excel at web-based tasks, surpassing GPT-4V by 18% on the WebLINX benchmark. Model inputs and outputs Llama-3-8B-Web takes in text as input and generates text as output. The model is designed for web-based applications, allowing agents to browse the web on behalf of users by generating appropriate actions and dialogue. Inputs Text representing the current state of the web browsing task Outputs Text representing the next action the agent should take to progress the web browsing task Capabilities Llama-3-8B-Web demonstrates strong performance on web-based tasks, outperforming GPT-4V by a significant margin on the WebLINX benchmark. This makes it well-suited for building powerful web browsing agents that can navigate and interact with web content on a user's behalf. What can I use it for? You can use Llama-3-8B-Web to build web browsing agents that can assist users by automatically navigating the web, retrieving information, and completing tasks. For example, you could create an agent that can book flights, make restaurant reservations, or research topics on the user's behalf. The model's strong performance on the WebLINX benchmark suggests it would be effective at such web-based applications. Things to try One interesting thing to try with Llama-3-8B-Web is building a web browsing agent that can engage in natural dialogue with the user to understand their needs and preferences, and then navigate the web accordingly. By leveraging the model's text generation capabilities, you could create an agent that feels more natural and human-like in its interactions, making the web browsing experience more seamless and enjoyable for the user.

Read more