What is Llama 2? How to use Llama 2?

From OpenAI's GPT-4 to Google's PalM 2, big language models are dominating the tech headlines. Each new model promises to be better and more powerful than the previous one, sometimes surpassing any existing competitors.

However, the number of existing models does not slow down the emergence of new ones. Now, Facebook's parent company, Meta, has released Llama 2, a powerful new language model. But what is unique about Llama 2? How is Llama 2 different from GPT-4, PaLM 2 and Claude 2 and why should you care about it?

What is Llama 2?

Llama 2, a major language model, is the product of an alliance between Meta and Microsoft, two competing tech giants at the forefront of Artificial Intelligence (AI) research. It is the successor to Meta's Llama 1 language model, which was released in the first quarter of 2023.

It can be said that Meta is equivalent to Google's PaLM 2, OpenAI GPT-4 and Anthropic's Claude 2. It was trained on a huge dataset of publicly available Internet data, enjoying the advantage of a new and more diverse dataset than the one used to train Llama 1. Llama 2 trained on 40% more data than its predecessor and had twice the context length (4k).

 

If you've had a chance to interact with Llama 1 before but weren't too impressed with its output, then Llama 2 outperforms its predecessor and might just be what you need. But how does Llama 2 compete with other competitors?

How does Llama 2 outperform its competitors?

Firstly, Llama 2 is an open source project. This means that Meta is publishing the entire model so that anyone can use it to build new models or applications. If you compare Llama 2 with other major open source language paradigms like Falcon or MBT, you'll find it outperforms them in a number of metrics. It can be said that Llama 2 is one of the most powerful open source large language models on the market today.

However, Llama 2 loses its edge when it comes to the likes of OpenAI's GPT and Google's PalM line of AI models. When it comes to handling creative tasks, Llama 2 is a bit different. Depending on the variant you test with, you may not get the same output as you get from the Anthropic and OpenAI models.

Llama 2 is primarily just a 'base model' rather than a 'refined' option. Foundational models are large language models built with future adaptability in mind. They are not fine-tuned for any particular field but are built to be able to tackle a wide range of tasks, although, at times, with limited capabilities.

A fine-tuned model, on the other hand, is a foundational model that is tweaked to increase its performance in a particular field. It's like taking a platform model like GPT and tweaking it into ChatGPT for the public to use.

 

How to use Llama 2 now

Although Llama 2 is not available on a public platform like ChatGPT, you can still own the model by downloading a copy and running it locally, or using access through the Hugging Face cloud-hosted instance.

To access Llama on Hugging Face, simply open the corresponding Hugging Face link below and start writing a prompt for the AI ​​chatbot.

1. 7B parameters Llama-2 chat
2. 13B parameters Llama-2 chat
3. 70B parameters Llama-2 chat

The Llama models above have been refined for conversational applications, so this is the closest model to ChatGPT you'll get for Llama-2. Not sure which version to try? The article recommends option 3, 70B parameters Llama-2 chat. You can still try with all 3 models to see which best suits your unique needs.

The article used 70B parameters Llama-2 chat by Meta and the results are impressive. To test the model's creativity and sense of humour, the author gave it a test of creativity and characteristic irony. The author asked an AI model to simulate a conversation between two people debating the value of going into space, and here are the results.

Next:

And finally:

 

It doesn't get all the details in our guide right, but the humor is impressive.

On the other hand, if you have the technical expertise to run the Llama model locally on your machine, you can request access to the model using Meta's form. After providing your name, email, location and organization name, Meta will review your application, then access will be denied or granted for a period of several hours to 2 days.