Chapter 1: Introduction to Llama 2

Llama 2 has rapidly become the leading large language model (LLM), and this surge in popularity is backed by several compelling factors. Its performance on public benchmarks surpasses that of earlier models, and unlike its predecessor, Llama 1, Llama 2 can be utilized in commercial settings. Numerous tutorials are available detailing how to implement and deploy Llama 2 effectively.

For more on this topic, feel free to explore my previous articles on The Kaitchup, where I've demonstrated running Llama 2 on consumer-grade GPUs and the process of quantizing it to minimize its footprint.

Section 1.1: The Power of Quantization

Quantization can significantly enhance Llama 2’s efficiency, enabling faster inference on personal systems.

Subsection 1.1.1: Visual Overview of Quantization Techniques

Chapter 2: Key Restrictions of the License

A significant caveat in the licensing terms is the restriction against using Llama 2 to enhance any other LLMs. Specifically, the agreement states:

You are prohibited from using the Llama Materials or any output derived from them to improve other large language models, excluding Llama 2 or its derivatives.

This definition of "Llama Materials" encompasses the model itself. The term "improve" is particularly noteworthy; it gives Meta considerable leeway in defining what constitutes an improvement. For example, generating datasets with Llama 2 for training or fine-tuning another LLM is not permitted, creating a more stringent limitation than what OpenAI enforces.

In contrast, OpenAI allows the use of generated datasets, provided they are not used to create competing models or services. I previously explored this topic in detail:

Can You Use the Falcon Models For Commercial Applications?

An Apache 2.0 license is not all we need

open.substack.com

With Llama 2, the generated datasets can only serve to enhance Llama 2 itself. Even for research purposes, the permissions granted by the license are inadequate. While enforcement of this clause remains uncertain, I anticipate that a substantial amount of data produced by Llama 2 will appear online, likely leading to its use in training other LLMs, which would infringe on the license.

This situation creates a paradox: Meta can harvest data from the internet, including your intellectual property, to refine its own models, yet users are restricted from utilizing data generated by Llama 2 for their own advancements.

I doubt that such restrictive licensing will endure in the long run, though this is merely my perspective. For those interested in a deeper dive, the complete licensing agreement for Llama 2 is available here.

The first video titled "Why You Can't Use Llama 2 (70B), Yet" provides insight into the challenges users face with Llama 2's licensing and usage limitations.

The second video, "How to Use Llama 2 for Free (Without Coding)," offers practical guidance on leveraging Llama 2's capabilities without extensive programming knowledge.