📰 News: Meta Unveils Llama 2, An Enhanced Text-Generating Models for Better Support

In response to the expanding generative AI field, Meta has unveiled Llama 2, an AI model update specifically designed to enhance support for modern chatbots, including OpenAI's ChatGPT, Bing Chat, and other advanced chatbots.

Llama is a collection of models capable of generating text and code in response to prompts like other chatbot systems. Llama 2, an upgraded version of the original Llama model, is trained on a combination of publicly accessible data and has demonstrated superior performance compared to earlier Llama models.

Unlike its predecessor, which was leaked online against the company policy, Llama 2 is freely accessible for research and commercial purposes and can be fine-tuned on AWS, Azure, and Hugging Face's AI model hosting platform in its pre-trained state.

Meta has confirmed that Llama 2 is designed to offer enhanced compatibility and optimized performance on Windows, smartphones, and PCs powered by Qualcomm Snapdragon chips, thanks to an extended partnership with Microsoft and Qualcomm.

In addition to the above-mentioned differences between Llama 2 and the Llama AI models, here are more prominent differences between the two models:

• Llama 2 is offered in two flavors: Llama 2 and Llama 2-Chat, with the latter specifically tailored for two-way communication.

• Llama 2 and Llama 2-Chat are subdivided into versions with different levels of sophistication: 7 billion, 13 billion, and 70 billion parameter models.

• Llama 2 was trained on two million tokens, nearly twice as many as Llama (1.4 trillion), which typically enhances generative AI proficiency as more tokens generally yield better outcomes.

• Meta's internal benchmarks reveal that Llama 2-Chat performs better than Llama 2 in terms of "helpfulness" and toxicity.

As for its drawbacks:

• Meta admits Llama 2 tests cannot cover all real-world scenarios and may lack benchmark diversity.

• Meta acknowledges that Llama 2, like other generative AI models, exhibits biases, resulting in a higher frequency of "he" pronouns relative to "she" due to training data imbalances.

• Llama 2 exhibits a Western bias stemming from data imbalances that include an overrepresentation of terms like "Christian," "Catholic," and "Jewish."

While Meta hasn't disclosed the exact sources of the training data in the whitepaper, it states that the data primarily originates from the web in English and not from Meta's own products or services.