Meta Unveils Game-Changing Open-Source LLaMa-2-Chat with Unprecedented Performance

LLaMa-2-Chat is a truly remarkable achievement created by the talented team at GenAI’s LLM department. With impressive parameters count of 70 billion, this model is comparable to and even outperforms the highly regarded GPT-3.5 on certain benchmarks.

Highlights:

• Commercial friendly
• Pretrained on 2T tokens
• Strong MMLU scores (i.e. strong reasoning)
• 4K context
• (Extendable) RoPE embeddings
• Coding performance is meh
• SFT/RLHF chat versions

The fact that LLaMa-2-Chat is the first model of its size to be fine-tuned using RLHF (Reinforcement Learning from Human Feedback) makes it even more remarkable. In an unprecedented move, Meta has made this model completely free for commercial use. Those interested can request the download link from their official website.

One of the most significant advantages of LLaMa-2-Chat is its potential to create ChatGPT analogues without the need to share any data with OpenAI. This empowers developers and researchers to harness the power of the model while maintaining complete control over their data. You can download new model here.

HUGE AI NEWS!!!🔥Llama 2 just came out! And guess what? It’s fully open-source and can be used for commercial purposes!!! 7-70B parameters are supported.

They also release fine-tuned variants optimized for dialogue use cases (LLaMA 2-Chat)!

The paper looks super detailed – 76… pic.twitter.com/yZahl7Jzya

— Aleksa Gordić 🍿🤖 (@gordic_aleksa) July 18, 2023

In terms of human evaluation metrics, LLaMa-2-Chat stands shoulder-to-shoulder with ChatGPT-3.5 in terms of quality. Notably, it showcases exceptional performance on mathematical problems, outperforming other models in this domain.

• In February, Meta has released LLaMA model, a large language model designed to support AI researchers. Available in various sizes (7B, 13B, 33B, and 65B parameters), LLaMA allows researchers to test new approaches and explore new use cases. The model is ideal for fine-tuning tasks and is built on a large set of unlabelled data. Despite its versatility, LLaMA faces risks of bias, toxic comments, and hallucinations. The model is released under a non-commercial license focused on research use cases, and eligibility for access is evaluated on a case-by-case basis.
• The LLaMa model, with 7 billion parameters, has achieved lightning-fast inference on a MacBook with the M2 Max chip. This achievement was made possible by Greganov’s successful implementation of model inference on the Metal GPU, a specialized accelerator found in Apple’s latest chips. The LLaMa model demonstrates 0% CPU utilization, harnessing the processing power of all 38 Metal cores. This vision of personalized AI assistance and localization on personal devices holds immense potential for a future where AI becomes an integral part of people’s lives, providing personalized assistance and streamlining routine tasks.

Read more about AI: