On April 18th, local time in the US, Meta released the highly anticipated open-source large language model Llama3, which offers two model sizes: 8B and 70B parameters, with a 400B version expected to be released in the future.
Meta mentioned in their blog that thanks to improvements in training techniques, the Llama3 models, both 8B and 70B, are the strongest among models of the same parameter scale.
As shown in the figure above, the performance of Llama3 70B even surpasses that of Google’s Gemini Pro1.5 and Anthropic’s Claude 3 Sonnet.
Llama3 significantly reduces the error rejection rate, enhances model alignment, and increases the diversity of model responses. Additionally, its capabilities in logical reasoning, code generation, and instruction following have also improved substantially, with overall controllability being stronger.
Besides the usual benchmark scores, Meta has also developed a testing method that is closer to real-world experiences.
In Meta’s testing, 1800 prompts were compiled and given to models including Meta3, Claude Sonnet, Mistral Medium, and GPT-3.5, which were then evaluated by humans based on their responses.
The results showed that the Llama3 70B instruction-tuned version is stronger than Claude Sonnet, Mistral Medium, OpenAI GPT-3.5, and Llama2.
It is clear that Llama3 is quite powerful. So, how can one experience this stronger new model?
Shortly after the release of Llama3, Amazon Web Services released a press release stating that unlike previous models listed on Amazon Bedrock, this time it is available on Amazon SageMaker JumpStart.
Amazon SageMaker JumpStart is a machine learning service specifically built by Amazon Web Services, offering pre-trained models, built-in algorithms, and pre-built solutions to help users quickly evaluate models and develop AI projects.
Deploying Llama 3 through Amazon SageMaker JumpStart is quite smooth. Users can see models from developers like HuggingFace, Meta, Stability.ai, etc., in the JumpStart backend.
Under the Meta tag, multiple versions of the Llama model are provided. By clicking to view the Llama3 70B instruction-tuned version, there is an introduction page for the model, and a prominent “Deploy” button in the upper right corner.
After clicking the “Deploy” button, users can choose the specifications of the cloud host and make some primary settings. Finally, they can click deploy to deploy it to the corresponding instance.
The model deployment process is very straightforward, and once deployed, users can develop in their preferred way.
During development, features like SageMaker Pipelines, SageMaker Debugger, or container logs can be used to enhance model performance and implement MLOps controls.
Since all these operations are performed on the Amazon Web Services platform, both the model itself and the related data are protected under the security measures provided by Amazon Web Services.
If you want to personally try how to deploy and develop, you can refer to the technical blog of Amazon Web Services:
https://aws.amazon.com/cn/blogs/machine-learning/meta-llama-3-models-are-now-available-in-amazon-sagemaker-jumpstart/
Meta stated in their blog that a larger 400B version will be open-sourced in the future, and the 400B version is still in training. If one day the Llama3 400B is suddenly released, you might consider using Amazon SageMaker JumpStart to get hands-on experience right away.
Currently, various new and stronger large models continue to emerge, and today’s strongest may be surpassed tomorrow. In the face of the explosion of large model technology, many people are eager to try while feeling confused and anxious.
In response, Amazon Web Services CEO Adam Selipsky said, “Things are developing so rapidly that adaptability is the most valuable skill you can possess in this environment. There will not be a single model that dominates everything, nor will there be a company that can provide a model for everyone to use.”
Since new technologies are constantly emerging, one should try to understand the newer technologies.
In the process, consider utilizing the convenient platforms and tools provided by Amazon Web Services, which has always aimed to make the use of various technologies simpler.
In the face of large models, Amazon SageMaker JumpStart and Amazon Bedrock are both excellent starting points. They can help alleviate the anxiety brought by large model technology.
● The strongest open-source commercial LLM requires at least 15GB of memory, just use SageMaker!
● Stable Diffusion model training with Amazon SageMaker
● Can’t afford a GPU for your graduation project? Amazon SageMaker offers it for free!
● Amazon SageMaker: A ready-to-use machine learning service