Llama 3.3 70B Overview
When utilised for text-only applications, Llama 3.3, a text-only 70B instruction-tuned model, offers better performance than Llama 3.1 70B and Llama 3.2 90B. Furthermore, in certain cases, Llama 3.3 70B performs similarly to Llama 3.1 405B. Meta has released a new, state-of-the-art 70B model that performs comparably to the Llama 3.1 405B model.
In 70B, the Meta Llama 3.3 multilingual large language model (LLM) is a generative model that has been pretrained and instruction adjusted. On widely used industry benchmarks, the Llama 3.3 70B instruction customised text only model performs better than many open source and closed chat models and is tailored for multilingual discussion use cases.
Which languages does Llama 3.3 officially support?
English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai are the eight languages that Llama 3.3 70B formally supports, according to the sources.
A pretrained and instruction-tuned generative model, the Meta Llama 3.3 70B multilingual large language model is optimized for multilingual discussion use cases throughout these supported languages. These eight languages are the ones that meet safety and helpfulness performance standards, even though the model has been trained on a larger set of languages.
The authors specifically advise against developers utilising Llama 3.3 70B to communicate in unsupported languages without first putting system restrictions and fine-tuning in place. It is deemed out-of-scope to use the model in languages other than these eight that do not include these measures. As long as they adhere to the Acceptable Use Policy and the Llama 3.3 Community License and guarantee responsible and safe use, developers can customise Llama 3.3 for other languages.
New capacities
A wider context window, multilingual inputs and outputs, and the potential for developer collaboration with third-party tools are among the new features included in this release. In addition to the recommended practices that are generally applicable to all use cases of generative AI, there are particular considerations that must be made while building with these new capabilities.
Utilising tools
Developers are in charge of integrating the LLM with the tools and services of their choosing, much like in traditional software development. In order to be aware of the safety and security constraints when using this feature, they should establish a clear policy for their use case and evaluate the reliability of the third-party services they utilise. For recommended practices on how to safely implement the third-party protections, consult the Responsible Use Guide.
Being multilingual
It’s possible that Llama can produce text in languages other than those that satisfy safety and usefulness performance requirements. In accordance with their policies and the best practices discussed in the Responsible Use Guide, it strongly advise developers not to use this model to communicate in non-supported languages without first putting in place system restrictions and fine-tuning.
Purpose of Use
- Targeted Use Cases Llama 3.3 70B is designed for multilingual commercial and research use. While pretrained models can be modified for a range of natural language generating tasks, instruction tailored text only models are designed for assistant-like chat. Additionally, the Llama 3.3 model facilitates the utilisation of its models’ outputs to enhance other models, such as distillation and synthetic data production. These usage situations are permitted by the Llama 3.3 Community License.
- Beyond the purview of Use in a way that contravenes any applicable laws or rules, including those pertaining to trade compliance. Use in any other manner that is forbidden by the Llama 3.3 Community License and the Acceptable Use Policy. Use in languages other than those that this model card specifically mentions as supported.
Note: The eight available languages are not the only set of languages that Llama 3.3 70B has been trained on. If developers follow the Acceptable Use Policy and the Llama 3.3 Community License, they can modify Llama 3.3 models for languages other than the eight supported languages. In these situations, they are in charge of making sure that any usage of Llama 3.3 in other languages is done responsibly and safely.
