| Model Type | | text generation, multilingual |
|
| Use Cases |
| Areas: | |
| Applications: | | Assistant-like chat and agentic applications, Knowledge retrieval and summarization, Writing assistants, Query and prompt rewriting |
|
| Primary Use Cases: | | Natural language generation tasks |
|
| Limitations: | | Use that violates applicable laws or regulations, Prohibited by the Acceptable Use Policy |
|
| Considerations: | | Developers must ensure compliance with laws and complete deployments safely. |
|
|
| Additional Notes | | Intended for use in constrained environments; involves elements of responsible AI development. |
|
| Supported Languages | | English (supported), German (supported), French (supported), Italian (supported), Portuguese (supported), Hindi (supported), Spanish (supported), Thai (supported) |
|
| Training Details |
| Data Sources: | | publicly available online data |
|
| Data Volume: | |
| Methodology: | | Using logits from the Llama 3.1 8B and 70B models in pretraining, knowledge distillation, Supervised Fine-Tuning, Rejection Sampling, Direct Preference Optimization. |
|
| Context Length: | |
| Hardware Used: | | Meta's custom built GPU cluster, H100-80GB |
|
| Model Architecture: | | auto-regressive language model using optimized transformer architecture |
|
|
| Safety Evaluation |
| Methodologies: | | Safety as a System, Red Teaming, CBRNE risk assessment, Child Safety risk assessments, Cyber Attacks risk assessment |
|
| Risk Categories: | | CBRNE, Child Safety, Cyber Attacks |
|
|
| Responsible Ai Considerations |
| Fairness: | | Inclusive development, acknowledging diverse user backgrounds. |
|
| Transparency: | | Open sourced and open to community contributions. |
|
| Mitigation Strategies: | | Implemented safety mitigations and tones to address safety and ethical concerns. |
|
|
| Input Output |
| Input Format: | |
| Accepted Modalities: | |
| Output Format: | |
|
| Release Notes |
| Version: | |
| Date: | |
| Notes: | | Release of Llama 3.2 multilingual models. |
|
|
|