| Model Type | | text-generation, instruct, auto-regressive |
|
| Use Cases |
| Areas: | | research, commercial applications |
|
| Applications: | | assistant-like chat, NLP tasks |
|
| Primary Use Cases: | | natural language generation, instruction following |
|
| Limitations: | | bias and objectionable responses possible, testing primarily in English |
|
| Considerations: | | Developers should perform safety testing before deploying applications. |
|
|
| Supported Languages | |
| Training Details |
| Data Sources: | | public online data, instruction datasets, human-annotated examples |
|
| Data Volume: | |
| Methodology: | | supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) |
|
| Context Length: | |
| Model Architecture: | | optimized transformer architecture |
|
|
| Safety Evaluation |
| Methodologies: | | red teaming, adversarial evaluations |
|
| Findings: | | residual risks reduced compared to previous models, emphasis on model refusals to benign prompts |
|
| Risk Categories: | | CRBNE, Cyber Security, Child Safety |
|
| Ethical Considerations: | | Model development involved mitigation strategies to limit misuse and harm. Overfocus on safety without impacting user experience. |
|
|
| Responsible Ai Considerations |
| Fairness: | | Inclusive and open approach. |
|
| Transparency: | | Responsible Use Guide provided. |
|
| Accountability: | | Meta provides resources and tools for developers. |
|
| Mitigation Strategies: | | Purple Llama tools, Llama Guard, and safety best practices |
|
|
| Input Output |
| Input Format: | |
| Accepted Modalities: | |
| Output Format: | |
| Performance Tips: | | Follow safe use guidelines and optimize system level safety with layered tools |
|
|