Model Type | text generation, dialogue use cases |
|
Use Cases |
Areas: | |
Applications: | Assistant-like chat, Natural language generation tasks |
|
Primary Use Cases: | Chat models, Instruction tuned applications |
|
Limitations: | Limited to English use, Out-of-scope for legal violations |
|
Considerations: | Developers may fine-tune for languages beyond English. |
|
|
Supported Languages | |
Training Details |
Data Sources: | Publicly available online data, Publicly available instruction datasets, Human-annotated examples |
|
Data Volume: | 15 trillion tokens for pretraining, 10 million human-annotated examples for fine-tuning |
|
Methodology: | Supervised fine-tuning, Reinforcement Learning with Human Feedback (RLHF) |
|
Context Length: | |
Hardware Used: | |
Model Architecture: | Optimized transformer architecture with Grouped-Query Attention (GQA) |
|
|
Safety Evaluation |
Methodologies: | Red-teaming exercises, Adversarial evaluations, Safety mitigations |
|
Findings: | Model is less likely to falsely refuse prompts than Llama 2 |
|
Risk Categories: | Misuse, Critical risks, such as CBRNE, Cybersecurity, Child Safety |
|
Ethical Considerations: | Developers should implement safety best practices and assess residual risks. |
|
|
Responsible Ai Considerations |
Fairness: | Model is designed to be open, inclusive, and helpful. |
|
Transparency: | Encourages community collaboration for transparency and safety |
|
Accountability: | Developers responsible for appropriate use and safety evaluations. |
|
Mitigation Strategies: | Purple Llama tools and Llama Guard safeguard measures are provided. |
|
|
Input Output |
Input Format: | |
Accepted Modalities: | |
Output Format: | |
|