| Model Type | | text generation, instruction-tuned |
|
| Use Cases |
| Areas: | | Commercial applications, Research |
|
| Applications: | | Assistant-like chat, Natural language generation tasks |
|
| Primary Use Cases: | | Dialogue, Instruction following |
|
| Limitations: | | Limited to English applications, Potential outputs: inaccurate, biased, objectionable |
|
| Considerations: | | Safety testing and tuning required before deployment |
|
|
| Supported Languages | |
| Training Details |
| Data Sources: | | publicly available online data, SlimPajama dataset, UltraChat dataset |
|
| Data Volume: | | 15 trillion tokens (pretraining), 10M human-annotated examples (fine-tuning) |
|
| Methodology: | | Supervised fine-tuning, Reinforcement learning with human feedback |
|
| Context Length: | |
| Hardware Used: | |
| Model Architecture: | | Auto-regressive transformer with optimized architecture |
|
|
| Safety Evaluation |
| Methodologies: | | Red teaming, Adversarial evaluations |
|
| Risk Categories: | | CBRNE, Cybersecurity, Child Safety |
|
| Ethical Considerations: | | Assess responses related to adversarial risks and CBRNE threats |
|
|
| Responsible Ai Considerations |
| Fairness: | | Fairness considerations are interwoven with model alignment strategies. |
|
| Transparency: | | Models openly released for safety evaluation and transparency. |
|
| Accountability: | | Model developers are accountable for ensuring alignment and safety. |
|
| Mitigation Strategies: | | Meta Llama Guard and Code Shield as safeguards. |
|
|
| Input Output |
| Input Format: | |
| Accepted Modalities: | |
| Output Format: | |
| Performance Tips: | | Apply supervised fine-tuning or RLHF for specific applications. |
|
|