| Model Type | | text generation, dialogue, instruction tuned |
|
| Use Cases |
| Areas: | | Commercial use, Research use |
|
| Applications: | | Dialog and assistant-like applications |
|
| Primary Use Cases: | | Assistant-like chat and instruction tasks |
|
| Limitations: | | Inapplicable in languages other than English without additional tuning., Requires appropriate safety tuning for specialized applications. |
|
| Considerations: | | Developers should employ additional safeguarding measures and consider context when deploying applications. |
|
|
| Additional Notes | | Quantized versions available for resource-efficient deployments. |
|
| Supported Languages | | en (Primary support for dialogue and instructional tasks in English.) |
|
| Training Details |
| Data Sources: | | publicly available online data, SlimPajama-627B, UltraChat |
|
| Data Volume: | | 15 trillion tokens (pretraining) |
|
| Methodology: | | Supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF). NTK-aware interpolation technique for context length adjustment. |
|
| Context Length: | |
| Training Time: | | Total of 7.7M GPU hours for pretraining across multiple models. |
|
| Hardware Used: | | Crusoe Energy high performance L40S cluster |
|
| Model Architecture: | | Optimized transformer using RoPE theta for extended contexts. |
|
|
| Safety Evaluation |
| Methodologies: | | red-teaming, adversarial evaluations |
|
| Findings: | | Improved refusal rate for false prompts in comparison to Llama 2., Limited residual risks assessed through community tools. |
|
| Risk Categories: | | Child safety, Cybersecurity vulnerabilities, CBRNE threats |
|
| Ethical Considerations: | | Guided by a Responsible Use Guide and supporting tools like Meta Llama Guard 2. |
|
|
| Responsible Ai Considerations |
| Fairness: | | Focused on openness, inclusivity, and helpfulness while respecting diverse values and perspectives. |
|
| Transparency: | | Incorporates feedback from the community to continually assess safety and alignment. |
|
| Accountability: | | Developers undertaking use are held to the standards under the Acceptable Use Policy. |
|
| Mitigation Strategies: | | Use community safeguards like Llama Guard to supplement model-level safety. |
|
|
| Input Output |
| Input Format: | | Text input following a conversational template. |
|
| Accepted Modalities: | |
| Output Format: | | Generated text and code responses. |
|
| Performance Tips: | | Leverage context window optimizations for handling extensive lengths. |
|
|
| Release Notes |
| Version: | |
| Date: | |
| Notes: | | Finalized assistant-like chat optimizations and context length extension techniques. |
|
|
|