Model Type | |
Use Cases |
Areas: | |
Applications: | |
Primary Use Cases: | natural language generation tasks |
|
Limitations: | Use in languages other than English, Designed for a broad range of applications and may not meet every developer safety preference out-of-the-box |
|
Considerations: | Developers should tailor safety testing and tools according to their use cases |
|
|
Additional Notes | 4-bit quantization increases accessibility. |
|
Supported Languages | |
Training Details |
Data Sources: | publicly available online data |
|
Data Volume: | |
Methodology: | supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) |
|
Context Length: | |
Hardware Used: | Meta's Research SuperCluster, third-party cloud compute |
|
Model Architecture: | optimized transformer architecture |
|
|
Safety Evaluation |
Methodologies: | |
Findings: | significant reduction in false refusals compared to Llama 2 |
|
Risk Categories: | CBRNE, Cyber Security, Child Safety |
|
Ethical Considerations: | Limited misuse and harm reported, developers encouraged to perform safety evaluations |
|
|
Responsible Ai Considerations |
Fairness: | Commitment to reduce residual risks and focus on alignment and helpfulness |
|
Transparency: | Benchmarking standards made transparent |
|
Accountability: | Developers are responsible for ensuring the use is in line with the Llama 3 Community License and Acceptable Use Policy |
|
Mitigation Strategies: | Steps to limit misuse, open source community tools provided |
|
|
Input Output |
Input Format: | |
Accepted Modalities: | |
Output Format: | |
Performance Tips: | Use with transformers for optimal integration and use cases. |
|
|
Release Notes |
Version: | |
Date: | |
Notes: | Release of Meta Llama 3 family of models, including instruction-tuned and pretrained versions. |
|
|
|