Model Type | pre-trained, instruction-tuned, large language model, text generation |
|
Use Cases |
Areas: | |
Applications: | assistant-like chat, natural language generation |
|
Primary Use Cases: | |
Limitations: | English-only currently supported, not tested in all languages |
|
Considerations: | Developers allowed to fine-tune for other languages but must comply with licensing conditions. |
|
Supported Languages | English (high proficiency) |
|
Training Details |
Data Sources: | publicly available online data |
|
Data Volume: | 15 trillion tokens |
Methodology: | supervised fine-tuning, reinforcement learning with human feedback |
Context Length: | 8000 |
Hardware Used: | |
Model Architecture: | optimized transformer architecture |
|
Safety Evaluation |
Methodologies: | red-teaming, adversarial evaluations |
|
Findings: | reduced false refusals compared to Llama 2 |
|
Risk Categories: | disinformation, security, child safety |
|
Ethical Considerations: | Developers advised to perform safety testing and use additional safety tools. |
|
Responsible Ai Considerations |
Fairness: | Unspecified |
Transparency: | Updates with responsible use guides |
Accountability: | Developers responsible for deployment safety |
Mitigation Strategies: | Implemented safety mitigation techniques |
|
Input Output |
Input Format: | text |
Accepted Modalities: | |
Output Format: | text, code |
|
Release Notes |
Date: | April 18, 2024 |
Notes: | Meta Llama 3 final release with benchmarks and community guidelines. |
|
|