Model Type | |
Use Cases |
Areas: | |
Applications: | Assistant-like chat with tuned models, Natural language generation |
|
Primary Use Cases: | Text generation for dialogue and other natural language tasks |
|
Limitations: | Limited to English, Not suitable for use that violates applicable laws or regulations |
|
Considerations: | Formatting needs to be followed for chat versions, including specific tokens and whitespace. |
|
Additional Notes | Foundation model with potential for further fine-tuning, especially for dialogue use cases. |
Supported Languages | |
Training Details |
Data Sources: | A new mix of publicly available online data |
|
Data Volume: | 2 trillion tokens |
Methodology: | Pretrained and fine-tuned, using supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) |
Context Length: | 4000 |
Training Time: | January 2023 to July 2023 |
Hardware Used: | Meta's Research Super Cluster and production clusters |
|
Model Architecture: | Optimized transformer architecture with Grouped-Query Attention (GQA) for improved inference scalability. |
|
Safety Evaluation |
Methodologies: | Internal evaluations using academic benchmarks |
|
Findings: | May produce inaccurate, biased, or objectionable responses in certain scenarios |
|
Risk Categories: | |
Ethical Considerations: | The model carries risks with use; outputs cannot be predicted in advance. |
|
Responsible Ai Considerations |
Fairness: | Uncertain; developers should perform safety testing and tuning tailored to their specific applications. |
Transparency: | The model's potential outputs cannot be predicted in advance. |
Accountability: | Meta is responsible for the initial deployment, but users must ensure safe application. |
Mitigation Strategies: | Perform safety testing and tuning tailored to specific applications. |
|
Input Output |
Input Format: | text |
Accepted Modalities: | |
Output Format: | text |
Performance Tips: | Adhere to specified input formatting for chat models. |
|
Release Notes |
Version: | 70B |
Date: | 2023-07-01 |
Notes: | 70 billion parameters for varied natural language generation tasks. |
|
|