| Model Type | |
| Use Cases |
| Areas: | | Synthetic Data Generation, building and customizing LLMs |
|
| Applications: | | Chat applications, AI assistant |
|
| Primary Use Cases: | |
| Limitations: | | Amplifies biases from training data, may generate socially undesirable text |
|
|
| Supported Languages | | languages_supported (Multilingual), proficiency_levels () |
|
| Training Details |
| Data Sources: | | 9 trillion tokens of English based texts, 50+ natural languages, and 40+ coding languages |
|
| Methodology: | | Supervised Fine-tuning (SFT), Direct Preference Optimization (DPO), Reward-aware Preference Optimization (RPO), Grouped-Query Attention (GQA), Rotary Position Embeddings (RoPE) |
|
| Context Length: | |
| Training Time: | |
| Model Architecture: | |
|
| Safety Evaluation |
| Methodologies: | | Adversarial testing via Garak, AEGIS content safety evaluation, Human Content Red Teaming |
|
| Risk Categories: | | Toxic language, unsafe content, societal biases |
|
| Ethical Considerations: | | NVIDIA believes Trustworthy AI is a shared responsibility. |
|
|
| Input Output |
| Input Format: | | Single Turn: System User {prompt} Assistant; Multi-Turn: User {prompt 1} Assistant {response 1} User {prompt 2} Assistant {response 2}... |
|
| Output Format: | |
|
