| Model Type | |
| Use Cases |
| Areas: | | Research, Commercial Applications |
|
| Applications: | | AI assistants, Text classification, Summarization, Question Answering, Retrieval Augmented Generation (RAG), Code-related tasks, Function-calling, Multilingual dialog |
|
| Primary Use Cases: | | General instruction response |
|
| Limitations: | | Performance may not match English tasks for multilingual dialog., Requires safety testing and tuning before deployment. |
|
| Considerations: | | Use with proper safety measures; introduce few-shot examples for improved outputs. |
|
|
| Additional Notes | | User testing across selected domains recommended to ensure safety. |
|
| Supported Languages | | English (Proficient), German (Intermediate), Spanish (Intermediate), French (Intermediate), Japanese (Intermediate), Portuguese (Intermediate), Arabic (Intermediate), Czech (Intermediate), Italian (Intermediate), Korean (Intermediate), Dutch (Intermediate), Chinese (Intermediate) |
|
| Training Details |
| Data Sources: | | publicly available datasets with permissive license, internal synthetic data, human-curated data |
|
| Data Volume: | |
| Methodology: | | Supervised finetuning, reinforcement learning for model alignment, model merging |
|
| Context Length: | |
| Training Time: | |
| Hardware Used: | | IBM's super computing cluster, Blue Vela, with NVIDIA H100 GPUs |
|
| Model Architecture: | | Decoder-only dense transformer architecture, including GQA, RoPE, MLP with SwiGLU, RMSNorm, and shared input/output embeddings |
|
|
| Safety Evaluation |
| Ethical Considerations: | | May produce inaccurate, biased, or unsafe responses. Usage requires proper safety testing and tuning. |
|
|
| Responsible Ai Considerations |
| Fairness: | | Aligned to ensure safety; however, multilingual performance might not match English task performance. |
|
| Transparency: | | Open source with comprehensive documentation, paper, and technical report. |
|
| Accountability: | |
| Mitigation Strategies: | | Few-shot examples can improve multilingual capabilities. |
|
|
| Input Output |
| Input Format: | | Structured chat format with prompts |
|
| Accepted Modalities: | |
| Output Format: | |
| Performance Tips: | | Fine-tuning with additional examples may improve specificity and accuracy. |
|
|
| Release Notes |
| Version: | |
| Date: | |
| Notes: | | Model released with refined capabilities and aligned for structured instruction following. |
|
|
|