| Model Type | | text generation, decoder-only |
|
| Use Cases |
| Areas: | | Research, Commercial applications |
|
| Applications: | | NLP research, Language learning tools, Content creation, Chatbots, Text summarization |
|
| Primary Use Cases: | | Text generation, Question answering, Summarization |
|
| Limitations: | | Biases from training data, Factual inaccuracies, Complex open-ended task challenges |
|
| Considerations: | | Consider using tools for de-biasing and content moderation |
|
|
| Additional Notes | | Gemma is part of the foundation models, offering benefits in Responsible AI development for accessibility and fostering innovation. |
|
| Supported Languages | | English (High proficiency) |
|
| Training Details |
| Data Sources: | | Web Documents, Code, Mathematics |
|
| Data Volume: | |
| Methodology: | |
| Hardware Used: | |
| Model Architecture: | |
|
| Safety Evaluation |
| Methodologies: | | Red-teaming, Structured evaluations |
|
| Findings: | | Within acceptable thresholds for Google's internal policies |
|
| Risk Categories: | | Content safety, Representational harms, Memorization, Large-scale harms |
|
| Ethical Considerations: | | Addressed filters for CSAM, sensitive information, aligning with Google AI principles |
|
|
| Responsible Ai Considerations |
| Fairness: | | Socio-cultural biases scrutinized, data pre-processing and evaluations reported. |
|
| Transparency: | | Summary details on architecture, capabilities, and limitations provided. |
|
| Accountability: | |
| Mitigation Strategies: | | Continuous monitoring, guidelines for content safety |
|
|
| Input Output |
| Input Format: | |
| Accepted Modalities: | |
| Output Format: | | Generated English-language text |
|
| Performance Tips: | | Use "$\tau$" with higher values for creative tasks and lower for educational tasks. |
|
|
| Release Notes |
| Version: | |
| Notes: | | Improvement over the original model, using RLHF leading to quality improvements. Addressed a bug for multi-turn conversations. |
|
|
|