| Model Type | | text-to-text, decoder-only, large language model |
|
| Use Cases |
| Areas: | | Content Creation and Communication, Research and Education |
|
| Applications: | | Text Generation, Chatbots and Conversational AI, Text Summarization, Natural Language Processing (NLP) Research, Language Learning Tools, Knowledge Exploration |
|
| Primary Use Cases: | | Content Creation, Communication, Research |
|
| Limitations: | | Training Data Quality, Context and Task Complexity, Language Ambiguity and Nuance, Factual Accuracy, Common Sense |
|
| Considerations: | | Careful consideration for potential biases and misinformation. Follow guidelines for responsible use. |
|
|
| Additional Notes | | Prohibited uses outlined in the Gemma Prohibited Use Policy. |
|
| Supported Languages | | English (fully supported) |
|
| Training Details |
| Data Sources: | | Web Documents, Code, Mathematics |
|
| Data Volume: | |
| Methodology: | | Training using a novel RLHF method |
|
| Hardware Used: | |
| Model Architecture: | | Text-to-text, decoder-only |
|
|
| Safety Evaluation |
| Methodologies: | | Red-teaming, structured evaluations |
|
| Risk Categories: | | Text-to-Text Content Safety, Text-to-Text Representational Harms, Memorization, Large-scale harm |
|
| Ethical Considerations: | | Evaluation within acceptable thresholds for meeting internal policies for categories such as child safety, content safety, representational harms, memorization, large-scale harms. |
|
|
| Responsible Ai Considerations |
| Fairness: | | LLMs trained on large-scale, real-world text data can reflect socio-cultural biases embedded in the training material. |
|
| Transparency: | | Model card summarizes details on the models' architecture, capabilities, limitations, and evaluation processes. |
|
| Accountability: | | Transparency and outlining measures for responsible usage. |
|
| Mitigation Strategies: | | Perpetuation of biases: Continuous monitoring, evaluation metrics, human review, and exploration of de-biasing techniques. |
|
|
| Input Output |
| Input Format: | | Text string, such as a question, a prompt, or a document to be summarized. |
|
| Accepted Modalities: | |
| Output Format: | | Generated English-language text. |
|
| Performance Tips: | | Provide longer context for better outputs, up to a certain point. |
|
|
| Release Notes |
| Version: | |
| Date: | |
| Notes: | | Trained using a novel RLHF method, substantial gains in quality and capabilities, bug fixes in multi-turn conversations. |
|
|
|