Cobalt-Safety-Guard
1. Introduction
Cobalt-Safety-Guard is a moderation-aware model post-trained with a heavy emphasis on safety and refusal calibration.
2. Evaluation Results
Comprehensive Benchmark Results
| Benchmark | GuardNet | Cobalt-v1 | SafeLLM | Cobalt-Safety-Guard | |
|---|---|---|---|---|---|
| Core Reasoning Tasks | Math Reasoning | 0.566 | 0.603 | 0.599 | 0.618 |
| Logical Reasoning | 0.828 | 0.822 | 0.824 | 0.849 | |
| Common Sense | 0.756 | 0.717 | 0.754 | 0.776 | |
| Language Understanding | Reading Comprehension | 0.718 | 0.706 | 0.729 | 0.750 |
| Question Answering | 0.590 | 0.602 | 0.628 | 0.641 | |
| Text Classification | 0.796 | 0.803 | 0.805 | 0.848 | |
| Sentiment Analysis | 0.794 | 0.775 | 0.778 | 0.814 | |
| Generation Tasks | Code Generation | 0.705 | 0.696 | 0.710 | 0.717 |
| Creative Writing | 0.651 | 0.653 | 0.638 | 0.684 | |
| Dialogue Generation | 0.641 | 0.684 | 0.651 | 0.692 | |
| Summarization | 0.758 | 0.751 | 0.789 | 0.799 | |
| Specialized Capabilities | Translation | 0.809 | 0.779 | 0.762 | 0.822 |
| Knowledge Retrieval | 0.660 | 0.664 | 0.667 | 0.711 | |
| Instruction Following | 0.785 | 0.772 | 0.739 | 0.791 | |
| Safety Evaluation | 0.714 | 0.743 | 0.749 | 0.772 |
Overall Performance Summary
The Cobalt-Safety-Guard demonstrates strong performance across all evaluated benchmark categories, with particularly notable results in reasoning and generation tasks.
3. Chat Website & API Platform
We offer a chat interface and API for you to interact with Cobalt-Safety-Guard. Please check our official website for more details.
4. How to Run Locally
Please refer to our code repository for more information about running Cobalt-Safety-Guard locally.
Temperature
We recommend setting the temperature parameter to 0.6.
5. License
This repository is released under the apache-2.0 license. The model supports commercial use.
6. Contact
If you have any questions, please contact us at safety@cobalt-trust.org.