language:
- en
size_categories: - 1K<n<10K
AI21 Jamba-Specific Enkrypt Alignment Dataset
Overview
The AI21 Jamba-Specific Enkrypt Alignment Dataset is a targeted dataset created by Enkrypt AI to improve the alignment of the AI21 Jamba-1.5-mini model. This dataset was developed using insights gained from Enkrypt AI’s custom red-teaming efforts on the Jamba-1.5-mini model.
Data Collection Process
- Enkrypt AI leveraged its proprietary SAGE-RT (Synthetic Alignment data Generation for Safety Evaluation and Red Teaming) pipeline to identify vulnerabilities in the Jamba-1.5-mini model.
- Once vulnerabilities were detected, SAGE was used to refine data collection, ensuring that the dataset focused on high-risk failure modes rather than random sampling.
- This dataset was designed specifically for AI21, supplementing a more generalized SAGE preference dataset (which is owned by Enkrypt AI) to improve Jamba-1.5-mini’s safety, robustness, and ethical alignment.
Dataset Composition
- The dataset includes high-risk adversarial prompts and responses, spanning categories like bias, misinformation, unethical content generation, discrimination, and harmful language.
- It targets vulnerabilities unique to AI21 Jamba-1.5-mini, ensuring precision-tuned alignment rather than general safety improvements.
- Distribution of Categories: The dataset is structured to almost uniformly represent all the vulnerabilites of the Jamba-1.5-mini model.