opencompass
opencompass copied to clipboard
[Update] Enhancements and Fixes in Needlebench
-
Optimized Needlebench Configurations
- Streamlined the writing of Needlebench config files by eliminating redundant code and improving structure.
- Reduced code duplication across multiple configuration files, making the setup more efficient and easier to manage.
-
Updated Multi-Needle-Reasoning Task
- Modified the Multi-Needle-Reasoning task to use ATC-specific needles, aligning the task with the ATC dataset requirements.
- Adjusted the configuration settings to better suit the ATC dataset, ensuring accurate and relevant results.
-
ATC Fixes
- The default configuration for ATC has been changed to use a fill-in-the-blank format \boxed{}.
- Additionally, the diversity of the tasks has been increased.