Loading request...
The user wishes for a better system to distribute benchmark data, particularly for LLMs, that prevents it from being inadvertently used in subsequent model training data. This is crucial for maintaining the integrity of benchmarks and avoiding data contamination.
[Weekend read] The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning - https://t.co/nrrHgIZUxc A new extensive benchmark to evaluate #LLM dangerous bio and cyber capabilities written by experts. I really like their approach to find precursor tasks to test without having dangerous data in the benchmark. This paper also highlights how expensive it is to develop complex benchmarks – over $200k for about 4k questions written by experts. Last but not least the authors carefully thought of making the data available as a zip with a password - that’s important to avoid model training on those benchmarks - I wish we had a better system for making benchmark data available while avoiding the risk of them ending up in subsequent models training data. Happy Reading 🙂 #AI #cybersecurity #LLM #Safety #LM