Concept16K contains training and inference data for 16K concepts randomly sampled from the released GemmaScope concept list for Gemma-2-2B-it at 20, and Gemma-2-9B-it at layers 20.
At the time of our AxBench release, Concept16K is by far the largest supervised dictionary learning (SDL) dataset for LLMs.
Cost: With our initial implementation, construct training data for each concept costs less than $0.01 with 72 positive examples and 72… See the full description on the dataset page:
https://huggingface.co/datasets/pyvene/axbench-concept16k.
