ConceptFD contains training and inference data for ~20 very high-quality concepts sampled from the released by Gur-Arieh et. al., 2025 for Gemma-2-2B at layers 10 and 20.
Each subset comes with the following format:
input: Instructions sampled from publicly available instruction-tuning datasets for three genres: text, code, and math.
output: Model or LLM-generated responses. If output_concept is not EEEEE, then the output is generated by the… See the full description on the dataset page:
https://huggingface.co/datasets/pyvene/axbench-conceptFD.
