Single-question activation-probing dataset built from the raw rollouts at
cds-jb/synthweb-gemma4-26b-a4b
(FineWeb prefix + google/gemma-4-26b-a4b continuations). A simpler, one-question-per-item
analog of cds-jb/synthweb-qwen3-8b-multiscale-inference: no atom/scope/lens machinery.
Each row asks one question whose answer the activation oracle M should
recover from a source model's (google/gemma-4-26b-a4b) activations over an extraction
region, sampled… See the full description on the dataset page:
https://huggingface.co/datasets/cds-jb/synthweb-qa-gemma4-26b-a4b.