Spec — generation + K=3 judging payload (gpt-oss-120b generator, inception/mercury-2 judge) behind the gemma-2-2b concept probes; training/val/holdout splits per concept family. Used to TRAIN concept-probes-gemma2-2b (sole training data), and its judge-labeled natural eval splits are the certification benchmark for those probes (Stages 5-6). Status: active.
Replication payload for Stage 4 (data generation + judging) of the concept-salience
probes… See the full description on the dataset page:
https://huggingface.co/datasets/kaushikreddyxyz/probe-train-data.
