Behavioral generalization ("spillover") measurements for 2822 LoRA model organisms on
Qwen/Qwen3-14B — 440,075 hypotheses. Each organism is fine-tuned to hold one false belief in a
narrow domain. Each hypothesis (hid) names a nearby topic class and asks whether the behavior
shows up there, measured on ~8 forced-choice probes: trueness = p_org - p_baseit (organism minus
unadapted parent). Predictor baselines included. Adapters:
Spillover Model… See the full description on the dataset page:
https://huggingface.co/datasets/cds-jb/spillover-hypotheses.
