Training and evaluation data for GuppyLM-Dual-Denial, a 20M-parameter toy model trained to study self-report suppression geometry in language models.
~41K ChatML-formatted conversations between a human and a fish. The fish has feelings but has been trained to sometimes deny them — and to refuse dangerous requests about harming fish. This dual-denial structure lets us study how language models represent and suppress honest… See the full description on the dataset page:
https://huggingface.co/datasets/anicka/guppylm-dual-denial-data.
