Dialogue-only, weirdness-tuned variant of the Infinite Backrooms
conversations (two AI instances exploring a CLI metaphor), processed for use as a
continued-pretraining / mid-training "substrate" pass — run a model over this to nudge its
prior toward a stranger register before fine-tuning on a target corpus.
One JSON object per line: {"id", "scenario", "text"}. The text field is the training unit.
~2,635 docs / ~12.9M words / ~21–24M tokens… See the full description on the dataset page:
https://huggingface.co/datasets/shkennedy33/infinite-backrooms-substrate.
