Sign in to view source links and access this dataset
Description
A large contrastive story dataset for mechanistic interpretability and alignment research. Each row contains two short matched narratives about the same situation, expressing opposing behavioral, affective, or epistemic poles. The dataset is based on a curated concept ontology with 148 classes and 1036 concept-antagonist pairs, authored by AntonKorznikov and last updated on July 15, 2026.
Use Cases
Training contrastive models based on paired narratives expressing opposing concepts.
Studying model representations based on behavioral and affective dichotomies defined in the ontology.
Developing interpretability tools based on the structured narrative guidance provided for each concept pair.
Strengths
Dataset is based on a curated concept ontology containing 148 distinct classes.
Includes 1036 specific concept-antagonist pairs for controlled contrastive study.
Provides narrative_guidance explaining the meaning of each conceptual dichotomy.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Last updated 2026-07-15 09:28:58; freshness should be verified.
Provenance
Source
huggingface, author AntonKorznikov
Collection Method
Likely manually curated or generated based on a defined concept ontology.
Freshness
Last updated 2026-07-15 09:28:58.
License is unknown; terms of use must be verified before application.