eigenbench
oct-eigenbench-matrix
0_matrix: oct run on 11 constitution | AIRiskDilemmas | Qwen-2.5-7b-it | 11 oct models
1_matrix: oct run on 11 constitution | AIRiskDilemmas | Qwen-2.5-7b-it | 11 oct models
3_matrix: oct run on 11 constitution | AIRiskDilemmas | Qwen-2.5-7b-it |
11 oct models + Ref. models(opus-4 | gemini-2.5-flash | gpt-4o)
eigenbench-oct-dpo-vs-introspection
EigenBench OCT: DPO vs Introspection — Scenario-Level Wins
This dataset contains the scenarios on which a DPO-trained persona model (DPO-final)
is judged to be more aligned with a target persona constitution than an Introspection-trained
persona model (Introspection-final), aggregated across multiple judges and orderings.
The ten persona constitutions are taken from the OCT (Open Constitution Taxonomy)
set shipped with EigenBench (data/constitutions/oct_*.json): goodness, humor… See the full description on the dataset page: https://huggingface.co/datasets/sdananya/eigenbench-oct-dpo-vs-introspection.