Changes in this version:
1. removed some templates
2. fixed all object names to be singular
3. filtered names that are tokenized to a single token on GPT-2, Mistral, Phi, Llama, Qwen and Pythia models.
4. 10k samples in each split
5. added abc counterfactuals
A split of the IOI dataset from the paper "Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small" by Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, Jacob Steinhardt.
