This repository contains the ARC dataset, including both the original human-annotated version and extended versions created using Large Language Models (LLMs).
As of now, only 10% of the complete dataset has been made public.
We plan to release the full dataset, including all human annotations and LLM-generated extensions, contingent upon the acceptance of our accompanying paper at LREC.
This folder houses the 10% sample of the core datasets. It is organized into three sub-folders, with each containing train.json, dev.json, and test.json files:
81_ARC_human_annotated: A sample of the original, human-annotated dataset.188_ARC_moderate_extension: A sample of the moderately extended version of the dataset.444_ARC_full_extension: A sample of the complete, fully extended dataset.
TANL/&DENIM/: These folders contain the scripts that were run on the datasets to generate performance metrics.Prompts/: This folder contains the specific prompts that were used with LLMs to extend the original dataset.