9,980 8-way multiple-choice grade school science questions are paired with a supporting corpus of 17 million sentences. These instances are partitioned into 8,134 training, 926 development, and 920 test samples designed to test multi-hop reasoning and sentence composition.
Use Cases
- Train retrieval-augmented generation (RAG) models to extract supporting evidence from the 17M sentence corpus
- Benchmark multi-hop reasoning by requiring models to compose two distinct facts to select the correct answer from the 8-way multiple-choice options
- Fine-tune classifiers to identify the correct answer among the 8-way multiple-choice candidates based on grade school science knowledge
Strengths
- 9,980 8-way multiple-choice questions focused on grade school science
- Supporting corpus of 17,000,000 sentences for fact retrieval and composition
- Data split into 8,134 training, 926 development, and 920 test instances