metadata
license: mit
task_categories:
- question-answering
language:
- en
configs:
- config_name: default
data_files:
- split: OmanicSynth
path: OmanicSynth.jsonl
- split: OmanicBench
path: OmanicBench.jsonl
Omanic
Omanic is an open-domain multi-hop question-answering (QA) resource designed for the step-wise evaluation of reasoning in Large Language Models (LLMs). It provides decomposed sub-questions and intermediate answers as structural annotations to help diagnose where reasoning failures occur.
- Paper: Omanic: Towards Step-wise Evaluation of Multi-hop Reasoning in Large Language Models
- Repository: https://github.com/XiaojieGu/Omanic
Dataset Summary
The dataset consists of two main components:
- OmanicSynth: A training set containing 10,296 machine-generated examples with structural reasoning chains, intended for supervised fine-tuning.
- OmanicBench: A high-quality evaluation benchmark containing 967 expert-reviewed, human-annotated examples for systematic assessment of multi-hop reasoning.
By providing intermediate steps rather than just final answers, Omanic allows researchers to analyze factual completeness and error propagation in the reasoning process of LLMs.