The dataset comprises question answer pairs generated by the Mistral-7B-Instruct-v0.3 model, over a sample of the documents available in the CiGi knowledge base. The generative Q-A creation for evaluating CiGi was preferred over manual annotation because it yields diverse queries whose distribution better reflects downstream user intents. The dataset presentes the question-answer pairs, along with a reference to the source snippet that was used by the model to construct the answer.
- Downloads last month
- 10