synth4indiref
Synthesize NL dialogues from algorithmically generated complex scenarios to create challenging comprehension questions.
A data synthesis pipeline developed as part of the Frame of Reference project. It algorithmically generates complex situated dialogue scenarios — involving spatial and temporal relational references — and produces natural language dialogues along with challenging comprehension questions as a proxy for grounding ability.
The goal is to create training and evaluation data for models that need to track shared context across extended conversations.