manifesta/scientific-chart-qa-17k

Scientific Chart QA, 17,070 rows A multimodal chart-interpretation dataset built around one idea: teaching a model when not to answer matters as much as teaching it to answer. One in seven questions here cannot be answered from its figure, and the correct response is cannot be determined. Baseline vision-language models overwhelmingly guess a plausible-looking number instead. That is the behaviour this set targets. The four things worth… See the full description on the dataset page: https://huggingface.co/datasets/manifesta/scientific-chart-qa-17k.