elliot-mllm/localize-general-objects

Elliot Localize General Objects Upstream training splits; known explicitly identified test/eval rows excluded. Cross-dataset benchmark overlap is not guaranteed. Published as a raw, manually gated release; source annotation caveats remain. Task views reuse original image archives. Coordinates are normalized 0–1000. Within each query targets are sorted left-to-right then top-to-bottom. HF preview configs contain 10 examples per view, not the complete training split. Full training… See the full description on the dataset page: https://huggingface.co/datasets/elliot-mllm/localize-general-objects.