UAVIT-1M UAV Visual Instruction Tuning Dataset

## Overview UAVIT-1M is an instruction-tuning dataset designed to enhance low-altitude UAV (drone) visual understanding in multimodal LLMs. It contains over 1 million samples (size category 1M<n<10M) in JSON format, covering 11 distinct image-level and region-level vision-language tasks. The dataset is in English and was released in May 2025 as part of an ongoing project that also includes the UAVBench evaluation benchmark. ## Schema - `image` — string — path/reference to the UAV image - `conversations` — list — multi-turn instruction/response pairs in standard LLaVA-style format - `task` — string — one of 11 task categories (image-level or region-level) - `id` — string — unique sample identifier - Additional fields may include region coordinates/bboxes for region-level tasks (Exact column set depends on task split; see source dataset page for per-file schema.) ## Sources - HuggingFace: https://huggingface.co/datasets/ZhanYang-nwpu/UAVIT-1M — License: CC-BY-4.0 - Publisher: ZhanYang-nwpu (Northwestern Polytechnical University) ## Methodology Per the source, UAVIT-1M was constructed to support instruction tuning of multimodal LLMs on low-altitude UAV imagery, covering 11 image-level and region-level tasks (e.g., captioning, VQA, grounding, region description). The authors aggregate UAV imagery and generate instruction-formatted conversations to enable visual instruction tuning. The companion UAVBench benchmark provides evaluation. Detailed data collection and curation methodology is described on the source dataset page and associated paper. ## Known gaps & limitations The dataset is described as ongoing and evaluation code plus inference tutorials are not yet released as of the source's last update. Coverage is limited to low-altitude UAV viewpoints and English-language instructions, which may not generalize to satellite, ground-level, or non-English use cases. Source does not extensively document sampling biases across geographies, scenes, or sensors; buyers should validate empirically for their target deployment. ## Intended use & out-of-scope - IS for: instruction tuning and fine-tuning of multimodal LLMs for UAV/drone visual understanding, region grounding, and VQA; supervised training data for low-altitude aerial vision tasks. - NOT for: general satellite/remote-sensing tasks, ground-level scene understanding, or evaluation without cross-checking overlap with the companion UAVBench (leakage risk if used as both train and eval). Original supplier listing: UAVIT-1M: UAV Visual Instruction Tuning Dataset (1M+) Largest instruction-tuning dataset for low-altitude UAV visual understanding, with 1M+ samples across 11 image- and region-level tasks. CC-BY-4.0.