VagueUHR corpus and evaluation outputs for WeaveAgent: a two-stage tool-routing agent for ultra-high-resolution remote sensing imagery

Data and code supporting the manuscript "WeaveAgent: A Two-Stage Tool-Routing Agent for Ultra-High-Resolution Remote Sensing Imagery" (preprint: arXiv:2609.31234). Contents: (1) data/ — the VagueUHR corpus splits: a 5,000-record synthesis base, the 3,273-record routing-training split, 1,740-row aligned training demonstrations, the 1,000-record dual-split test set (600 intrinsic + 400 tool-requiring rows with tool-call trajectories), a clear-register control test (1,000) and an LLM-rewritten test variant (992); each row links image references, queries, tool-call trajectories and ground truths. (2) evaluation_outputs/ — per-record predictions and result summaries for every headline experiment in the paper (routing table, five-point repair ladder, oracle attribution, +/-image ablation, phrasing sensitivity, query-register matrix, GCC recall audit, native-tool baseline). (3) code/ — initial code release: the routing-first reward (R_WA2), answer-matching utilities, exact-McNemar and Wilson-interval statistics, and unit tests. Provenance: derived from the imagery and annotations of three public benchmarks (LRS-VQA, MME-RealWorld-RS, XLRS-Bench); imagery files are not redistributed — obtain them from the source benchmarks and join on the image references. The remaining pipeline components are released at github.com/Annafier/WeaveAgent upon paper acceptance.