GeoHop — Multi-Hop Remote-Sensing VQA (benchmarks + weights)

Clean v1/v2/v3 layout for syncing experiment progress across machines.

Layout

• benchmarks/v1, v2, v3 — question/answer splits (see each version's README).
• images/<dataset> — image corpora. geollava / lrs_gro / xlrs_vg are shared by all versions; bright / xbd / sn7 are NEW v3 candidate sources.
• weights/<date_tag> — best checkpoint(s) per milestone (see weights/WEIGHTS_INDEX.md).
• _archive/ — historical bench versions, build inputs (mhrs_bench/{relations,generated}), and non-bench experiments (uhrcoz, eval_outputs). Nothing was deleted in the reorg.

Benchmark versions

ver	what	splits	class format
v1	mhrs_bench v1_5_1 — historical ruler (greedy 31.44% / Pass@4 46.51%)	rl train 1100 / test 458 · sft train 3055 / val 276	open-vocab
v2	frozen 2026-06-15, granularity-refined (class coarsened to a shared taxonomy)	same	open-vocab coarse
v3	MCQ: class → 4-option single-answer; color/shape/distance unchanged. + candidate sources bright/xbd/sn7 (images+questions only, NOT in train/eval splits)	same + candidates/	4-option MCQ

v3 candidate taxonomy: aircraft/airfield/building/containers/road/ship/sports facility/vehicle. v3 image references resolve under images/<dataset>/ (see benchmarks/v3/README.md).

Weights — see weights/WEIGHTS_INDEX.md.