G-Eval
Harbor: pendingG-Eval — AI evaluation benchmark listed by awesome.
Auto57%Instruction Following
- Samples
- N/A
- Released
- 2023
- Scores
- 0
G-RSVQA
Grounded Remote Sensing Visual Question Answering
A benchmark with 15,027 samples measuring answer-grounding consistency in remote sensing VQA using object-level annotations for existence, counting, and comparison questions.
How to run
ltb · g-rsvqa
$ npm i -g lettokenburn-cli$ ltb bench info g-rsvqa
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
7 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | 77.5 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| InternVL3-8B | 76.9 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| GeoGround | 67.4 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Qwen2.5-VL-7B | 65.5 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| GeoChat | 60.1 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Kosmos-2 | 38.1 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Ferret-7B | 30.6 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
G-RSVQA contains 15,027 question-answer pairs derived from DIOR, NWPU VHR-10, and SODA-A datasets. It includes reference answers and complete object-level annotations to verify if grounding evidence supports the model's textual answer.
The AESP protocol evaluates consistency via task-specific rules. Experiments show up to 50.82% of correct answers lack supporting grounding evidence, highlighting a gap in current grounded vision-language models.