RsEvi-248K

RsEvi-248K contains 247,528 remote-sensing examples used for supervised fine-tuning of RsEvi-8B. Each example pairs an image and an instruction with a reasoning trace and a final answer. Bounding boxes in the traces identify the visual evidence used to answer the question.

Paper: Grounding What Matters: Evidence-Aware Credit Assignment for Remote Sensing Reasoning (EMNLP 2026)

Data

Task	Records
Visual question answering	100,000
Visual grounding	59,436
Scene classification	40,662
Object detection	24,448
Object counting	22,982

train.jsonl has one record per line, with messages, images, and task fields. Reasoning is enclosed in <think>...</think>, with visual evidence marked by <box>...</box>. Image paths are relative to the dataset root; extract the image_shards/images-*.tar archives there before using the JSONL file.

Examples

Three examples from the training set. Questions are shortened for display. The original reasoning is split into paragraphs, with box tags shown as labels and coordinates.

Aircraft at a terminal

Object detection · 4 evidence boxes

Question: Detect all airplanes and return their bounding boxes.

Reasoning

I scan the image to detect all airplanes.

Starting from the upper left, I observe a large airplane parked near the terminal: airplane [163,46,416,445].

Moving to the right side of the image, I identify another airplane positioned near the jet bridge: airplane [825,343,973,545].

Further down in the lower right area, I spot a third airplane: airplane [719,573,854,759].

Finally, near the bottom center, I detect a fourth airplane parked at a gate: airplane [570,751,722,919].

These are all the airplanes visible in the scene.

Answer

[
  [163, 46, 416, 445],
  [825, 343, 973, 545],
  [719, 573, 854, 759],
  [570, 751, 722, 919]
]

---

Tennis courts beside a marina

Object counting · 6 evidence boxes

Question: How many tennis courts are there in the image?

Reasoning

I examine the land area adjacent to the marina for recreational facilities. In the lower-left portion of the image, there is a distinct red rectangular complex containing multiple green courts marked with white lines.

Scanning this area, I identify two courts in the upper section of the complex:

tennis court [87,650,133,738] and
tennis court [123,637,168,716].

Below these, I see a second row of four courts arranged side-by-side:

tennis court [97,764,147,860],
tennis court [133,741,179,840],
tennis court [174,723,219,818], and
tennis court [207,708,251,800].

Counting these individual courts results in a total of 6.

Answer: 6

---

Identifying a storage facility

Scene classification · 3 evidence boxes

Question: What type of scene is shown in this image?

Reasoning

The image shows an aerial view of a facility containing multiple large cylindrical structures.

I can identify several white circular tanks, such as the one at storage tanks [685,229,825,356], along with others scattered across the site like storage tanks [232,423,333,511] and storage tanks [359,251,459,338].

These structures are connected by a network of pipes and surrounded by containment berms, which is characteristic of an industrial storage facility for liquids or gases.

The presence of these specific structures strongly points to the classification of storage tanks.

Answer: storage tanks

Citation

@inproceedings{miao-etal-2026-grounding,
  title = {Grounding What Matters: Evidence-Aware Credit Assignment for Remote {S}ensing Reasoning},
  author = {Miao, Ziqi and Xie, Zhanxu and Han, Liangzhe and Xiong, Yuan and Qian, Chen and Li, Lijun and Lv, Weifeng and Sun, Leilei},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year = {2026},
  pages = {51598--51627},
  publisher = {Association for Computational Linguistics},
  url = {https://aclanthology.org/2026.emnlp-main.2269/}
}