AwaRes Training Dataset

Look Where It Matters: High-Resolution Crops Retrieval for Efficient VLMs

Paper | Project Page | GitHub

This is the training dataset for AwaRes — a spatial-on-demand VLM inference framework that processes low-resolution images first and selectively retrieves high-resolution crops via tool-calling. AwaRes achieves comparable performance to Qwen2.5-VL-7B across six benchmarks while using only ~36% of the visual tokens, offering a 4.4× speedup over dynamic high-resolution methods.

Dataset Description

The dataset contains ~48.7K visual question-answering samples with oracle crop annotations, curated for training VLMs to learn a coupled decision policy: whether high-resolution detail is needed and where to acquire it.

Columns

Column	Type	Description
image	list[Image]	Low-res and high-res crop images
problem	string	VQA question posed to the model
solution	string	Ground-truth answer
original_question	string	Original question before prompt formatting
original_answer	string	Original answer
hr_helps	int	Binary label — whether high-resolution crops help answer the question
helped_crops	list	Which crop indices were useful
image_paths	list	Original file paths for the crop images
modified	int	Whether the sample instruction was rephrased during curation

Crop Region Index

The model can request 9 crop regions by index:

Index	Region	Coordinates
0	top-left	(0, 0, W/2, H/2)
1	top-right	(W/2, 0, W, H/2)
2	bottom-left	(0, H/2, W/2, H)
3	bottom-right	(W/2, H/2, W, H)
4	center	(W/4, H/4, 3W/4, 3H/4)
5	top	(0, 0, W, H/2)
6	bottom	(0, H/2, W, H)
7	left	(0, 0, W/2, H)
8	right	(W/2, 0, W, H)

Usage

from datasets import load_dataset
ds = load_dataset("NimrodShabtay1986/awares-temp", split="train")
sample = ds[0]
lr_image = sample["image"][0]  # low-res
hr_image = sample["image"][1]  # high-res

Extracting Crop Regions from High-Res Image

from PIL import Image
CROP_REGIONS = {
    0: ("top-left",     lambda w, h: (0,     0,     w//2,   h//2)),
    1: ("top-right",    lambda w, h: (w//2,  0,     w,      h//2)),
    2: ("bottom-left",  lambda w, h: (0,     h//2,  w//2,   h)),
    3: ("bottom-right", lambda w, h: (w//2,  h//2,  w,      h)),
    4: ("center",       lambda w, h: (w//4,  h//4,  3*w//4, 3*h//4)),
    5: ("top",          lambda w, h: (0,     0,     w,      h//2)),
    6: ("bottom",       lambda w, h: (0,     h//2,  w,      h)),
    7: ("left",         lambda w, h: (0,     0,     w//2,   h)),
    8: ("right",        lambda w, h: (w//2,  0,     w,      h)),
}
def get_crop(hr_image: Image.Image, crop_index: int) -> Image.Image:
    """Extract a crop region from the high-res image.
    Args:
        hr_image: The high-resolution PIL Image.
        crop_index: Integer 0-8 specifying which crop region to extract.
    Returns:
        Cropped PIL Image for the requested region.
    """
    if crop_index not in CROP_REGIONS:
        raise ValueError(f"crop_index must be 0-8, got {crop_index}")
    w, h = hr_image.size
    _, box_fn = CROP_REGIONS[crop_index]
    return hr_image.crop(box_fn(w, h))
# Example: extract the crops that help for a given sample
sample = ds[0]
hr_image = sample["image"][1]
helped = sample["helped_crops"]  # e.g. ['3'] or ['all']
if helped and helped != ["all"]:
    crops = [get_crop(hr_image, int(idx)) for idx in helped]

Data Curation Pipeline

Samples were constructed automatically via a three-stage pipeline:

1. Resolution comparison: Qwen2.5-VL-7B generates answers at both low and high resolution. A LLaMA-3.3-70B judge compares outputs to label whether high-resolution crops help (hr_helps).
2. Oracle grounding: For samples where HR helps, an oracle grounding model(Qwen3-VL-235B-A22B) localizes the visual evidence for the correct answer, which is mapped to a discrete crop set.
3. Trajectory construction: Labels are converted into multi-turn tool-use trajectories for SFT cold-start training, followed by multi-turn GRPO with composite rewards (semantic correctness + crop-cost penalty).

Citation

@article{shabtay2026look,
  title={Look Where It Matters: High-Resolution Crops Retrieval for Efficient VLMs},
  author={Shabtay, Nimrod and Kimhi, Moshe and Spector, Artem and Haray, Sivan and Rivlin, Ehud and Baskin, Chaim and Giryes, Raja and Schwartz, Eli},
  journal={arXiv preprint arXiv:2603.16932},
  year={2026}
}

Links

• arXiv Paper
• Project Page
• GitHub