ViRA-Train

This repository holds ViRA-Train, the training data of the paper "ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relative Policy Optimization".

Set accuracy on ViRA-Bench and ViRA-Bench-Real at five pixel budgets, and accuracy on general benchmarks at each model's maximum resolution. The bars of each group are sorted by height. Claude Haiku 5.5 and GPT-6-Luna run at their highest thinking level.

ViRA-Train is the training data of ViRA-9B and ViRA-27B, visual agents that zoom into high-resolution images. Each row is one question asked on a set of num_versions = K images, with a correct answer on each image:

Code for training: github.com/huawei-lin/ViRA (python -m vira.hf train downloads this dataset and writes the training files of ViRA-9B and ViRA-27B).

• K = 3 or 2: contrast sets. Versions of one photo from Wikimedia Commons (21 to 120 MP, median 30 MP): the original O and two variants A and B. In A and B one small region was changed (a piece of text, or an object's color, symbol, shape, direction or state), so the versions are the same photo except at that region, and the question has a different correct option on each version. The region covers a median of 0.022% of the photo; its box diff_bbox is given in full-resolution pixels. Its shorter side is at least 20 px at full resolution, and its longer side has a median of 16 px when the photo is reduced to 1 MP. Sets with K = 2 have the variants A and B only; they ask for reasoning about the changed words, for example what a sign allows.
• K = 1: single images. Questions about small details in Commons photos, questions about several objects in Commons photos and maps, synthetic charts and tables, and math problems.

The train split has 6,457 contrast sets on 3,797 photos, of which 5,329 concern text and 1,128 concern objects, and 2,804 single images. No photo is in ViRA-Bench. The contrast sets also exclude every photo that shares a title series with a ViRA-Bench photo, has a 64-bit perceptual hash within distance 10 of a ViRA-Bench photo, or has the same author and a Commons page number within 2,000 of a ViRA-Bench photo. The dev split is the validation set of training, and its photos are disjoint from the train photos.

	train	dev
K = 3	4,110	126
K = 2	2,347	0
K = 1	2,804	200
questions	9,261	326
image files	14,454	456

type	K	content	train	dev
contrast_object	3	contrast set in which an object differs between the versions	1,128	126
contrast_text	3 or 2	contrast set in which a piece of text differs between the versions	5,329	0
detail	1	open question about one small detail of a photo	967	100
multi_evidence_photo	1	question about several objects in one photo	1,094	57
multi_evidence_map	1	question about several objects in one map or plan	201	13
chart	1	question about a synthetic chart or table (12-16 MP)	293	15
math	1	math problem (geometry, tables, arithmetic)	249	15

Every row has one image per version: multi_evidence_photo and multi_evidence_map ask about several places of a single image, not about several input images.

Training subsets

in_vira9b_train and in_vira27b_train mark the rows that ViRA-9B (trained from Qwen3.5-9B) and ViRA-27B (trained from Qwen3.8-27B) were trained on:

	ViRA-9B	ViRA-27B
contrast sets	1,664	1,338
detail	599	600
multi_evidence_photo	859	949
multi_evidence_map	170	175
chart	200	200
math	150	139
total	3,642	3,401

Each subset holds the data that give its base model a learning signal. Its contrast sets are the sets that the base model does not answer correctly on every version when it views the whole photo at 12.8 MP without a zoom tool. Its single images are neither too easy nor unsolvable for the base model in 8 attempts: detail questions are answered correctly in 1 to 6 attempts with the zoom tool; multi_evidence_photo, multi_evidence_map and chart questions in fewer than 7 attempts without the tool and in at least 1 with it; math problems in 1 to 6 attempts without the tool.

data/vira9b_train_order.txt and data/vira27b_train_order.txt list the qids in training order: each row was used once, 16 rows per step, in this order. The dev split (126 contrast sets, 100 detail questions and 100 questions of the other single-image parts) was evaluated every 20 steps during training.

Files

data/*.parquet holds one row per question; images/<photo_group>/ holds the full-resolution images. Contrast sets: <photo>_O.jpg (or <photo>_O_<pair>.jpg), <pair>_A.jpg, <pair>_B.jpg (JPEG quality 95, no chroma subsampling). Single images: <photo_group>_O.<ext>, the file as downloaded (Commons photos and maps) or as generated (charts).

from datasets import load_dataset
from huggingface_hub import hf_hub_download
ds = load_dataset("huaweilin/ViRA-Train", split="train")
row = ds[0]
paths = [row[f"image_{v}"] for v in "oab" if row[f"image_{v}"]]          # K = len(paths) = row["num_versions"]
files = [hf_hub_download("huaweilin/ViRA-Train", p, repo_type="dataset") for p in paths]   # full resolution

column	meaning
qid, group_id	question id (the id used in training); group_id = qid (the rollouts on all versions of a question form one set)
num_versions	K, the number of images the question is asked on (3: O, A, B; 2: A, B; 1: O only)
pair_id, photo_group	contrast sets: id of the image set (an image set can have several questions); every row: source photo or image
type	type of the row (see above)
kind	contrast sets: object / text; single images: the question type (see below)
obj_kind, qtype	contrast sets only: change kind of the object (given for a part of the contrast_object rows); perception or reasoning
question, options	the question and its options (one order for all versions; empty for open questions)
answer_type	letter (multiple choice, the answer is a letter) or open (free-form answer)
answer_o, answer_a, answer_b	correct answer on O, A, B (answer_o is empty for K = 2)
image_o, image_a, image_b	repo paths of the full-resolution images (empty when the image is not used)
zoom_o, zoom_a, zoom_b, preview	display only: the evidence enlarged; the image downscaled with the evidence outlined
image_size	[width, height]
evidence_boxes	[left, top, right, bottom] boxes of the evidence in full-resolution pixels (contrast sets: [diff_bbox]; empty for math)
diff_bbox, core_bbox	contrast sets only: region that differs between the versions; box of the object / text the question is about
right_native_<model>, right_1mp_<model>	contrast sets only: versions (letters o, a, b) the model answered correctly, see Difficulty
hard_native_unsolved	contrast sets only: no model answered every version correctly at native resolution (hard, or possibly a label error)
seam_shift_px	contrast sets: measured offset of the pasted region against O (px; 0 = none found; empty when not measured)
in_vira9b_train, in_vira27b_train	the row was in the training data of ViRA-9B / ViRA-27B
license, attribution, source_url	license, credit and source page of the image
meta	everything else, as JSON

Question types (kind) of the single images: detail rows give the kind of detail (text, object, number, shape, count, position, symbol, direction, commonsense, reasoning); multi_evidence_photo and multi_evidence_map rows are position (left/right or above/below of named objects), chain (a detail of an object singled out through another object), order (which of several similar objects, counting in a given direction, has a property), count, attribute_or_absence (color, material or shape of a small object, or that it is not in the image) or reading_reasoning. The options of position, order and attribute_or_absence were computed by a program from the object boxes, so that their wording cannot reveal the answer.

Scoring

A letter answer is correct when it equals the correct letter. Open answers were judged in training by GPT-4o (gpt-4o-2024-11-20, temperature 0) with the judge prompt of Mini-o3.

Difficulty

Three vision-language models answered every contrast set on each version, without tools and with greedy decoding, once with the whole photo at native resolution (at most 12.8 MP) and once at 1 MP: Qwen2.5-VL-7B-Instruct, Qwen3.5-9B (thinking off) and Qwen3-VL-8B-Instruct. The columns right_native_<model> and right_1mp_<model> give the versions that each model answered correctly. ViRA-Train contains no contrast set that Qwen2.5-VL-7B-Instruct or Qwen3.5-9B answers correctly on every version at 1 MP.

Paper

ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relative Policy Optimization

Visual agents answer questions about ultra-high-resolution photos by zooming into the relevant regions. However, existing agents often answer from common sense instead of reading the evidence. Reinforcement learning (RL) on the final answer cannot correct this, because a guess earns the same reward as reading. We propose Cross-Input Relative Policy Optimization (CrossRPO), which trains agents on contrast sets. A contrast set asks one question about several versions of an image that differ only in a small region with the evidence and have different answers, so only reading the evidence solves the whole set. CrossRPO rewards solving the whole set, and we prove that it does not reinforce a set answered from common sense. We build contrast sets for training (ViRA-Train) and evaluation (ViRA-Bench) by editing one small object in images of 21 to 120 megapixels, and we photograph further sets with real changes (ViRA-Bench-Real) to exclude editing traces. We train two Visual Reasoning Agents, ViRA-9B and ViRA-27B, with CrossRPO on only about 3,500 items of ViRA-Train each. CrossRPO achieves 63.3% set accuracy on ViRA-Bench at the maximum resolution, against 59.3% for GRPO. ViRA-9B is the most accurate open model of its size on all benchmarks, and ViRA-27B outperforms Claude Haiku 5.5 and GPT-6-Luna, both at their highest thinking level, on most benchmarks.

License

Contrast sets: O is the Commons photo at source_url; A and B are modified versions of it. Each image keeps the photo's license (license: CC BY, CC BY-SA, CC0 or public domain). Single images: Commons photos and maps keep their license (license, attribution, source_url); math images and problems come from ThinkLite-VL-hard-11k (MIT), which collects them from public datasets; the synthetic charts are made by the authors. Credit attribution, link the source and the license, and say that A and B were modified; share adaptations of CC BY-SA photos under the same license. The questions, the annotations and the synthetic charts are released by the authors under the Apache License 2.0.