adopd/DocCount

Project | Code | ADOPD2026

DocCount is a manually curated document visual-counting benchmark with 442 original high-resolution images. Every row asks for the number of entities that match one explicit semantic definition, with an integer ground-truth answer.

The first release contains only benchmark inputs and answers. Model responses, reasoning traces, evaluation logs, masks, and semantic tagging artifacts are not part of the dataset.

Benchmark Summary

Questions	Unique images	Target classes	Answer sum	Parquet shards
442	442	4	2,555	2

The four categories are deliberately defined semantically rather than by a single visual template:

Class	Questions	What is counted
photograph	180	photographic image regions meeting the row's class definition
brand_logo	163	visible brand or organization logo instances
table	75	distinct tabular structures
chart_graph	24	distinct charts, plots, or graphs

Loading

from datasets import load_dataset

dataset = load_dataset("adopd/DocCount", split="test")
sample = dataset[0]

image = sample["image"]
prompt = sample["prompt"]
answer = sample["answer"]

Streaming is useful for evaluation runners that do not need a local copy:

dataset = load_dataset("adopd/DocCount", split="test", streaming=True)
for sample in dataset:
    response = model(sample["image"], sample["prompt"])

Evaluation Protocol

Send the original image and the row's prompt to a vision-language model. Parse exactly one final integer and compare it with answer:

correct = int(predicted_count) == sample["answer"]

The primary metric is exact-count accuracy over all 442 rows:

accuracy = number of exactly correct integer answers / 442

Request failures and unparseable responses count as incorrect in the primary metric and should also be reported separately. Do not replace the denominator with only successfully parsed rows. Keep raw responses so parsing and model errors remain distinguishable.

The prompt asks the model to reason carefully and permits a rationale, but the final line must contain one integer in the machine-readable answer tag defined by the row. Models should receive the row prompt verbatim.

Selected Results

Model	Backend	Selected setting	Correct	Exact accuracy
Kimi K2.5	API	reasoning on	322 / 442	72.85%
Qwen3.6 35B A3B	self-hosted	reasoning off	319 / 442	72.17%
Qwen3.5 397B A17B	API	reasoning on	313 / 442	70.81%
GPT-5.5	API	best complete run	304 / 442	68.78%
Claude Sonnet 4.5	API	best complete run	301 / 442	68.10%
GPT-5.2	API	reasoning on	293 / 442	66.29%
GLM-4.6V	self-hosted	reasoning on	285 / 442	64.48%
Gemma 4 31B IT	self-hosted	reasoning on	269 / 442	60.86%

These are exact integer accuracies on the same 442-row test set. The chart uses the best complete result available for each displayed model; reasoning settings are therefore not uniform across rows. Use the released prompt and report model version, inference backend, reasoning mode, sampling parameters, parse failures, and request errors when adding comparisons.

Row Schema

Field	Type	Description
sample_id	string	Stable identifier for the benchmark row
image	Image	Original high-resolution JPEG bytes
image_width	int32	Original width in pixels
image_height	int32	Original height in pixels
image_sha256	string	SHA-256 of the exact embedded image bytes
target_class	string	Human-readable semantic class name
target_class_slug	string	Stable machine-readable class key
class_definition	string	Inclusion/exclusion definition used to determine what counts
question	string	Natural-language counting question
prompt	string	Complete evaluation prompt to send to the model
answer	int64	Manually curated exact count

sample_id and image_sha256 are unique across all 442 rows. There is one question per image in this release.

Data Integrity

• The release contains exactly 442 rows and 442 unique images.
• All answers are integers and sum to 2,555.
• The two Parquet shards contain 229 and 213 rows.
• Embedded images retain their original JPEG bytes.
• Evaluation outputs and model-specific metadata are excluded from the release.

Intended Uses And Limitations

DocCount is intended for evaluating semantic visual counting in document images, including repeated logos, photographs, tables, and charts. It is not a general object-counting training corpus. The category distribution is intentionally non-uniform, and aggregate accuracy should be accompanied by per-class results.

The benchmark is small enough for detailed error analysis but not large enough to characterize every document domain, language, or layout. Models may exploit OCR, visual repetition, and layout cues differently; raw response inspection is recommended.

License

DocCount uses ADOPD source images and is released under CC BY-NC-ND 4.0. It is intended for non-commercial research use. Redistribution must preserve the dataset without adaptations and provide appropriate attribution.

Citation

@misc{zhu2026thinkingwithanchors,
  title={Thinking with Anchors: Grounded and Efficient Document Reasoning},
  author={Sichen Zhu and Yuchen Zhu and Wenzhuo Xu and Jason Kuen and Wanrong Zhu and Jing Shi and Xuan Shen and Quanyi Wang and Yiwei Wang and Yujun Cai and Bing Shuai and Qin Zhang and Yongxin Chen and Shilong Liu and Molei Tao and Jiuxiang Gu},
  year={2026}
}