time-wizard-bench

time-wizard-bench holds 3196 photographs of analog clocks. We cropped each photograph to one clock. Each crop carries the hour and minute that clock shows. The photographs come from COCO and OpenImages. The labels come from the paper "It's About Time: Analog Clock Reading in the Wild" by Charig Yang, Weidi Xie, and Andrew Zisserman. The dataset adds fixed test, dev, and train splits so that anyone can compare two models on identical photographs.

The code that built the dataset, trains on it, and scores against it lives at github.com/jadidbourbaki/time-wizard.

Examples

Three crops from the test split with their labels.

3:28	1:20	1:00

The first two crops are typical of the sharper photographs. The third shows the low resolution of many COCO clocks.

Fields

Field	Type	Meaning
key	string	Source and image id, such as coco_000000009172
image	image	The clock crop, 448 by 448 pixels, RGB
hours	int	Hour on the 12 hour dial, 1 to 12
minutes	int	Minute, 0 to 59

A key starting with coco_ names an image in COCO 2017. A key starting with openimages_ names an image in OpenImages. The part after the underscore is the image id in the source dataset. Every crop therefore traces back to its original photograph.

Splits

Split	Photographs	Purpose
test	200	Score a finished model once
dev	200	Choose hyperparameters and select checkpoints
train	2796	Fine-tune

Each split holds COCO and OpenImages photographs in the same proportion as the whole set, 59 percent COCO. A random draw with seed 0 assigned the photographs. The split files in the GitHub repository record the assignment as image ids and labels.

Crop pipeline

The It's About Time authors found 1911 clocks in COCO photographs and 1317 in OpenImages photographs. The authors read each clock by eye and recorded the hour and minute. The object detector CBNetV2 drew a box around each clock. The authors published the labels and the boxes as CSV files under the MIT licence.

The build script in the GitHub repository turns those files into this dataset in five steps:

1. Download each photograph from COCO or OpenImages.
2. Crop the photograph to the detector box plus a margin of 20 percent of the box on each side. The paper crops with the same margin.
3. Pad the crop to a square with black bars and resize it to 448 pixels.
4. Drop near duplicates. A perceptual hash summarises each crop as a 64 bit fingerprint. Two crops whose fingerprints differ in four bits or fewer show the same clock. The hash check removed 32 photographs.
5. Shuffle each source with seed 0. Assign 200 photographs to test, 200 to dev, and the remaining 2796 to train.

Scoring

Show the model one crop and ask for the time. We fix the system prompt and the question:

Reply with ONLY the requested JSON, no preface and no code block.

What time does this analog clock show? Reply as JSON: {"hours": H, "minutes": M} with H from 1 to 12.

Take the first JSON object in the reply as the prediction. A reply with no JSON object counts as wrong. A prediction counts as correct when it lands within one minute of the label on the 12 hour dial. The distance wraps at twelve o'clock. A prediction of 12:59 against a label of 1:00 is therefore one minute of error.

The timewizard.reading module in the GitHub repository holds the two strings, the parser, and the metric. The snippet below scores any function that maps a crop and the two prompts to a reply string:

from datasets import load_dataset
from timewizard.reading import PROMPT, SYSTEM, Time, parse_time, score

rows = load_dataset("jadidbourbaki/time-wizard-bench", split="test")
predictions = [parse_time(my_model(row["image"], SYSTEM, PROMPT)) for row in rows]
truths = [Time(hours=row["hours"], minutes=row["minutes"]) for row in rows]
print(score(predictions, truths))

parse_time returns the parsed time or None. score returns the share within one minute, the share exactly right, the share with the correct hour, the mean circular error in minutes over the readable replies, and the count of unreadable replies. The command just bench --model <id> --split test in the repository runs the same procedure against any model that pydantic-ai can reach.

The It's About Time paper computes its top-1 accuracy with the same one minute rule. The paper's numbers on the full COCO and OpenImages sets sit on the same scale as numbers on this test split.

Results

GPT-5.6 Sol, Claude Fable 5.1, and Claude Opus 5 answered the 200 test crops through their APIs. Opus 5 ran on 2026-09-03. GPT-5.6 Sol and Fable 5.1 ran on 2026-09-04.

Model	Effort	Within 1 min	Exact	Hour	Mean error	Unreadable
GPT-5.6 Sol	max	61.5%	31.0%	68.5%	65 min	0
Claude Fable 5.1	max	21.5%	11.5%	42.5%	107 min	2
Claude Opus 5	max	15.5%	6.0%	31.0%	141 min	0

Each column reads as follows.

• Within 1 min, the headline metric, is the share of clocks where the prediction lands within one minute of the label on the 12 hour dial. The distance wraps at twelve. A reading of 12:59 against 1:00 is one minute off.
• Exact is the share of clocks where the prediction matches the label to the minute.
• Hour is the share of clocks where the predicted hour matches the labelled hour, whatever the minute.
• Mean error is the wrapped distance in minutes between prediction and label, averaged over the readable replies. The largest possible distance is 360 minutes.
• Unreadable is the count of replies with no valid time in them, such as a refusal or malformed JSON. An unreadable reply counts as wrong in the three shares above and is left out of the mean error.

Every model saw the same crops, sent as 448 pixel PNG images, with the system prompt and question above. Each model ran at its highest reasoning setting. The reply budget was 32000 tokens per clock, covering reasoning and answer. We allowed each request 600 seconds and retried a failed request up to five times. Eight requests ran at once.

Model	API id	Provider	Input tokens	Output tokens	Cost
GPT-5.6 Sol	openai.gpt-5.6-sol	Amazon Bedrock, us-east-1	79,200	686,146	$14.04
Claude Fable 5.1	claude-fable-5-1	Anthropic API	65,934	971,794	$49.25
Claude Opus 5	claude-opus-5	Anthropic API	66,200	365,711	$9.47

Cost multiplies the token counts by the list price on the day of the run: $4 and $20 per million input and output tokens for GPT-5.6 Sol on Bedrock, $10 and $50 for Fable 5.1, and $5 and $25 for Opus 5. Fable 5.1 used its whole budget on two clocks without answering. The runner records no tokens for a failed request. The two failed requests add about 64,000 output tokens and $3.20 that the table leaves out.

The GitHub repository holds every reply in runs/bench/ and the score files beside them.

Limitations

COCO clocks are small in their source photographs. The median COCO clock spans 80 pixels. The median OpenImages clock spans 235 pixels. Four in five COCO crops cover under 224 pixels of the source photograph before the resize to 448. At that size one minute of arc is under a pixel wide. No model can read those clocks to the minute.

The labels record hours and minutes only. There is no second hand label, no AM or PM, and no flag for a clock that shows an impossible time. The authors read the clocks by eye. Some labels therefore carry a minute or two of error. We score every model against the same labels. Comparisons between models therefore hold.

The photographs have been public on the internet for years. A model trained on web data may have seen them along with nearby text. A model trained on the train split alone has seen only those 2796 crops.

Licence

The labels, the crop boxes, and the split assignments are MIT. The It's About Time repository publishes its labels under MIT. time-wizard-bench keeps that licence for everything derived from them.

The photographs keep the licence of their source. OpenImages photographs are CC BY 2.0. The key field identifies each one for attribution. COCO photographs come from Flickr under the licence each photographer chose. COCO distributes them subject to the Flickr terms of use. Use the photographs on the same terms as their sources.

Citation

Cite the paper that produced the labels:

@inproceedings{yang2022itsabouttime,
  title     = {It's About Time: Analog Clock Reading in the Wild},
  author    = {Yang, Charig and Xie, Weidi and Zisserman, Andrew},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2022}
}

Cite the dataset for the splits and the scores:

@misc{timewizardbench2026,
  title  = {time-wizard-bench},
  author = {Tirmazi, Hayder},
  year   = {2026},
  url    = {https://huggingface.co/datasets/jadidbourbaki/time-wizard-bench}
}