graph2table sets a new state of the art on the ChartX benchmark
July 29, 2026
We published a technical report about the accuracy of graph2table. The report measures the system on ChartX. ChartX is the reference benchmark for chart data extraction. This post gives the results and the links to the full material.
The test
The ChartX test set has 1,152 charts of 18 different types. The types include bar charts, line charts, pie charts, heatmaps, radar charts, and candlestick charts. For each chart, the system reads the image and returns the data table. We used the official ChartX metric, SCRM, and the official scorer. We did not change the scorer. Each chart got exactly one attempt. We did not use retries and we did not select between multiple answers.
The results
SCRM reports precision at three tolerance levels. The table below shows the official test set. Higher is better. The baseline numbers come from the published papers.
| System | AP, slight tolerance | AP, strict tolerance |
|---|---|---|
| DePlot (1.3B) | 19.0 | 8.9 |
| ChartAssistant (13B) | 22.8 | 11.4 |
| GPT-4V | 26.0 | 20.9 |
| ChartVLM-L (8.3B) | 30.7 | 23.2 |
| graph2table | 77.7 | 63.6 |
We also measured the system on the ChartX-SE variant. This variant has 2,360 charts of 5 types. Recent specialized models report their results on it. graph2table scores 77.3 AP at slight tolerance on this set. The best published result is 53.5 (ChartVSR).
A problem in the benchmark labels
The report also describes a problem in the ChartX ground truth. Many reference tables contain units that are not in the chart image. For example, a reference header can be "Salary(USD)" when the axis in the image only shows "Salary". A system that reads only the image cannot know the missing unit. The metric then counts a correct extraction as an error. We removed this effect with one neutral transform. We applied the same transform to the reference tables and to our tables. Without the transform, our score is 65.9 instead of 77.7. We released both versions of the data, so everyone can examine the effect.
You can check every number
We released the extracted table for each of the 3,512 evaluated charts. The official ChartX scorer is public. You can score our tables yourself and get each number in the report. You do not need any code from us.
- The report: doi.org/10.5281/zenodo.21669621
- The data: doi.org/10.5281/zenodo.21667548
- GitHub: graph2table/graph2table-chartx-evaluation
- Hugging Face: graph2table/chartx-extractions
Try it
The system in the report is the same system that runs on this site. Upload a chart on the converter page and get your table.