--- dataset_info: features: - name: question_id dtype: string - name: qtype dtype: string - name: figure_path dtype: image - name: visual_figure_path list: image - name: question dtype: string - name: answer dtype: string - name: instructions dtype: string - name: prompt dtype: string - name: options list: string splits: - name: info num_bytes: 9389294399.0 num_examples: 55091 - name: plain num_bytes: 15950918129.0 num_examples: 55091 - name: visual_metaphor num_bytes: 144053150.0 num_examples: 450 - name: visual_basic num_bytes: 1254942699.466 num_examples: 7297 download_size: 20376840742 dataset_size: 26739208377.466 configs: - config_name: default data_files: - split: info path: data/info-* - split: plain path: data/plain-* - split: visual_metaphor path: data/visual_metaphor-* - split: visual_basic path: data/visual_basic-* --- # InfoChartQA: Benchmark for Multimodal Question Answering on Infographic Charts ![xbhs3](teaser.jpg)🤗[Dataset](https://huggingface.co/datasets/Jietson/InfoChartQA) # Dataset You can find our dataset on huggingface: 🤗[InfoChartQA Dataset](https://huggingface.co/datasets/Jietson/InfoChartQA) # Usage Each question entry is arranged as: ``` --question_id: int --qtype: int --figure_path: image --visual_figure_path: list of image --question: str --answer: str --instructions: str --prompt: str --options: list of dict ("A/B/C/D":"option_content") ``` Each question is built as: ``` image_input: figure_path, visual_figure_path_1...visual_figure_path_n (if any) text_input: prompt (if any) + question + options (if any) + instructions (if any) ``` # Evaluate You should store and evaluate model's response as: ```python # Example code for evaluate def build_question(query):#to build the question question = "" if "prompt" in query: question = question + f"{query["prompt"]}\n" question = question + f"{query["question"]}\n" if "options" in query: for _ in query["options"]: question = question + f"{_} {query['options'][_]}\n" if "instructions" in query: question = question + query["instructions"] return question with open("visual_basic.json","r",encode="utf-8") as f: queries = json.load(f) for idx in range(queries): question = build_question(queries[idx]) figure_path = [queries[idx]['figure_path']] visual_figure_path = queries[idx]['visual_figure_path'] response = model.generate(question, [figure_path, visual_figure_path])# generate model's response based on queries[idx]["response"] = reponse with open("model_reponse.json","w",encode="utf-8") as f: json.dump(queries, f) from checker import evaluate evaluate("model_reponse.json", "path_to_save_the_result") ``` Or simply use after your answer is generated: ```python python -c "import checker; checker.evaluate(sys.argv[1], sys.argv[2])" PATH_TO_INPUT_FILE PATH_TO_INPUT_FILE ``` # LeaderBoard | Model | Infographic | Plain | Δ | Basic | Metaphor | Avg. | | ------------------------ | ----------- | ------- | ----- | ------- | -------- | ----- | | **Baselines** | | | | | | | | Human | 95.35\* | 96.28\* | 0.93 | 93.17\* | 88.69 | 90.93 | | **Proprietary Models** | | | | | | | | OpenAI O4-mini | 79.41 | 94.61 | 15.20 | 92.12 | 54.76 | 73.44 | | GPT-4o | 66.09 | 81.77 | 15.68 | 81.77 | 47.19 | 64.48 | | Claude 3.5 Sonnet | 65.67 | 83.11 | 17.44 | 90.36 | 55.33 | 72.85 | | Gemini 2.5 Pro Preview | 83.31 | 93.88 | 10.07 | 90.01 | 60.42 | 75.22 | | Gemini 2.5 Flash Preview | 71.91 | 84.66 | 12.75 | 82.02 | 56.28 | 69.15 | | **Open-Source Models** | | | | | | | | Qwen2.5-VL-72B | 62.06 | 78.47 | 16.41 | 77.34 | 54.64 | 65.99 | | Llama-4 Scout | 67.41 | 84.84 | 17.43 | 81.76 | 51.89 | 66.83 | | Intern-VL3-78B | 66.38 | 82.18 | 15.80 | 79.46 | 51.52 | 65.49 | | Intern-VL3-8B | 56.82 | 73.50 | 16.68 | 74.26 | 49.57 | 61.92 | | Janus Pro | 29.61 | 45.29 | 15.68 | 41.18 | 42.21 | 41.69 | | DeepSeek VL2 | 39.81 | 47.01 | 7.20 | 58.72 | 44.54 | 51.63 | | Phi-4 | 46.20 | 66.97 | 20.77 | 61.87 | 38.31 | 50.09 | | LLaVA OneVision Chat 78B | 47.78 | 63.66 | 15.88 | 62.11 | 50.22 | 56.17 | | LLaVA OneVision Chat 7B | 38.41 | 54.43 | 16.02 | 61.03 | 45.67 | 53.35 | | Pixtral | 44.70 | 60.88 | 16.11 | 64.23 | 50.87 | 57.55 | | Ovis1.6-Gemma2-9B | 50.56 | 64.52 | 13.98 | 60.96 | 34.42 | 47.69 | | ChartGemma | 19.99 | 33.81 | 13.82 | 30.52 | 33.77 | 32.15 | | TinyChart | 26.34 | 44.73 | 18.39 | 14.72 | 9.03 | 11.88 | | ChartInstruct-LLama2 | 20.55 | 27.91 | 7.36 | 33.86 | 33.12 | 33.49 | # License Our original data contributions (all data except the charts) are distributed under the [CC BY-SA 4.0](https://github.com/princeton-nlp/CharXiv/blob/main/data/LICENSE) license. Our code is licensed under [Apache 2.0](https://github.com/princeton-nlp/CharXiv/blob/main/LICENSE) license. The copyright of the charts belong to the original authors. ## Cite If you use our work and are inspired by our work, please consider cite us (available soon): ``` @misc{lin2025infochartqabenchmarkmultimodalquestion, title={InfoChartQA: A Benchmark for Multimodal Question Answering on Infographic Charts}, author={Minzhi Lin and Tianchi Xie and Mengchen Liu and Yilin Ye and Changjian Chen and Shixia Liu}, year={2025}, eprint={2505.19028}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2505.19028}, } ```