--- license: mit dataset_info: - config_name: de features: - name: item_id dtype: string - name: id_hash dtype: string - name: pre_text dtype: string - name: post_text dtype: string - name: table list: list: string - name: images list: image - name: images_roles list: string - name: question dtype: string - name: answer dtype: string splits: - name: train num_bytes: 1232550828 num_examples: 1239 download_size: 1187742792 dataset_size: 1232550828 - config_name: en features: - name: item_id dtype: string - name: id_hash dtype: string - name: pre_text dtype: string - name: post_text dtype: string - name: table list: list: string - name: images list: image - name: images_roles list: string - name: question dtype: string - name: answer dtype: string splits: - name: train num_bytes: 2638340272 num_examples: 3313 download_size: 2614792468 dataset_size: 2638340272 - config_name: es features: - name: item_id dtype: string - name: id_hash dtype: string - name: pre_text dtype: string - name: post_text dtype: string - name: table list: list: string - name: images list: image - name: images_roles list: string - name: question dtype: string - name: answer dtype: string splits: - name: train num_bytes: 1219399524 num_examples: 1238 download_size: 1114127869 dataset_size: 1219399524 - config_name: fr features: - name: item_id dtype: string - name: id_hash dtype: string - name: pre_text dtype: string - name: post_text dtype: string - name: table list: list: string - name: images list: image - name: images_roles list: string - name: question dtype: string - name: answer dtype: string splits: - name: train num_bytes: 1105747447 num_examples: 1239 download_size: 1090202705 dataset_size: 1105747447 - config_name: it features: - name: item_id dtype: string - name: id_hash dtype: string - name: pre_text dtype: string - name: post_text dtype: string - name: table list: list: string - name: images list: image - name: images_roles list: string - name: question dtype: string - name: answer dtype: string splits: - name: train num_bytes: 1087071034 num_examples: 1238 download_size: 1055442857 dataset_size: 1087071034 configs: - config_name: de data_files: - split: train path: de/train-* - config_name: en data_files: - split: train path: en/train-* - config_name: es data_files: - split: train path: es/train-* - config_name: fr data_files: - split: train path: fr/train-* - config_name: it data_files: - split: train path: it/train-* --- # Multi-FinQA **Multi-FinQA** is a **multilingual** extension of the original [**FinQA**](https://finqasite.github.io) dataset. FinQA is a large-scale expert-annotated dataset with 2.8k financial reports for 8k Question-Answering pairs over financial reports to study numerical reasoning with structured and unstructured evidence. The dataset was introduced in the 2021 EMNLP paper [FinQA: A Dataset of Numerical Reasoning over Financial Data](https://aclanthology.org/2021.emnlp-main.300/), where details about its implementation can be found. The FinQA dataset version used in **Multi-FinQA** is the one available at [The Cauldron](https://huggingface.co/datasets/HuggingFaceM4/the_cauldron), a massive collection of 50 vision-language datasets (training sets only) that were used for the fine-tuning of the vision-language model Idefics2. ### Dataset Description **Multi-FinQA** extends the support to four languages, i.e. French, German, Italian, and Spanish, using a **permissively licensed** model: [**RedHatAI gemma-4-31B-it-FP8-dynamic**](https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic). Each financial report consists of a financial table preceded and followed by text, as shown in the following dataset entry excerpt: ```json "pre_text": the texts before the table; "post_text": the text after the table; "table": the table; ... "qa": { "question": the question; "answer": the answer; ... } ``` `"pre_text"`, `"post_text"`, and `"table"` represent the fields of the financial report itself. For each entry in the dataset, the model was prompted to translate the above-mentioned JSON fields to the desired target language. Once having the financial reports translated, they were rendered back as .png images. Images were built as plain HTML/CSS and rendered using randomized styling (fonts, colors, spacing) for visual diversity. According to the length of the `"pre_text"` and `"post_text"` fields, the images were split into separate images or combined into a single one. Information about the splitting policy per each dataset entry is to be found in the .parquet `image_roles` column: `combined` refers to a single .png image, otherwise a list shows the order in which the images are shown in column (e.g. `"table"`, `"pre_text"`, `"post_text"`). ## Dataset Structure Each entry in every language subset is structured as follows: | Field | Type | Description | | ----- | ---- | ----------- | | `item_id` | string | unique example id. composed by the original report name plus example index for this report | | `id_hash` | string | hash of the dataset entry | | `pre_text` | string | the text(s) before the table | | `post_text` | string | the text(s) after the table | | `table` | list | the financial table | | `images` | list | the image(s) of the financial table | | `image_roles` | list | the role of the image: `combined` if a single image, otherwise the ordered image role is shown | | `question` | string | the question | | `answer` | string | the answer | ## Language distribution The dataset is composed of five subsets, one per language, distributed as follows: | Language | Subset Name | Percentage | Number of Examples | | -------- | ----------- | ---------- | ------------------ | | English | en | 40% | 3313 | | French | fr | 15% | 1239 | | German | de | 15% | 1239 | | Italian | it | 15% | 1238 | | Spanish | es | 15% | 1238 | English constitutes 40% of the dataset. French, German, Italian, and Spanish collectively account for the remaining 60%, distributed evenly. ## Loading the dataset To load a specific subset: ```python from datasets import load_dataset dataset = load_dataset("VillanovaAI/Multi-FinQA", "fr") ``` ### Dataset Sources - **Repository:** [The Cauldron](https://huggingface.co/datasets/HuggingFaceM4/the_cauldron) - **Paper:** [https://aclanthology.org/2021.emnlp-main.300.pdf](https://aclanthology.org/2021.emnlp-main.300.pdf) ``` @article{chen2021finqa, title={FinQA: A Dataset of Numerical Reasoning over Financial Data}, author={Chen, Zhiyu and Chen, Wenhu and Smiley, Charese and Shah, Sameena and Borova, Iana and Langdon, Dylan and Moussa, Reema and Beane, Matt and Huang, Ting-Hao and Routledge, Bryan and Wang, William Yang}, journal={Proceedings of EMNLP 2021}, year={2021} } ``` ## Licensing Information The FinQA dataset is distributed under a **dual-layered licensing structure** based on its components: 1. **Dataset Annotations & Code:** The annotations created by the authors and the evaluation code are licensed under the **MIT License** (as per the [official FinQA GitHub repository](https://github.com)). 2. **Underlying Financial Data:** The raw financial reports are derived from IBM's *FinTabNet* dataset, which is distributed under the **CDLA-Permissive-1.0** (Community Data License Agreement) license. This combination allows for free usage, modification, and redistribution, including commercial applications, provided that the respective license notices and copyright attributions are preserved.