|
|
--- |
|
|
license: cc-by-nc-4.0 |
|
|
--- |
|
|
# Dataset Summary |
|
|
This dataset contains a large-scale collection of tables from the financial earnings reports of Fortune 500 companies. It is derived from the ComTQA dataset, which in turn is based on the original FinTabNet dataset released by IBM Research. |
|
|
|
|
|
This version has been modified to make each question specific to a single document. This change transforms the dataset into a benchmark for question answering over a large corpus of different PDFs, designed to evaluate a system's ability to first retrieve the correct document and then extract and reason over its contents. |
|
|
|
|
|
# Supported Tasks |
|
|
- **Question Answering over a Document Corpus:** The primary task is to find the correct financial report from the entire corpus and then answer a specific question using the tables and text within that document. For example: |
|
|
``` |
|
|
For PNR, What is the % point change of segment income from 2014 to 2015? |
|
|
``` |
|
|
|
|
|
# Dataset Structure |
|
|
- **Corpus:** The dataset consists of a large corpus of PDF documents, each representing a distinct financial report. |
|
|
- **Questions:** The question-answer pairs have been curated to be answerable from a single, specific PDF within the corpus and are provided in the repository's data files. |
|
|
|
|
|
# Licensing and Attribution |
|
|
This dataset is a subset of the [ComTQA dataset](https://huggingface.co/datasets/ByteDance/ComTQA) and its use is subject to the original license, which requires attribution. The underlying data comes from the original [FinTabNet dataset](https://dataplatform.cloud.ibm.com/analytics/notebooks/v2/f57cf3f6-e972-48ff-ab7b-3771ba7b9683/view?access_token=317644327d84f5d75b4782f97499146c78d029651a7c7ace050f4a7656033c30) from IBM Research. |
|
|
|
