Amian/FinLongDocQA

---
license: other
task_categories:
- question-answering
language:
- en
tags:
- financial
- numerical-reasoning
- long-document
- table-qa
- multi-table
- annual-reports
pretty_name: FinLongDocQA
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: test
path: dataset_qa.jsonl
---

# FinLongDocQA

**Numerical Reasoning across Multiple Tables for Document-Level Financial Question Answering**

[](https://huggingface.co/datasets/Amian/FinLongDocQA)

## Dataset Description


*An example QA instance from FinLongDocQA. The figure shows only the relevant tables and text for presentation; in practice, the model must retrieve them from the full annual report before computing the answer.*

FinLongDocQA is a benchmark for financial numerical reasoning over long, structured annual reports. It covers both **single-table** and **cross-table** settings where answering a question requires integrating evidence scattered across multiple tables and narrative text.

Financial annual reports commonly exceed 129k tokens, making it challenging for LLMs to (1) locate the relevant tables (*context rot*) and (2) perform accurate multi-step arithmetic once the evidence is found. FinLongDocQA is designed to stress-test both capabilities.

### Dataset Summary

| Field | Value |
|---|---|
| Examples | 7,527 |
| Companies | 489 |
| Fiscal years | 2022, 2023, 2024 |
| Question types | `mixed` (5,951), `table` (1,319), `text` (257) |

### Question Types

| Type | Description |
|---|---|
| `table` | Evidence comes entirely from one or more financial tables |
| `text` | Evidence comes entirely from narrative text |
| `mixed` | Evidence spans both tables and narrative text |

## Dataset Structure

Each record in `dataset_qa.jsonl` contains:

```json
{
"id": "1",
"company": "A",
"year": "2022",
"question": "On average, how many manufacturing facilities does each business segment have?",
"type": "mixed",
"thoughts": "Thought: Page 4 cites 3 segments. Page 11 lists 4 U.S. and 4 non-U.S. manufacturing facilities = 8 total. Average = 8/3.",
"page_numbers": [4, 11],
"python_code": "total_facilities=8\nsegments=3\navg=total_facilities/segments\nround(avg,2)",
"answer": 2.67
}
```

### Fields

| Field | Type | Description |
|---|---|---|
| `id` | string | Unique example identifier |
| `company` | string | Anonymized company ticker |
| `year` | string | Fiscal year of the annual report |
| `question` | string | Natural-language financial question |
| `type` | string | Question type: `table`, `text`, or `mixed` |
| `thoughts` | string | Chain-of-thought reasoning trace with page references |
| `page_numbers` | list[int] | Pages in the annual report that contain the relevant evidence |
| `python_code` | string | Executable Python snippet that computes the answer |
| `answer` | float | Ground-truth numerical answer |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Amian/FinLongDocQA")
print(ds["test"][0])
```

## License

This dataset is released under the **AI²Lab Source Code License (National Taiwan University)**.
See the full license [here](LICENSE).

---
许可证：其他
任务类别：
- 问答（question-answering）
语言：
- 英语（en）
标签：
- 金融（financial）
- 数值推理（numerical-reasoning）
- 长文档（long-document）
- 表格问答（table-qa）
- 多表格（multi-table）
- 年度报告（annual-reports）
美观名称：FinLongDocQA
规模类别：
- 1K<n<10K
配置项：
- 配置名称：default
数据文件：
- 拆分集：test
路径：dataset_qa.jsonl
---

# FinLongDocQA

**面向多表格的文档级金融问答数值推理基准**

[](https://huggingface.co/datasets/Amian/FinLongDocQA)

## 数据集说明


*FinLongDocQA中的一则问答示例。本图仅展示相关表格与文本以作演示；实际应用中，模型需先从完整年度报告中检索到对应内容，再计算最终答案。*

FinLongDocQA是一款面向长格式结构化年度报告的金融数值推理基准数据集。其涵盖**单表格**与**跨表格**两类场景，在这类场景中，回答问题需要整合分散在多个表格及叙述性文本中的证据。

金融年度报告的Token数通常超过129000，这给大语言模型（Large Language Model）带来了两大挑战：(1) 定位相关表格（即*上下文迷失（context rot）*）；(2) 在找到证据后执行准确的多步算术运算。FinLongDocQA正是为了对这两项能力进行压力测试而设计的。

### 数据集概览

| 字段 | 数值 |
|---|---|
| 示例总数 | 7,527 |
| 覆盖企业 | 489家 |
| 财年范围 | 2022、2023、2024 |
| 问题类型 | 混合型（5,951条）、表格型（1,319条）、文本型（257条） |

### 问题类型

| 类型 | 说明 |
|---|---|
| `table` | 答案证据完全来自一个或多个金融表格 |
| `text` | 答案证据完全来自叙述性文本 |
| `mixed` | 答案证据同时涵盖表格与叙述性文本 |

## 数据集结构

`dataset_qa.jsonl`中的每条记录均包含以下内容：

json
{
"id": "1",
"company": "A",
"year": "2022",
"question": "各业务板块平均拥有多少家生产工厂？",
"type": "mixed",
"thoughts": "思考过程：第4页提及3个业务板块。第11页列出了4家美国本土及4家海外生产工厂，总计8家。平均数量=8/3。",
"page_numbers": [4, 11],
"python_code": "total_facilities=8
segments=3
avg=total_facilities/segments
round(avg,2)",
"answer": 2.67
}

### 字段说明

| 字段名 | 数据类型 | 说明 |
|---|---|---|
| `id` | 字符串 | 唯一的示例标识符 |
| `company` | 字符串 | 匿名化处理后的企业股票代码 |
| `year` | 字符串 | 对应年度报告的财年 |
| `question` | 字符串 | 自然语言形式的金融问题 |
| `type` | 字符串 | 问题类型：`table`、`text`或`mixed` |
| `thoughts` | 字符串 | 带页码引用的链式思考推理轨迹 |
| `page_numbers` | 整数列表 | 年度报告中包含相关证据的页码 |
| `python_code` | 字符串 | 可执行的Python代码片段，用于计算最终答案 |
| `answer` | 浮点数 | 标注的真实数值答案 |

## 使用方法

python
from datasets import load_dataset

ds = load_dataset("Amian/FinLongDocQA")
print(ds["test"][0])

## 许可证

本数据集采用**AI²Lab源代码许可协议（国立台湾大学）**发布。完整许可证内容请参见[LICENSE](LICENSE).