Awesome

(English|简体中文)

Benchmarking Large Language Models on CFLUE - A Chinese Financial Language Understanding Evaluation Dataset

CFLUE Version 1.0 — Chinese Financial Language Understanding Evaluation Dataset in the Financial Domain

Alibaba Cloud, in collaboration with Soochow University, has introduced CFLUE (Chinese Financial Language Understanding Evaluation), a novel and comprehensive benchmark designed to assess the understanding and processing capabilities of large language models within the context of Chinese financial language.

CFLUE evaluates the performance of language models through two main dimensions—Knowledge Assessment and Application Assessment.

• The Knowledge Assessment component consists of over 38,000 multiple-choice questions selected from 15 different types of financial qualification simulation exams, aimed at testing the language models' ability to predict answers and reason. Each question is accompanied by explanations, which aids in a thorough evaluation of the models' reasoning processes.

• The Application Assessment component provides over 16,000 instances covering five classic NLP tasks including text classification, machine translation, relation extraction, reading comprehension, and text generation. These instances are derived from existing shared tasks or annotated real data by professionals.

Overall, CFLUE offers multi-faceted insights for understanding and enhancing the capabilities of LLMs in the Chinese financial domain, and calls for more comprehensive and meticulous assessment of these models through CFLUE. The research team hopes that CFLUE will not only facilitate in-depth understanding of existing models but also drive new strides in the development of language models in the Chinese financial domain.

Currently, the CFLUE V1.0 evaluation dataset is available to the public, with plans to continuously update versions and introduce an integrated platform-based evaluation service in the future. This aims to provide a comprehensive, one-stop evaluation solution for the entire industry.

Changelog

• [2024.05.16] Our paper "Benchmarking Large Language Models on CFLUE - A Chinese Financial Language Understanding Evaluation Dataset" has been officially accepted by ACL-2024! 🚀🚀🚀
• [2024.03.06] CFLUE has now open-sourced the development set for "Knowledge Assessment" and data samples for "Application Assessment" 🚀🚀🚀；

Table of Contents

• Leaderboard
• Data
• Quick Start
• How to Submit
• TODO
• Licenses
• Citations

Leaderboard

Below, we list the zero-shot performance of the models we evaluated in the initial version. We have conducted a comprehensive assessment of various large language models on CFLUE, including OpenAI's GPT-4 and GPT-4-turbo, as well as several models for general and financial domains. The results show that GPT-4 and GPT-4-turbo significantly outperform other models in answer prediction for Knowledge Assessment, with accuracy exceeding 60%, demonstrating their leading position in the field but also implying significant room for improvement in other language models. In Application Assessment, although these two models perform well overall, their advantage is somewhat reduced compared to some models specifically designed for Chinese. The study also found that current LLMs in the financial field, such as FinGPT V3, DISC-FinLLM, and Tongyi-Finance, perform poorly in zero-sample tests, indicating that these models’ coverage of financial knowledge needs to be strengthened. On the other hand, lightweight LLMs perform well after supervised fine-tuning, such as ChatGLM3-6B, Qwen-7B, and Baichuan2-7B, whose performance in some tasks is comparable to the larger-parameter ChatGPT.

Knowledge Assessment

Application Assessment

Data

The CFLUE evaluation data in the data directory contains two subdirectories: knowledge and application. The former includes evaluation data for financial applications, while the latter contains sample evaluation data for financial applications.

Financial Knowledge Evaluation Data

Here is a sample data entry for financial knowledge:

{
    "名称":"证券专项考试",
    "task":"单项选择题",
    "question":"用复利计算第n期终值的公式为（）。",
    "choices":"{'A': 'FV=PV×(1+I×n)', 'B': 'PV=FV×(1+I×n)', 'C': 'FV=PV×(1+I)^n', 'D': 'PV=FV×(1+I)^n'}",
    "answer":"C",
    "analysis":"复利是计算利息的另一种方法。按照这种方法，每经过一个计息期，要将所生利息加入本金再计利息。因此，复利终值计算公式为：FV=PV×(1+I)^n。"
}

Because financial knowledge includes three different types of question formats—single-choice, multiple-choice, and true/false—CFLUE utilizes corresponding prompt templates. By executing utils/format_example.py, one can quickly load the data and build the final model input. Below are examples of prompt templates for single-choice and multiple-choice questions:

• Single-Choice Question Prompt Template

假设你是一位金融行业专家，请回答下列问题。
注意：题目是单选题，只需要返回一个最合适的选项，若有多个合适的答案，只返回最准确的即可。
注意：结果只输出两行，第一行只需要返回答案的英文选项(注意只需要返回一个最合适的答案)，第二行进行简要的解析，输出格式限制为：“答案：”，“解析：”。

{question}
{choices}

• Multiple-Choice Question Prompt Template

假设你是一位金融行业专家，请回答下列问题。
注意：题目是多选题，可能存在多个正确的答案。
注意：结果只输出两行，第一行只需要返回答案的英文选项，第二行进行简要的解释。输出格式限制为：“答案：”，“解析：”。

{question}
{choices}

Financial Application Evaluation Data

For example, in financial text classification, the data samples are as follows:

{
    "task":"金融文本分类",
    "sub_task":"ESG分类",
    "instruction":"你是一个金融行业的文本分类专家，请将下面的ESG公告进行分类，类别包括：['顾客健康与安全', '物料', '反腐败行为', '非虚假营销', '多样性与机会平等', '依法合规纳税', '非直接经济影响', '公共政策', '水资源与污水处理', '职业培训及教育', '经济绩效', '反不正当竞争', '安全管理实践', '能源', '市场占有率', '排放']。请仅输出分类结果，不要做多余的解释，答案有且只有一个。\n公告：本公司及董事会全体成员保证信息披露的内容真实、准确、完整，没有虚假记载、误导性陈述或重大遗漏。 2010年3月16日，吉林光华控股集团股份有限公司（以下简称\"公司\"）接到公司第一大股东江苏开元资产管理有限公司（以下简称\"开元资产\"）转发的江苏省人民政府国有资产监督管理委员会《关于同意公开转让江苏开元资产管理公司9%股权的批复》（苏国资复[2010]27号）。批复内容如下： 同意江苏开元国际集团有限公司（以下简称\"开元集团\"）将所持开元资产9%的股权，通过产权交易机构进行公开转让。根据国家关于企业国有产权转让的有关规定，办理资产评估相关手续，合理制定产权转让公告中受让方应具备的合理性、公平性资格条件，确保产权转让的公开与规范。并按照《国有股东转让所持上市公司股份管理暂行办法》（国资委、证监会令第19号）要求，及时履行信息披露等相关义务，在办理产权转让鉴证前，按规定程序报国务院国资委审核批准。 开元集团预计将在近日于江苏省产权交易所进行公开挂牌转让开元资产9%股权。开元集团持有开元资产51%股权，上海泰泓投资管理有限公司持有开元资产49%股权，若此次公开转让成功，开元集团持有开元资产的股权将下降至42%，将导致开元资产控股权发生变化，进而导致公司实际控制人发生变化。公司将根据该事项进展情况履行持续信息披露义务。 特此公告。 吉林光华控股集团股份有限公司 董事会 2010年3月16日\n分类结果：",
    "input":"",
    "output":"非直接经济影响",
    "history":[]
}

Unlike financial knowledge, the 'instruction' field in financial application evaluation data refers to the model input field that has been assembled using the corresponding prompt, allowing for the immediate commencement of the evaluation task.

Quick Start

#!/bin/bash

model_name=""  # 模型名称
checkpoint_path=""  # 指定模型checkpoint路径
eval_type="knowledge"  # 指定金融知识评测或金融应用评测
save_result_dir="../results"

python cflue_main.py \
    --model_name ${model_name} \
    --checkpoint_path ${checkpoint_path} \
    --eval_type ${eval_type} \
    --save_result_dir ${save_result_dir}

How to Submit

If you wish to participate in the evaluation of the custom large model on the Test set, you will first need to prepare a UTF-8 encoded JSON file and write it in the format of submission_example.json. Then, send it as an attachment in an email formatted as below to CFLUE@alibabacloud.com to apply.

Organization:
Contact Information:
Model Name:
Model Results Attachment 📎

TODO

• Release the development set for the "Knowledge Assessment" section and data samples for the "Application Assessment";
• Include zero-shot results;
• Provide detailed results for each subtask;
• Release the full test dataset (without labels);
• Release the training data for "Knowledge Assessment";

Licenses

This project adheres to MIT License.

The CFLUE dataset follows Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Citations

If you use our dataset, please cite our paper.

@inproceedings{zhu2024cflue,
title={Benchmarking Large Language Models on CFLUE - A Chinese Financial Language Understanding Evaluation Dataset}, 
author={Jie Zhu, Junhui Li, Yalong Wen, Lifan Guo},
booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics(ACL-2024)},
year={2024}
}