GitHub - The-FinAI/PIXIU: This repository introduces PIXIU, an open-source resource featuring the first financial large language models (LLMs), instruction tuning data, and evaluation benchmarks to holistically assess financial LLMs. Our goal is to continually push forward the open-source development of financial artificial intelligence (AI).

Qianqian Xie¹ Weiguang Han² Zhengyu Chen² Ruoyu Xiang¹ Xiao Zhang¹ Yueru He¹ Mengxi Xiao² Dong Li² Yongfu Dai⁷ Duanyu Feng⁷ Yijing Xu¹ Haoqiang Kang⁵ Ziyan Kuang¹² Chenhan Yuan³ Kailai Yang³ Zheheng Luo³ Tianlin Zhang³ Zhiwei Liu³ Guojun Xiong¹⁰ Zhiyang Deng⁹ Yuechen Jiang⁹ Zhiyuan Yao⁹ Haohang Li⁹ Yangyang Yu⁹ Gang Hu⁸ Jiajia Huang¹¹ Xiao-Yang Liu⁵ Alejandro Lopez-Lira⁴ Benyou Wang⁶ Yanzhao Lai¹³ Hao Wang⁷ Min Peng^2* Sophia Ananiadou³ Jimin Huang¹

¹The Fin AI ²Wuhan University ³The University of Manchester ⁴University of Florida ⁵Columbia University ⁶The Chinese University of Hong Kong, Shenzhen ⁷Sichuan University ⁸Yunnan University ⁹Stevens Institute of Technology ¹⁰Stony Brook University ¹¹Nanjin Audit University ¹²Jiangxi Normal University ¹³Southwest Jiaotong University

Pixiu Paper | FinBen Leaderboard

Disclaimer

This repository and its contents are provided for academic and educational purposes only. None of the material constitutes financial, legal, or investment advice. No warranties, express or implied, are offered regarding the accuracy, completeness, or utility of the content. The authors and contributors are not responsible for any errors, omissions, or any consequences arising from the use of the information herein. Users should exercise their own judgment and consult professionals before making any financial, legal, or investment decisions. The use of the software and information contained in this repository is entirely at the user's own risk.

By using or accessing the information in this repository, you agree to indemnify, defend, and hold harmless the authors, contributors, and any affiliated organizations or persons from any and all claims or damages.

📢 Update (Date: 09-22-2023)

🚀 We're thrilled to announce that our paper, "PIXIU: A Comprehensive Benchmark, Instruction Dataset and Large Language Model for Finance", has been accepted by NeurIPS 2023 Track Datasets and Benchmarks!

📢 Update (Date: 10-08-2023)

🌏 We're proud to share that the enhanced versions of FinBen, which now support both Chinese and Spanish!

📢 Update (Date: 02-20-2024)

🌏 We're delighted to share that our paper, "The FinBen: An Holistic Financial Benchmark for Large Language Models", is now available at FinBen.

📢 Update (Date: 05-02-2024)

🌏 We're pleased to invite you to attend the IJCAI2024-challenge, "Financial Challenges in Large Language Models - FinLLM", the starter-kit is available at Starter-kit.

Checkpoints:

Languages

Papers

Evaluations:

English Evaluation Datasets (More details on FinBen section)
Spanish Evaluation Datasets
Chinese Evaluation Datasets

Sentiment Analysis

Classification

Knowledge Extraction

Number Understanding

Text Summarization

Credit Scoring

Forecasting

Overview

Welcome to the PIXIU project! This project is designed to support the development, fine-tuning, and evaluation of Large Language Models (LLMs) in the financial domain. PIXIU is a significant step towards understanding and harnessing the power of LLMs in the financial domain.

Structure of the Repository

The repository is organized into several key components, each serving a unique purpose in the financial NLP pipeline:

FinBen: Our Financial Language Understanding and Prediction Evaluation Benchmark. FinBen serves as the evaluation suite for financial LLMs, with a focus on understanding and prediction tasks across various financial contexts.
FIT: Our Financial Instruction Dataset. FIT is a multi-task and multi-modal instruction dataset specifically tailored for financial tasks. It serves as the training ground for fine-tuning LLMs for these tasks.
FinMA: Our Financial Large Language Model (LLM). FinMA is the core of our project, providing the learning and prediction power for our financial tasks.

Key Features

Open resources: PIXIU openly provides the financial LLM, instruction tuning data, and datasets included in the evaluation benchmark to encourage open research and transparency.
Multi-task: The instruction tuning data and benchmark in PIXIU cover a diverse set of financial tasks, including four financial NLP tasks and one financial prediction task.
Multi-modality: PIXIU's instruction tuning data and benchmark consist of multi-modality financial data, including time series data from the stock movement prediction task. It covers various types of financial texts, including reports, news articles, tweets, and regulatory filings.
Diversity: Unlike previous benchmarks focusing mainly on financial NLP tasks, PIXIU's evaluation benchmark includes critical financial prediction tasks aligned with real-world scenarios, making it more challenging.

FinBen 2.0: Financial Language Understanding and Prediction Evaluation Benchmark

In this section, we provide a detailed performance analysis of FinMA compared to other leading models, including ChatGPT, GPT-4, and BloombergGPT et al. For this analysis, we've chosen a range of tasks and metrics that span various aspects of financial Natural Language Processing and financial prediction. All model results of FinBen can be found on our leaderboard!

Tasks

Data	Task	Raw	Data Types	Modalities	License	Paper
FPB	sentiment analysis	4,845	news	text	CC BY-SA 3.0	[1]
FiQA-SA	sentiment analysis	1,173	news headlines, tweets	text	Public	[2]
TSA	sentiment analysis	561	news headlines	text	CC BY-NC-SA 4.0	[3]
FOMC	hawkish-dovish classification	496	FOMC transcripts	text	CC BY-NC 4.0	[4]
Headlines	news headline classification	11,412	news headlines	text	CC BY-SA 3.0	[5]
FinArg-ECC-Task1	argument unit classification	969	earnings conference call	text	CC BY-NC-SA 4.0	[6]
FinArg-ECC-Task2	argument relation classification	690	earnings conference call	text	CC BY-NC-SA 4.0	[6]
Multifin EN	multi-class classification	546	article headlines	text	Public	[7]
M&A	deal completeness classification	500	news articles, tweets	text	Public	[8]
MLESG EN	ESG Issue Identification	300	news articles	text	CC BY-NC-ND	[9]
NER	named entity recognition	1,366	financial agreements	text	CC BY-SA 3.0	[10]
Finer Ord	named entity recognition	1,080	news articles	text	CC BY-NC 4.0	[11]
FinRED	relation extraction	1,070	earning call transcipts	text	Public	[12]
FinCausual 2020 Task1	causal classification	8,630	news articles, SEC	text	CC BY 4.0	[13]
FinCausual 2020 Task2	causal detection	226	news articles, SEC	text	CC BY 4.0	[13]
FinQA	question answering	8,281	earnings reports	text, table	MIT License	[14]
TatQA	question answering	1,670	financial reports	text, table	MIT License	[15]
FNXL	numeric labeling	318	SEC	text	Public	[16]
FSRL	token classification	97	news articles	text	MIT License	[17]
ECTSUM	text summarization	495	earning call transcipts	text	Public	[18]
EDTSUM	text summarization	2000	news articles	text	Public	[19]
German	credit scoring	1000	credit records	table	CC BY 4.0	[20]
Australian	credit scoring	690	credit records	table	CC BY 4.0	[21]
Lending Club	credit scoring	1,3453	financial information	table	CC0 1.0	[22]
BigData22	stock movement prediction	7,164	tweets, historical prices	text, time series	Public	[23]
ACL18	stock movement prediction	27,053	tweets, historical prices	text, time series	MIT License	[24]
CIKM18	stock movement prediction	4,967	tweets, historical prices	text, time series	Public	[25]
ConvFinQA	multi-turn question answering	1,490	earnings reports	text, table	MIT License	[26]
Credit Card Fraud	Fraud Detection	11,392	financial information	table	(DbCL) v1.0	[22]
ccFraud	Fraud Detection	10,485	financial information	table	Public	[22]
Polish	Financial Distress Identification	8,681	financial status features	table	CC BY 4.0	[22]
Taiwan Economic Journal	Financial Distress Identification	6,819	financial status features	table	CC BY 4.0	[22]
PortoSeguro	Claim Analysis	11,904	claim and financial information	table	Public	[22]
Travel Insurance	Claim Analysis	12,665	claim and financial information	table	(ODbL) v1.0	[22]

1. Pekka Malo, Ankur Sinha, Pekka Korhonen, Jyrki Wallenius, and Pyry Takala. 2014. Good debt or bad debt: Detecting semantic orientations in economic texts. Journal of the Association for Information Science and Technology 65, 4 (2014), 782–796.

2. Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. 2018. Www’18 open challenge: financial opinion mining and question answering. In Companion proceedings of the the web conference 2018. 1941–1942.

3. Keith Cortis, André Freitas, Tobias Daudert, Manuela Huerlimann, Manel Zarrouk, Siegfried Handschuh, and Brian Davis. 2017. SemEval-2017 Task 5: Fine-Grained Sentiment Analysis on Financial Microblogs and News. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 519–535, Vancouver, Canada. Association for Computational Linguistics.

4. Agam Shah, Suvan Paturi, and Sudheer Chava. 2023. Trillion Dollar Words: A New Financial Dataset, Task & Market Analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6664–6679, Toronto, Canada. Association for Computational Linguistics.

5. Ankur Sinha and Tanmay Khandait. 2021. Impact of news on the commodity market: Dataset and results. In Advances in Information and Communication: Proceedings of the 2021 Future of Information and Communication Conference (FICC), Volume 2. Springer, 589–601.

6. Chen C C, Lin C Y, Chiu C J, et al. Overview of the NTCIR-17 FinArg-1 Task: Fine-grained argument understanding in financial analysis[C]//Proceedings of the 17th NTCIR Conference on Evaluation of Information Access Technologies, Tokyo, Japan. 2023.

7. Rasmus Jørgensen, Oliver Brandt, Mareike Hartmann, Xiang Dai, Christian Igel, and Desmond Elliott. 2023. MultiFin: A Dataset for Multilingual Financial NLP. In Findings of the Association for Computational Linguistics: EACL 2023, pages 894–909, Dubrovnik, Croatia. Association for Computational Linguistics.

8. Yang, L., Kenny, E.M., Ng, T.L., Yang, Y., Smyth, B., & Dong, R. (2020). Generating Plausible Counterfactual Explanations for Deep Transformers in Financial Text Classification. International Conference on Computational Linguistics.

9. Chung-Chi Chen, Yu-Min Tseng, Juyeon Kang, Anaïs Lhuissier, Min-Yuh Day, Teng-Tsai Tu, and Hsin-Hsi Chen. 2023. Multi-lingual esg issue identification. In Proceedings of the Fifth Workshop on Financial Tech- nology and Natural Language Processing (FinNLP) and the Second Multimodal AI For Financial Fore- casting (Muffin).

10. Julio Cesar Salinas Alvarado, Karin Verspoor, and Timothy Baldwin. 2015. Domain adaption of named entity recognition to support credit risk assessment. In Proceedings of the Australasian Language Technology Association Workshop 2015. 84–90.

11. Shah A, Vithani R, Gullapalli A, et al. Finer: Financial named entity recognition dataset and weak-supervision model[J]. arXiv preprint arXiv:2302.11157, 2023.

12. Sharma, Soumya et al. “FinRED: A Dataset for Relation Extraction in Financial Domain.” Companion Proceedings of the Web Conference 2022 (2022): n. pag.

13. Dominique Mariko, Hanna Abi-Akl, Estelle Labidurie, Stephane Durfort, Hugues De Mazancourt, and Mahmoud El-Haj. 2020. The Financial Document Causality Detection Shared Task (FinCausal 2020). In Proceedings of the 1st Joint Workshop on Financial Narrative Processing and MultiLing Financial Summarisation, pages 23–32, Barcelona, Spain (Online). COLING.

14. Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan R Routledge, et al . 2021. FinQA: A Dataset of Numerical Reasoning over Financial Data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 3697–3711.

15. Zhu, Fengbin, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng and Tat-Seng Chua. “TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance.” ArXiv abs/2105.07624 (2021): n. pag.

16. Soumya Sharma, Subhendu Khatuya, Manjunath Hegde, Afreen Shaikh, Koustuv Dasgupta, Pawan Goyal, and Niloy Ganguly. 2023. Financial Numeric Extreme Labelling: A dataset and benchmarking. In Findings of the Association for Computational Linguistics: ACL 2023, pages 3550–3561, Toronto, Canada. Association for Computational Linguistics.

17. Matthew Lamm, Arun Chaganty, Christopher D. Manning, Dan Jurafsky, and Percy Liang. 2018. Textual Analogy Parsing: What’s Shared and What’s Compared among Analogous Facts. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 82–92, Brussels, Belgium. Association for Computational Linguistics.

18. Rajdeep Mukherjee, Abhinav Bohra, Akash Banerjee, Soumya Sharma, Manjunath Hegde, Afreen Shaikh, Shivani Shrivastava, Koustuv Dasgupta, Niloy Ganguly, Saptarshi Ghosh, and Pawan Goyal. 2022. ECTSum: A New Benchmark Dataset For Bullet Point Summarization of Long Earnings Call Transcripts. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10893–10906, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.

19. Zhihan Zhou, Liqian Ma, and Han Liu. 2021. Trade the Event: Corporate Events Detection for News-Based Event-Driven Trading. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2114–2124, Online. Association for Computational Linguistics.

20. Hofmann,Hans. (1994). Statlog (German Credit Data). UCI Machine Learning Repository. https://doi.org/10.24432/C5NC77.

21. Quinlan,Ross. Statlog (Australian Credit Approval). UCI Machine Learning Repository. https://doi.org/10.24432/C59012.

22. Duanyu Feng, Yongfu Dai, Jimin Huang, Yifang Zhang, Qianqian Xie, Weiguang Han, Alejandro Lopez-Lira, Hao Wang. 2023. Empowering Many, Biasing a Few: Generalist Credit Scoring through Large Language Models. ArXiv abs/2310.00566 (2023): n. pag.

23. Yejun Soun, Jaemin Yoo, Minyong Cho, Jihyeong Jeon, and U Kang. 2022. Accurate Stock Movement Prediction with Self-supervised Learning from Sparse Noisy Tweets. In 2022 IEEE International Conference on Big Data (Big Data). IEEE, 1691–1700.

24. Yumo Xu and Shay B Cohen. 2018. Stock movement prediction from tweets and historical prices. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1970–1979.

25. Huizhe Wu, Wei Zhang, Weiwei Shen, and Jun Wang. 2018. Hybrid deep sequential modeling for social text-driven stock prediction. In Proceedings of the 27th ACM international conference on information and knowledge management. 1627–1630.

26. Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. 2022. ConvFinQA: Exploring the Chain of Numerical Reasoning in Conversational Finance Question Answering. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6279–6292, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.

Evaluation

Preparation

Locally install

git clone https://github.com/The-FinAI/PIXIU.git --recursive
cd PIXIU
pip install -r requirements.txt
cd src/financial-evaluation
pip install -e .[multilingual]

Docker image

sudo bash scripts/docker_run.sh

Above command starts a docker container, you can modify docker_run.sh to fit your environment. We provide pre-built image by running sudo docker pull tothemoon/pixiu:latest

docker run --gpus all --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 \
    --network host \
    --env https_proxy=$https_proxy \
    --env http_proxy=$http_proxy \
    --env all_proxy=$all_proxy \
    --env HF_HOME=$hf_home \
    -it [--rm] \
    --name pixiu \
    -v $pixiu_path:$pixiu_path \
    -v $hf_home:$hf_home \
    -v $ssh_pub_key:/root/.ssh/authorized_keys \
    -w $workdir \
    $docker_user/pixiu:$tag \
    [--sshd_port 2201 --cmd "echo 'Hello, world!' && /bin/bash"]

Arguments explain:

[] means ignoreable arguments
HF_HOME: huggingface cache dir
sshd_port: sshd port of the container, you can run ssh -i private_key -p $sshd_port root@$ip to connect to the container, default to 22001
--rm: remove the container when exit container (ie.CTRL + D)

Automated Task Assessment

Before evaluation, please download BART checkpoint to src/metrics/BARTScore/bart_score.pth.

For automated evaluation, please follow these instructions:

Huggingface Transformer

To evaluate a model hosted on the HuggingFace Hub (for instance, finma-7b-full), use this command:

python eval.py \
    --model "hf-causal-llama" \
    --model_args "use_accelerate=True,pretrained=TheFinAI/finma-7b-full,tokenizer=TheFinAI/finma-7b-full,use_fast=False" \
    --tasks "flare_ner,flare_sm_acl,flare_fpb"

More details can be found in the lm_eval documentation.

Commercial APIs

Please note, for tasks such as NER, the automated evaluation is based on a specific pattern. This might fail to extract relevant information in zero-shot settings, resulting in relatively lower performance compared to previous human-annotated results.

export OPENAI_API_SECRET_KEY=YOUR_KEY_HERE
python eval.py \
    --model gpt-4 \
    --tasks flare_ner,flare_sm_acl,flare_fpb

Self-Hosted Evaluation

To run inference backend:

bash scripts/run_interface.sh

Please adjust run_interface.sh according to your environment requirements.

To evaluate:

python data/*/evaluate.py

Create new tasks

Creating a new task for FinBen involves creating a Huggingface dataset and implementing the task in a Python file. This guide walks you through each step of setting up a new task using the FinBen framework.

Creating your dataset in Huggingface

Your dataset should be created in the following format:

{
    "query": "...",
    "answer": "...",
    "text": "..."
}

In this format:

query: Combination of your prompt and text
answer: Your label

For Multi-turn tasks (such as )

For Classification tasks (such as FPB (FinBen_fpb)), additional keys should be defined:

choices: Set of labels
gold: Index of the correct label in choices (Start from 0)

For Sequential Labeling tasks (such as Finer Ord (FinBen_finer_ord)), additional keys should be defined:

label: List of token labels
token: List of tokens

For Extractive Summarization tasks (such as ECTSUM (FinBen_ectsum)), additional keys should be defined:

label: List of sentence labels

For abstractive Summarization and Question Answering tasks (such as EDTSUM (FinBen_edtsum)), no additional keys should be defined

Implementing the task

Once your dataset is ready, you can start implementing your task. Your task should be defined within a new class in flare.py or any other Python file located within the tasks directory.

To cater to a range of tasks, we offer several specialized base classes, including Classification, SequentialLabeling, RelationExtraction, ExtractiveSummarization, AbstractiveSummarization and QA.

For instance, if you are embarking on a classification task, you can directly leverage our Classification base class. This class allows for efficient and intuitive task creation. To better demonstrate this, let's delve into an example of crafting a task named FinBen-FPB using the Classification base class:

class flareFPB(Classification):
    DATASET_PATH = "flare-fpb"

And that's it! Once you've created your task class, the next step is to register it in the src/tasks/__init__.py file. To do this, add a new line following the format "task_name": module.ClassName. Here is how it's done:

TASK_REGISTRY = {
    "flare_fpb": flare.FPB,
    "your_new_task": your_module.YourTask,  # This is where you add your task
}

Predefined task metrics

Task	Metric	Illustration
Classification	Accuracy	This metric represents the ratio of correctly predicted observations to total observations. It is calculated as (True Positives + True Negatives) / Total Observations.
Classification	F1 Score	The F1 Score represents the harmonic mean of precision and recall, thereby creating an equilibrium between these two factors. It proves particularly useful in scenarios where one factor bears more significance than the other. The score ranges from 0 to 1, with 1 signifying perfect precision and recall, and 0 indicating the worst case. Furthermore, we provide both 'weighted' and 'macro' versions of the F1 score.
Classification	Missing Ratio	This metric calculates the proportion of responses where no options from the given choices in the task are returned.
Classification	Matthews Correlation Coefficient (MCC)	The MCC is a metric that assesses the quality of binary classifications, producing a score ranging from -1 to +1. A score of +1 signifies perfect prediction, 0 denotes a prediction no better than random chance, and -1 indicates a completely inverse prediction.
Sequential Labeling	F1 score	In the context of Sequential Labeling tasks, we utilize the F1 Score as computed by the `seqeval` library, a robust entity-level evaluation metric. This metric mandates an exact match of both the entity's span and type between the predicted and ground truth entities for a correct evaluation. True Positives (TP) represent correctly predicted entities, False Positives (FP) denote incorrectly predicted entities or entities with mismatched spans/types, and False Negatives (FN) signify missed entities from the ground truth. Precision, recall, and F1-score are then computed using these quantities, with the F1 Score representing the harmonic mean of precision and recall.
Sequential Labeling	Label F1 score	This metric evaluates model performance based solely on the correctness of the labels predicted, without considering entity spans.
Relation Extraction	Precision	Precision measures the proportion of correctly predicted relations out of all predicted relations. It is calculated as the number of True Positives (TP) divided by the sum of True Positives and False Positives (FP).
Relation Extraction	Recall	Recall measures the proportion of correctly predicted relations out of all actual relations. It is calculated as the number of True Positives (TP) divided by the sum of True Positives and False Negatives (FN).
Relation Extraction	F1 score	The F1 Score is the harmonic mean of precision and recall, and it provides a balance between these two metrics. The F1 Score is at its best at 1 (perfect precision and recall) and worst at 0.
Extractive and Abstractive Summarization	Rouge-N	This measures the overlap of N-grams (a contiguous sequence of N items from a given sample of text) between the system-generated summary and the reference summary. 'N' can be 1, 2, or more, with ROUGE-1 and ROUGE-2 being commonly used to assess unigram and bigram overlaps respectively.
Extractive and Abstractive Summarization	Rouge-L	This metric evaluates the longest common subsequence (LCS) between the system and the reference summaries. LCS takes into account sentence level structure similarity naturally and identifies longest co-occurring in-sequence n-grams automatically.
Question Answering	EmACC	EMACC assesses the exact match between the model-generated response and the reference answer. In other words, the model-generated response is considered correct only if it matches the reference answer exactly, word-for-word.

Additionally, you can determine if the labels should be lowercased during the matching process by specifying LOWER_CASE in your class definition. This is pertinent since labels are matched based on their appearance in the generated output. For tasks like examinations where the labels are a specific set of capitalized letters such as 'A', 'B', 'C', this should typically be set to False.

FIT: Financial Instruction Dataset

Our instruction dataset is uniquely tailored for the domain-specific LLM, FinMA. This dataset has been meticulously assembled to fine-tune our model on a diverse range of financial tasks. It features publicly available multi-task and multi-modal data derived from the multiple open released financial datasets.

The dataset is multi-faceted, featuring tasks including sentiment analysis, news headline classification, named entity recognition, question answering, and stock movement prediction. It covers both textual and time-series data modalities, offering a rich variety of financial data. The task specific instruction prompts for each task have been carefully degined by domain experts.

Modality and Prompts

The table below summarizes the different tasks, their corresponding modalities, text types, and examples of the instructions used for each task:

Task	Modalities	Text Types	Instructions Examples
Sentiment Analysis	Text	news headlines,tweets	"Analyze the sentiment of this statement extracted from a financial news article.Provide your answer as either negative, positive or neutral. For instance, 'The company's stocks plummeted following the scandal.' would be classified as negative."
News Headline Classification	Text	News Headlines	"Consider whether the headline mentions the price of gold. Is there a Price or Not in the gold commodity market indicated in the news headline? Please answer Yes or No."
Named Entity Recognition	Text	financial agreements	"In the sentences extracted from financial agreements in U.S. SEC filings, identify the named entities that represent a person ('PER'), an organization ('ORG'), or a location ('LOC'). The required answer format is: 'entity name, entity type'. For instance, in 'Elon Musk, CEO of SpaceX, announced the launch from Cape Canaveral.', the entities would be: 'Elon Musk, PER; SpaceX, ORG; Cape Canaveral, LOC'"
Question Answering	Text	earnings reports	"In the context of this series of interconnected finance-related queries and the additional information provided by the pretext, table data, and post text from a company's financial filings, please provide a response to the final question. This may require extracting information from the context and performing mathematical calculations. Please take into account the information provided in the preceding questions and their answers when formulating your response:"
Stock Movement Prediction	Text, Time-Series	tweets, Stock Prices	"Analyze the information and social media posts to determine if the closing price of {tid} will ascend or descend at {point}. Please respond with either Rise or Fall."

Dataset Statistics

The dataset contains a vast amount of instruction data samples (136K), allowing FinMA to capture the nuances of the diverse financial tasks. The table below provides the statistical details of the instruction dataset:

Data	Task	Raw	Instruction	Data Types	Modalities	License	Original Paper
FPB	sentiment analysis	4,845	48,450	news	text	CC BY-SA 3.0	[1]
FiQA-SA	sentiment analysis	1,173	11,730	news headlines, tweets	text	Public	[2]
Headline	news headline classification	11,412	11,412	news headlines	text	CC BY-SA 3.0	[3]
NER	named entity recognition	1,366	13,660	financial agreements	text	CC BY-SA 3.0	[4]
FinQA	question answering	8,281	8,281	earnings reports	text, table	MIT License	[5]
ConvFinQA	question answering	3,892	3,892	earnings reports	text, table	MIT License	[6]
BigData22	stock movement prediction	7,164	7,164	tweets, historical prices	text, time series	Public	[7]
ACL18	stock movement prediction	27,053	27,053	tweets, historical prices	text, time series	MIT License	[8]
CIKM18	stock movement prediction	4,967	4,967	tweets, historical prices	text, time series	Public	[9]

Pekka Malo, Ankur Sinha, Pekka Korhonen, Jyrki Wallenius, and Pyry Takala. 2014. Good debt or bad debt: Detecting semantic orientations in economic texts. Journal of the Association for Information Science and Technology 65, 4 (2014), 782–796.
Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. 2018. Www’18 open challenge: financial opinion mining and question answering. In Companion proceedings of the the web conference 2018. 1941–1942
Ankur Sinha and Tanmay Khandait. 2021. Impact of news on the commodity market: Dataset and results. In Advances in Information and Communication: Proceedings of the 2021 Future of Information and Communication Conference (FICC), Volume 2. Springer, 589–601
Julio Cesar Salinas Alvarado, Karin Verspoor, and Timothy Baldwin. 2015. Domain adaption of named entity recognition to support credit risk assessment. In Proceedings of the Australasian Language Technology Association Workshop 2015. 84–90.
Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan R Routledge, et al . 2021. FinQA: A Dataset of Numerical Reasoning over Financial Data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 3697–3711.
Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. 2022. Convfinqa: Exploring the chain of numerical reasoning in conversational finance question answering. arXiv preprint arXiv:2210.03849 (2022).
Yejun Soun, Jaemin Yoo, Minyong Cho, Jihyeong Jeon, and U Kang. 2022. Accurate Stock Movement Prediction with Self-supervised Learning from Sparse Noisy Tweets. In 2022 IEEE International Conference on Big Data (Big Data). IEEE, 1691–1700.
Yumo Xu and Shay B Cohen. 2018. Stock movement prediction from tweets and historical prices. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1970–1979.
Huizhe Wu, Wei Zhang, Weiwei Shen, and Jun Wang. 2018. Hybrid deep sequential modeling for social text-driven stock prediction. In Proceedings of the 27th ACM international conference on information and knowledge management. 1627–1630.

Generating Datasets for FIT

When you are working with the Financial Instruction Dataset (FIT), it's crucial to follow the prescribed format for training and testing models.

The format should look like this:

{
    "id": "unique id",
    "conversations": [
        {
            "from": "human",
            "value": "Your prompt and text"
        },
        {
            "from": "agent",
            "value": "Your answer"
        }
    ],
    "text": "Text to be classified",
    "label": "Your label"
}

Here's what each field means:

"id": a unique identifier for each example in your dataset.
"conversations": a list of conversation turns. Each turn is represented as a dictionary, with "from" representing the speaker, and "value" representing the text spoken in the turn.
"text": the text to be classified.
"label": the ground truth label for the text.

The first turn in the "conversations" list should always be from "human", and contain your prompt and the text. The second turn should be from "agent", and contain your answer.

FinMA v0.1: Financial Large Language Model

We are pleased to introduce the first version of FinMA, including three models FinMA-7B, FinMA-7B-full, FinMA-30B, fine-tuned on LLaMA 7B and LLaMA-30B. FinMA-7B and FinMA-30B are trained with the NLP instruction data, while FinMA-7B-full is trained with the full instruction data from FIT covering both NLP and prediction tasks.

FinMA v0.1 is now available on Huggingface for public use. We look forward to the valuable contributions that this initial version will make to the financial NLP field and encourage users to apply it to various financial tasks and scenarios. We also invite feedback and shared experiences to help improve future versions.

How to fine-tune a new large language model using PIXIU based on FIT?

Coming soon.

FinMem: A Performance-Enhanced LLM Trading Agent

FinMem is a novel LLM-based agent framework devised for financial decision-making, encompasses three core modules: Profiling, to outline the agent's characteristics; Memory, with layered processing, to aid the agent in assimilating realistic hierarchical financial data; and Decision-making, to convert insights gained from memories into investment decisions. Currently, FinMem can trade single stocks with high returns after a simple mode warm-up. Below is a quick start for a dockerized version framework, with TSLA as sample input.

Step 1: Set environmental variables in .env add HUGGINGFACE TOKEN and OPENAI API KEY as needed.

OPENAI_API_KEY = "<Your OpenAI Key>"
HF_TOKEN = "<Your HF token>"

Step 2: Set endpoint URL in config.toml Use endpoint URL to deploy models based on the model of choice (OPENAI, Gemini, open source models on HuggingFace, etc.). For open-source models on HuggingFace, one choice for generating TGI endpoints is through RunPod.

[chat]
model = "tgi"
end_point = "<set the your endpoint address>"
tokenization_model_name = "<model name>"
...

Step 3: Build Docker Image and Container

docker build -t test-finmem .devcontainer/.

start container:

docker run -it --rm -v $(pwd):/finmem test-finmem bash

Step 4: Start Simulation!

 Usage: run.py sim [OPTIONS]                                                                                                                
                                                                                                                                            
 Start Simulation                                                                                                                           
                                                                                                                                            
╭─ Options ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ --market-data-path    -mdp      TEXT  The environment data pickle path [default: data/06_input/subset_symbols.pkl]                       │
│ --start-time          -st       TEXT  The training or test start time [default: 2022-06-30 For Ticker 'TSLA']                                                               │
│ --end-time            -et       TEXT  The training or test end time [default: 2022-10-11]                                                                 │
│ --run-model           -rm       TEXT  Run mode: train or test [default: train]                                                           │
│ --config-path         -cp       TEXT  config file path [default: config/config.toml]                                                     │
│ --checkpoint-path     -ckp      TEXT  The checkpoint save path [default: data/10_checkpoint_test]                                             │
│ --result-path         -rp       TEXT  The result save path [default: data/11_train_result]                                               │
│ --trained-agent-path  -tap      TEXT  Only used in test mode, the path of trained agent [default: None. Can be changed to data/05_train_model_output OR data/06_train_checkpoint]                                  │
│ --help                                Show this message and exit.                                                                        │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

Example Usage:

python run.py sim --market-data-path data/03_model_input/tsla.pkl --start-time 2022-06-30 --end-time 2022-10-11 --run-model train --config-path config/tsla_tgi_config.toml --checkpoint-path data/06_train_checkpoint --result-path data/05_train_model_output

There are also checkpoint functionalities. For more details please visit FinMem Repository directly.

Citation

If you use PIXIU in your work, please cite our paper.

@misc{xie2023pixiu,
      title={PIXIU: A Large Language Model, Instruction Data and Evaluation Benchmark for Finance}, 
      author={Qianqian Xie and Weiguang Han and Xiao Zhang and Yanzhao Lai and Min Peng and Alejandro Lopez-Lira and Jimin Huang},
      year={2023},
      eprint={2306.05443},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{xie2024FinBen,
      title={The FinBen: An Holistic Financial Benchmark for Large Language Models}, 
      author={Qianqian Xie and Weiguang Han and Zhengyu Chen and Ruoyu Xiang and Xiao Zhang and Yueru He and Mengxi Xiao and Dong Li and Yongfu Dai and Duanyu Feng and Yijing Xu and Haoqiang Kang and Ziyan Kuang and Chenhan Yuan and Kailai Yang and Zheheng Luo and Tianlin Zhang and Zhiwei Liu and Guojun Xiong and Zhiyang Deng and Yuechen Jiang and Zhiyuan Yao and Haohang Li and Yangyang Yu and Gang Hu and Jiajia Huang and Xiao-Yang Liu and Alejandro Lopez-Lira and Benyou Wang and Yanzhao Lai and Hao Wang and Min Peng and Sophia Ananiadou and Jimin Huang},
      year={2024},
      eprint={2402.12659},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

License

PIXIU is licensed under [MIT]. For more details, please see the MIT file.

Name		Name	Last commit message	Last commit date
Latest commit History 255 Commits
.cache		.cache
FinMem-LLM-StockTrading @ 0aae217		FinMem-LLM-StockTrading @ 0aae217
docker		docker
notebooks		notebooks
scripts		scripts
src		src
static		static
.gitignore		.gitignore
.gitmodules		.gitmodules
LICENSE		LICENSE
README.es.md		README.es.md
README.ijcai_challenge.md		README.ijcai_challenge.md
README.md		README.md
README.zh.md		README.zh.md
requirements.txt		requirements.txt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Overview

Structure of the Repository

Key Features

FinBen 2.0: Financial Language Understanding and Prediction Evaluation Benchmark

Tasks

Evaluation

Preparation

Locally install

Docker image

Automated Task Assessment

Create new tasks

Creating your dataset in Huggingface

Implementing the task

Predefined task metrics

FIT: Financial Instruction Dataset

Modality and Prompts

Dataset Statistics

Generating Datasets for FIT

FinMA v0.1: Financial Large Language Model

How to fine-tune a new large language model using PIXIU based on FIT?

FinMem: A Performance-Enhanced LLM Trading Agent

Citation

License

Star History

About

Packages

Contributors 13

Languages

License

The-FinAI/PIXIU

Folders and files

Latest commit

History

Repository files navigation

Overview

Structure of the Repository

Key Features

FinBen 2.0: Financial Language Understanding and Prediction Evaluation Benchmark

Tasks

Evaluation

Preparation

Locally install

Docker image

Automated Task Assessment

Create new tasks

Creating your dataset in Huggingface

Implementing the task

Predefined task metrics

FIT: Financial Instruction Dataset

Modality and Prompts

Dataset Statistics

Generating Datasets for FIT

FinMA v0.1: Financial Large Language Model

How to fine-tune a new large language model using PIXIU based on FIT?

FinMem: A Performance-Enhanced LLM Trading Agent

Citation

License

Star History

About

Topics

Resources

License

Stars

Watchers

Forks

Packages 0

Contributors 13

Languages

Packages