TheFinAI/plutus-multifin

---

Dataset Card for Plutus Multifin

Dataset Summary

Plutus Multifin is a topic classification dataset that presents financial queries in a multiple-choice format. Each instance includes a financial headline (query), an associated answer text, additional context, a set of candidate topics (choices), and a gold label indicating the correct topic. This resource is designed to benchmark the performance of language models in categorizing Greek financial news headlines into one of six predefined thematic categories.

Supported Tasks

• Task: Topic Classification
• Evaluation Metrics: Accuracy

Task Definition

The topic classification task is derived from MultiFin with a focus on categorizing financial news headlines into predefined financial topics.

Languages

• Greek

Dataset Structure

Data Instances

Each instance in this dataset consists of the following five fields:

• query: A financial news headline serving as the input text.
• answer: The detailed answer text associated with the headline.
• text: Additional context or background information related to the news headline.
• choices: A list of candidate topics representing various financial categories.
• gold: An integer indicating the index of the correct topic within the choices.

Data Fields

• query: String – Represents the financial news headline.
• answer: String – Provides supplementary detail or explanation corresponding to the headline.
• text: String – Supplemental context offered to aid in topic determination.
• choices: Sequence of strings – A list of possible topics to which the headline may belong.
• gold: Int64 – Specifies the index of the correct topic within the list of choices.

Data Splits

The dataset is divided into three splits:

• Train: 171 examples (146,842 bytes)
• Validation: 43 examples (37,590 bytes)
• Test: 54 examples (46,372 bytes)

Dataset Creation

Curation Rationale

The Plutus Multifin dataset was developed to extend financial topic classification benchmarks within the Plutus collection. By focusing on classifying Greek financial news headlines into coherent topics, it aims to evaluate and enhance the performance of language models on this challenging task.

Source Data

Initial Data Collection and Normalization

• This task derives from the MultiFin dataset, which comprises 10,048 financial article headlines in 15 languages.
• For the GRMultiFin subset, the Greek financial headlines were extracted and standardized, ensuring consistency in language and topic labels.

Who are the Source Language Producers?

• The source data originates from the MultiFin dataset, in which validated financial news headlines from various sources were used. The Greek subset forms the basis for this dataset.

Annotations

Annotation Process

• We adopted the original Greek part of the MultiFin dataset for this task. No further annotation was performed on the Greek subset.

Personal and Sensitive Information

• This dataset does not contain any personally identifiable information (PII) and focuses solely on the classification of public financial news headlines.

Considerations for Using the Data

Social Impact of Dataset

This dataset supports advancements in financial NLP by providing a challenging benchmark for topic classification of Greek financial news. Improved topic classification can enhance automated financial decision-making and support academic research in multilingual financial analysis.

Discussion of Biases

• The dataset's emphasis on Greek financial news may introduce domain-specific linguistic patterns and biases.
• The inherent ambiguity in financial headlines and overlap between topics could affect model performance.
• As we adopted the original annotations from MultiFin without further modification, any biases inherent in the source data are retained.

Other Known Limitations

• Pre-processing may be necessary to address variability in headline phrasing and context.
• Being specialized in the financial domain, the dataset's applicability to more general classification tasks might be limited.

Additional Information

Dataset Curators

• Original dataset: Rasmus Jørgensen, Oliver Brandt, Mareike Hartmann, Xiang Dai, Christian Igel, Desmond Elliott

• Adapted Dataset:

  • Xueqing Peng
  • Triantafillos Papadopoulos
  • Efstathia Soufleri
  • Polydoros Giannouris
  • Ruoyu Xiang
  • Yan Wang
  • Lingfei Qian
  • Jimin Huang
  • Qianqian Xie
  • Sophia Ananiadou
  • The research is supported by NaCTeM, Archimedes RC, and The Fin AI.

Licensing Information

• License: CC BY-NC 4.0

Citation Information

Original Dataset:

@inproceedings{jorgensen2023multifin,
  title={MultiFin: A dataset for multilingual financial NLP},
  author={J{\o}rgensen, Rasmus and Brandt, Oliver and Hartmann, Mareike and Dai, Xiang and Igel, Christian and Elliott, Desmond},
  booktitle={Findings of the Association for Computational Linguistics: EACL 2023},
  pages={894--909},
  year={2023}
}

Adapted Version (Plutus):

@misc{peng2025plutusbenchmarkinglargelanguage,
      title={Plutus: Benchmarking Large Language Models in Low-Resource Greek Finance}, 
      author={Xueqing Peng and Triantafillos Papadopoulos and Efstathia Soufleri and Polydoros Giannouris and Ruoyu Xiang and Yan Wang and Lingfei Qian and Jimin Huang and Qianqian Xie and Sophia Ananiadou},
      year={2025},
      eprint={2502.18772},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.18772}, 
}