README.md

---
tags:
  - financial NLP
  - named entity recognition
  - XBRL
task_categories:
  - token-classification
  - text-classification
task_ids:
  - named-entity-recognition
pretty_name: "HiFi-KPI Lite: Expert-Curated Financial KPI Extraction"
dataset_name: "HiFi-KPI Lite"
size_categories:
  - 10K<n<100K
language:
  - en
---

# HiFi-KPI Lite: Expert-Curated Financial KPI Extraction

## Dataset Summary
HiFi-KPI Lite is a manually curated subset of the HiFi-KPI dataset, designed for evaluating structured financial KPI extraction. Unlike the full **[HiFi-KPI dataset](https://huggingface.co/datasets/AAU-NLP/HiFi-KPI)**, HiFi-KPI Lite maps financial entities to a much reduced, expert-defined label space. The dataset consists of **∼8K paragraphs** and **∼25K entities**, making it suitable for rapid model evaluation.

## Supported Tasks
The dataset is optimized for:
- **Named Entity Recognition (NER)**: Identifying financial KPIs from text.
- **Structured Data Extraction**: Extracting numerical values, currencies, and corresponding time periods from financial statements.
- **Text Classification**: Associating financial statements with expert-mapped KPI labels.

## Languages
The dataset is in **English**, sourced from SEC 10-K and 10-Q filings.

## Dataset Structure
### Data Fields
Each entry in HiFi-KPI Lite includes:
- **form_type**: "10-K" or "10-Q"
- **accession_number**: Unique filing identifier
- **filing_date**: Timestamp of the filing
- **company_name**: Name of the reporting entity
- **text**: Extracted paragraph from the filing
- **entities** (list of extracted entities):
  - **label**: Expert-defined financial KPI category(Ebit, revnues ..)
  - **start_date_for_period** / **end_date_for_period**: Time period of the financial figure
  - **currency/unit**: Currency (e.g., USD, EUR)
  - **value**: Extracted numerical figure

### Dataset Statistics
| Split  | # Paragraphs | # Entities |
|--------|------------|------------|
| Train  | 6,359      | 19,749     |
| Dev    | 768        | 2,601      |
| Test   | 856        | 2,437      |

## Baseline Model Performance
We establish baselines using:
- **Sequence Labeling**: fine-tuning **BERT (bert-base-uncased)** with a token classification head.
- **LLM-based Structured Extraction**: Few-shot prompting for **NuExtract, Qwen-2.5-14B, and DeepSeek-V3**.

### Macro F1 Performance on HiFi-KPI Lite
| Model          | Precision | Recall | Micro F1 |
|---------------|----------|--------|----------|
| [BERT (SL)](https://huggingface.co/AAU-NLP/Lite-BERT-SL)    | 89.2     | 91.8   | 89.1     |
| Qwen-2.5-14B | 63.7     | 60.2   | 49.5     |
| DeepSeek-V3  | 67.8     | 65.9   | 46.4     |

## Uses and Applications
HiFi-KPI Lite is useful for:
- **Benchmarking KPI Extraction**: Evaluating model performance on structured financial data extraction.
- **Fine-Grained Financial NLP Tasks**: Training models on structured financial entity recognition.
- **Evaluation of Large Language Models (LLMs)**: Testing how well LLMs generalize to financial entity extraction.

## Citation
If you use HiFi-KPI Lite in your research, please cite:
```
@article{aavang2025hifi,
  title={HiFi-KPI: A Dataset for Hierarchical KPI Extraction from Earnings Filings},
  author={Aavang, Rasmus and Rizzi, Giovanni and Bøggild, Rasmus and Iolov, Alexandre and Zhang, Mike and Bjerva, Johannes},
  journal={arXiv preprint arXiv:2502.15411},
  year={2025}
}
```

## Access
Example code and repo with links to all models at [GitHub Repository](https://github.com/rasmus393/HiFi-KPI)