README.md

---
task_categories:
- feature-extraction
- sentence-similarity
language:
- en
size_categories:
- 1K<n<10K
license: mit
---

# Dataset Card for FinSTS Golden Dataset

## Table of Contents
- [Dataset Summary](#dataset-summary)
- [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
- [Languages](#languages)
- [Dataset Structure](#dataset-structure)
- [Data Splits](#data-splits)
- [Dataset Creation](#dataset-creation)
- [Citing & Authors](#citing--authors)

## Dataset Summary

The FinSTS Golden Dataset is designed for financial semantic textual similarity tasks. It contains a development set of 2001 sentence pairs and a test set of 1999 sentence pairs, annotated to reflect the degree of semantic similarity between financial texts. The sentence pairs are collected from earnings call transcripts and 10-K filings, making this dataset ideal for evaluating models in financial text analysis.

## Supported Tasks and Leaderboards

- **Text Similarity**: This dataset can be used to evaluate models measuring the semantic similarity between pairs of financial sentences.

## Languages

The dataset is in English.

## Dataset Structure

### Data Fields

- `split`: Indicates whether the data point belongs to the development or test set.
- `sentence1`: The first sentence in the pair.
- `sentence2`: The second sentence in the pair.
- `gpt_score`: An integer score between 0 and 5 indicating the degree of semantic similarity between the two sentences, annotated by GPT-4 Turbo API.
- `score`: An integer score between 0 and 5 indicating the degree of semantic similarity between the two sentences, annotated by 4 human experts.
### Example

```json
{
  "split": "dev",
  "sentence1": "Unlike many industrial companies, substantially all of our assets and virtually all of our liabilities are monetary in nature.",
  "sentence2": "Unlike most industrial companies, virtually all of our assets and liabilities are monetary in nature.",
  "gpt_score": 5,
  "score": 4
}