| --- |
| dataset_info: |
| features: |
| - name: text1 |
| dtype: string |
| - name: text2 |
| dtype: string |
| - name: label |
| dtype: float32 |
| splits: |
| - name: train |
| num_bytes: 214876315.2 |
| num_examples: 45044 |
| - name: test |
| num_bytes: 53719078.8 |
| num_examples: 11261 |
| download_size: 153023068 |
| dataset_size: 268595394 |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-* |
| - split: test |
| path: data/test-* |
| license: apache-2.0 |
| language: |
| - en |
| task_categories: |
| - sentence-similarity |
| - text-classification |
| tags: |
| - finance |
| - economics |
| size_categories: |
| - 10K<n<100K |
| --- |
| |
| # Dataset card |
|
|
| The dataset econo-pairs-v2 is the second version of "econo-pairs". This dataset is built on top of the ["samchain/bis_central_bank_speeches"](https://huggingface.co/datasets/samchain/bis_central_bank_speeches) dataset. |
| It is made of pairs of sequences of 512 tokens. The pairs are built following this approach: |
| - If both sequences are coming from the same speech, the label is 1.0 |
| - If the sequences are coming from two different speeches but from the same central bank, the label is 0.5 |
| - If the sequences are coming from two different speeches and two different central banks, the label is 0.0 |
|
|
| ## Why a new version ? |
| This dataset presents two improvements vs the prior: |
| - More data cleaning, especially we made our best effort to prevent having non-sense speeches or wrongly utf decoded characters |
| - A better nuance of similarity and dissimilarity as we introduce a middle ground with the 0.5 labels |
|
|
| ## Credits |
| Full credits to the [Bank of International Settlements](https://www.bis.org/) for making the resource available through the internet. |
| Credits to Samuel Chaineau for making it available on HF Hub. |
