README.md

---
extra_gated_fields:
  First name: text
  Last name: text
  Affiliation: text
  Job title: text
  Email: text
  Country: country
  I want to use this dataset for:
    type: select
    options: 
      - Research
      - Education
      - label: Other
        value: other

language:
- en
- fr
- es
- it
- de
license: mit
size_categories:
- 1K<n<10K
task_categories:
- translation
pretty_name: DOLFIN
tags:
- finance
- legal
dataset_info:
  features:
  - name: source_text
    dtype: string
  - name: target_text
    dtype: string
  - name: sub_domain
    dtype: string
  - name: date
    dtype: string
  - name: id
    dtype: int64
  - name: Annotation
    struct:
    - name: ctxpro_annotations
      dtype: string
    - name: ctxpro_rules
      dtype: string
    - name: llm_analysis
      dtype: string
    - name: llm_score
      dtype: float64
  - name: src_lang
    dtype: string
  - name: trg_lang
    dtype: string
  splits:
  - name: test
    num_bytes: 55487535
    num_examples: 9759
  download_size: 25566431
  dataset_size: 55487535
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---
# Dataset Card

<!-- This dataset card aims to be a base template for new datasets. It has been generated using [this raw template](https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/templates/datasetcard_template.md?plain=1). -->

DOLFIN was created as part of an academic research project led by [Dragon LLM](https://dragonllm.ai) and the [Laboratoire d'Informatique de Grenoble](https://www.liglab.fr/fr), group [GETALP](https://lig-getalp.imag.fr/fr/accueil/). 

It is built from publicly available documents that are specialised in finance, containing mainly financial regulatory texts, that underwent a long processing pipeline.
The aim of this novel test-set is the evaluation of document-level machine translation in the domain of finance.
DOLFIN makes a step towards true document-level MT by abandoning the paradigm of aligned sentences and presenting data in units of sections rather than sentences. 
The test-set consists of an average of 1950 aligned sections for five language pairs.
It contains automatic annotations about the presence of context-sensitive phenomena that require extra-sentential context to be correctly translated.

## Uses

This dataset is designed to test the capabilities of LLMs and MT systems to deal with longer contexts and to translate sentences while taking the context into account.

## Dataset Structure

Every segment of the test-set contains the following fields:

- `source_text`: the segment in source language
- `target_text`: the segment in target language
- `src_lang`: source language
- `trg_lang`: target language
- `sub_domain`: document type referring to the sub-domain of finance
- `date`: date of publication of the document
- `comet_slide`: Comet-kiwi-slide score
- `Annotation`: annotations of context-sensitive phenomena (obtained by CTXPRO and Llama-3-70b)
- `id`: unique id of the segment


## Dataset Creation

### Curation Rationale

Despite the strong research interest in document-level Machine Translation (MT), the test-sets dedicated to this task are still
scarce. The existing test-sets mainly cover topics from the general domain and fall short on specialised domains, such as legal
and financial. Also, despite their document-level aspect, they still follow a sentence-level logic that doesn’t allow for including 
certain linguistic phenomena such as information reorganisation. We aim to fill this gap by proposing a novel test-set: DOLFIN.

### Additional information

For additional information about this test-set (such as: the data sources, data processing pipeline, annotation process, etc.), please refer to the paper.


## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

If you use this dataset, please cite:

[**DOLFIN - Document-Level Financial Test-Set for Machine Translation** Mariam Nakhle, Marco Dinarelli, Raheel Qader, Emmanuelle Esperança-Rodier, Hervé Blanchon](https://arxiv.org/abs/2502.03053)

```
@misc{nakhlé2025dolfindocumentlevelfinancial,
      title={DOLFIN -- Document-Level Financial test set for Machine Translation}, 
      author={Mariam Nakhlé and Marco Dinarelli and Raheel Qader and Emmanuelle Esperança-Rodier and Hervé Blanchon},
      year={2025},
      eprint={2502.03053},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.03053}, 
}
```

## Dataset Card Contact

For any enquiries, please contact mariam.nakhle(AT)univ-grenoble-alpes(DOT)fr