Datasets:
Download README.md from LangTech-MT/IMPaCT-APE: direct link, hf CLI and curl.
- Browser
- Download file 5.95 kB
-
https://huggingface.co/datasets/LangTech-MT/IMPaCT-APE/resolve/main/README.md
- Command line
-
hf download hf://datasets/LangTech-MT/IMPaCT-APE/README.md
-
curl -L -o README.md https://huggingface.co/datasets/LangTech-MT/IMPaCT-APE/resolve/main/README.md
task_categories:
- translation
language:
- es
- ca
- en
tags:
- catalan
- spanish
- english
- post-editing
- machine-translation
- ape
- biomedical
- health
pretty_name: IMPaCT APE
size_categories:
- n<1K
license: cc-by-4.0
IMPaCT APE
Dataset Summary
IMPaCT APE is a human post-edited dataset of 149 examples at the sentence and paragraph level that can be used to aid machine translation models in automatic post-editing tasks.
Each example provides an authentic Spanish source text, an English machine translation produced by SalamandraTA-7b-instruct (v3), and a human post-edited final version of that translation.
Additionally, the dataset comes with a synthetically generated Catalan source text that can be used to provide ca→en signal.
Supported Tasks
Automatic Post-Editing (APE): given a Spanish source text and its English machine translation (produced by SalamandraTA-7b-instruct), generate a corrected English translation. The human post-edited version serves as the reference output. The synthetic Catalan source can be used as auxiliary input, but note that the machine translation was produced from the Spanish source, not the Catalan.
Machine Translation (es→en): the Spanish source paired with the human post-edited English text can be used as parallel data for training or evaluating Spanish→English translation.
Machine Translation (ca→en): the synthetically generated Catalan source paired with the post-edited English text provides Catalan→English signal. Because the Catalan side is synthetic, it is better suited to training than to evaluation.
Quality Estimation / MT Evaluation: the difference between the raw MT output and its post-edited version (e.g. measured with HTER) can serve as a quality signal for SalamandraTA-7b-instruct outputs.
Languages
- Catalan (
ca) - Spanish (
es) - English (
en)
Dataset Structure
Data Instances
The dataset is provided in parquet format.
Each row is one multi-parallel example containing four aligned texts:
| Column | Description |
|---|---|
src_es |
Authentic Spanish source text |
src_ca_synth |
Synthetic Catalan source text |
tgt_en_mt |
Raw English machine translation (SalamandraTA-7b-instruct) |
tgt_en_pe |
Human post-edited English translation |
Data Splits
| Split | Examples |
|---|---|
| train | 149 |
The dataset is released as a single split. Users may use it for training or evaluation as suits their needs.
Dataset Creation
Curation Rationale
Created to provide human-verified post-editing signal for SalamandraTA-7b-instruct. Post-edits capture the corrections a human translator makes to real model output, which makes them useful both for training automatic post-editing systems and for analysing the model's translation errors. The source texts are biomedical and health-domain texts from IMPaCT-Data.
Source Data / Generation Process
- Source collection: Authentic Spanish source texts were obtained from the website of IMPaCT-Data (Programa de Ciencia de Datos de la Infraestructura de Medicina de Precisión asociada a la Ciencia y la Tecnología).
- Machine translation: Each Spanish sentence/paragraph was translated into English with SalamandraTA-7b-instruct.
- Human post-editing: The English machine translations were post-edited by human annotators to produce the final reference translations (see Annotations).
- Synthetic Catalan source: A Catalan version of each Spanish source was generated with SalamandraTA-7b-instruct to provide ca→en signal.
Annotations
The English post-edits are human annotations, produced by staff of the BSC Technologies for Biomedical Research Laboratory (TechBioLab).
The Catalan source is not authentic; it is machine-generated.
Considerations for Using the Data
Social Impact of Dataset
By providing human post-edits of biomedical and health-domain texts, this dataset aims to support more reliable machine translation in a domain where translation errors can have real consequences, and to help improve the accessibility of health information across languages. The inclusion of a Catalan source also contributes to promoting the use of Catalan across NLP tasks, improving the accessibility and visibility of the Catalan language.
Texts contain no patient or personal data.
Discussion of Biases
No specific bias mitigation strategies were applied to this dataset. Inherent biases may exist within the data.
Known Limitations
- Small dataset (149 examples) – suited to targeted fine-tuning, analysis, or evaluation rather than general-purpose MT training.
- The post-edits correct the errors of a single system (SalamandraTA-7b-instruct), so the edit patterns may not transfer directly to other MT systems.
- The Catalan source is synthetic and was not human-verified; it may contain translation errors or unnatural phrasing.
Licensing Information
CC-BY 4.0
Additional Information
Dataset Curators
BSC-LT Machine Translation team & BSC Technologies for Biomedical Research Laboratory (TechBioLab)
Funding
The project IMPaCT-Data 2 - PDI (Exp. PMP24/00024) has been funded by the Institute of Health Carlos III (ISCIII), co-funded by the European Union – NextGeneration EU.
This work was funded by the Departament de la Vicepresidència i de Polítiques Digitals i Territori de la Generalitat de Catalunya within the framework of Projecte AINA.
Related Resources
Citation Information
No citation is available yet.