# WMT 2023 Original Data

The file `wmttest2023.eng` is the original source dataset, obtained by concatenating various sources from WMT 2023 campaigns, including the general task, all challenge sets and the biomedical translation task.

Folders contain machine translations to Italian (`ita`) and Dutch (`nld`) for every segment produced by [NLLB-600M](https://huggingface.co/facebook/nllb-200-distilled-600M) and [NLLB-3B](https://huggingface.co/facebook/nllb-200-3.3B), using beam search with `n_beams = 5`. Every translation in both target languages is scored using both [`XCOMET-XL`](https://huggingface.co/Unbabel/XCOMET-XL) and [`XCOMET-XXL`](https://huggingface.co/Unbabel/XCOMET-XXL).

Only the NLLB-3B translations and the XCOMET-XXL error spans are used for the QE4PE study.

## Libraries versions

```shell
transformers == 4.22.0
unbabel-comet == 2.2.0
```