locailabs/legislation-gov-uk-en-cy

UK Legislation — Welsh–English SFT Dataset

Processed instruction-tuning dataset derived from techiaith/legislation-gov-uk_en-cy, a Welsh–English parallel translation memory published by the Bangor University Language Technologies Unit (Techiaith). Formatted for supervised fine-tuning (SFT) of language models.

Source

Field	Value
Source dataset	techiaith/legislation-gov-uk_en-cy
Domain	UK statute law (legislation.gov.uk)
Raw pairs	64,726
Processed examples	17,886
Licence	Open Government Licence v3.0 (OGL)

All translations were produced by professional translators working within Welsh public sector institutions.

Format

The dataset uses the messages chat format. Two example types are included.

Single-turn (~70% of examples):

{
  "messages": [
    {
      "role": "user",
      "content": "Translate the following English text into Welsh:\n\nThe application must be submitted before the deadline."
    },
    {
      "role": "assistant",
      "content": "Rhaid cyflwyno'r cais cyn y dyddiad cau."
    }
  ],
  "source_dataset": "techiaith/legislation-gov-uk_en-cy"
}

Multi-turn (~30% of examples):

{
  "messages": [
    {
      "role": "user",
      "content": "I'd like you to translate a series of English sentences into Welsh. I'll give you one sentence at a time.\n\nThe committee has considered this matter in detail."
    },
    {
      "role": "assistant",
      "content": "Mae'r pwyllgor wedi ystyried y mater hwn yn fanwl."
    },
    {
      "role": "user",
      "content": "An amendment was proposed by the Member for Ynys Môn."
    },
    {
      "role": "assistant",
      "content": "Cynigiwyd gwelliant gan yr Aelod dros Fôn."
    }
  ],
  "source_dataset": "techiaith/legislation-gov-uk_en-cy"
}

Fields:

• messages: Translation task in chat format
• source_dataset: HuggingFace ID of the originating source corpus

The dataset is balanced: ~50% English→Welsh and ~50% Welsh→English translations. Instruction prompts are drawn from a diverse pool of 21 template phrasings in both English and Welsh to reduce overfitting to a single prompt pattern.

Curation Pipeline

Raw pairs from the source dataset were processed through the following stages:

1. Length filter — pairs where either side is fewer than 20 characters are removed
2. Artefact filter — pairs containing URLs, emoji, bullet/list markers, or excessive repetition are removed
3. Exact deduplication — normalised string deduplication (NFC, lowercased, whitespace-collapsed)
4. MinHash deduplication — 1-gram MinHash LSH (128 permutations, Jaccard threshold 0.9) to remove near-identical surface-form variants
5. Semantic deduplication — embedding-based deduplication via SemHash (minishlab/potion-multilingual-8M, cosine threshold 0.85) to remove semantically equivalent pairs with different surface forms
6. Instruction formatting — conversion to chat format with template diversity and multi-turn conversation grouping

Limitations

All source data is drawn from formal institutional domains. The dataset covers formal Welsh well but underrepresents colloquial, spoken, and informal registers. Source translation memories may contain a small number of misaligned pairs that are not detectable without a dedicated quality scorer.

Citation

If you use this dataset, please cite the original Techiaith source:

@misc{techiaith_legislation_gov_uk,
  author       = {{Bangor University Language Technologies Unit (Techiaith)}},
  title        = {legislation.gov.uk Welsh–English Translation Memory},
  year         = {2023},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/techiaith/legislation-gov-uk_en-cy}},
}

Acknowledgements

• Techiaith (Bangor University) for producing and releasing the source translation memories
• Locai Labs for the curation pipeline