SynCED-EnDe — Pipeline, Raw Data & Scripts (GitHub Companion)

This repository provides pipeline scripts and raw/intermediate files for building the SynCED-EnDe dataset.

The clean, benchmark-ready release (train silver + eval gold) is on Hugging Face.
This repo is for reproducibility and transparency.

---

📂 Repository Structure

scripts/                    # Core dataset creation pipeline
  inject.py                 # inject controlled errors
  reinject.py               # retry injection for missing cases
  skim.py                   # preview dataset (sanity check)
  block_evalrows.py         # filter out scraped evaluation rows from train set 
  final.py                  # assemble final dataset TSVs
  data_check.py             # sanity checks (counts, distributions, label balance)
  data_scrape.py            # scrape GOV.UK + Stack Exchange

data/
  train/
    synced_ende_train_silver.tsv     # final train split
    raw/                             # raw/intermediate training data
      ced_final_injected_multi_class.tsv
      error_injected_rows_with_correctMT.tsv
      rows_for_error_injection.tsv
      sources_2024_2025.tsv

  eval/
    synced_ende_eval_gold.tsv        # final eval split
    eval_judged_quantified_annotated.tsv
    raw/                             # raw/intermediate eval data
      ced_final_injected_multi_class.tsv
      error_injected_rows_with_correctMT.tsv
      rows_for_error_injection.tsv
      sources_2024_2025.tsv

---

🛠️ Pipeline Usage

Run the scripts in order:

1. Scrape Data

   python scripts/data_scrape.py

2. Inject errors (first pass)

   python scripts/inject.py

3. Re-inject (fix empty rows, optional second pass)

   python scripts/reinject.py

4. Preview dataset (skim head rows)

   python scripts/skim.py

5. Block evaluation rows (avoid leakage)

   python scripts/block_evalrows.py

6. Judge & quantify translations (LLM-based)

   python scripts/judge_quantify.py
   

7. Assemble final dataset

   python scripts/final.py

   This produces:

   • synced_ende_train_silver.tsv
   • synced_ende_eval_gold.tsv
   • judged_quantified_annotated.tsv

---

📊 Raw / Intermediate Data

Available in data/:

• ced_final_injected_multi_class.tsv
• error_injected_rows_with_correctMT.tsv
• rows_for_error_injection.tsv
• sources_2024_2025.tsv

⚠️ Note: These files may contain earlier/unverified labels.
They are not for benchmarking — use only for reproducing pipeline steps.

---

📜 License

This dataset and accompanying scripts are licensed under:
Creative Commons Attribution- 4.0 (CC-BY 4.0)
Attribution required.

---

📖 Citation

@article{chopra2025syncedende,
  title={SynCED-EnDe 2025: A Synthetic and Curated English - German Dataset for Critical Error Detection in Machine Translation},
  author={Chopra, M. and others},
  journal={Hugging Face Papers},
  year={2025},
  url={https://huggingface.co/papers/2510.05144}
}