README.md

---
license: cc-by-sa-4.0
dataset_info:
  features:
  - name: id
    dtype: string
  - name: lp
    dtype: string
  - name: source
    dtype: string
  - name: reference
    dtype: string
  - name: gold_clean
    dtype: string
  - name: translation
    dtype: string
  - name: src_lang
    dtype: string
  - name: tgt_lang
    dtype: string
  - name: label
    dtype: int64
  - name: noise_category
    dtype: string
  - name: noise_pattern
    dtype: string
  - name: split
    dtype: string
  splits:
  - name: train
    num_bytes: 9755128
    num_examples: 8800
  download_size: 3139078
  dataset_size: 9755128
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
language:
- en
- ar
- zh
- he
- fr
- ru
- it
- de
- ko
- cs
- pl
- ta
- km
---


## Dataset Summary

Name: TransClean — A benchmark for detecting and extracting clean translations from LLM outputs.

Description: This dataset contains **synthetic** examples used to evaluate methods for detecting and extracting clean translations produced by large language models. It supports research on translation quality estimation, extraction, and filtering of clean model outputs. More details can be found in our [paper](https://arxiv.org/abs/2609.11399) at WMT26.

Data columns: `translation` is the output generated by LLMs, which might contain noise if directly used for evaluation. Hence, it is the input of noise extraction method to get clean translation. `gold_clean` is the gold clean translation that can be used for computing accuracy metric. `noise_pattern` indicates what kind of noise the `translation` contains. When `noise_pattern` is `none`, it mean it's already a clean translation. When it is `off_topic`, then there is no clean translation to be extracted, hence **an empty string**.

Included files: `synthetic.jsonl` for the synthetic subset only. Please contact us if you need the curated subset `curated_noisy_1100_with_silver.jsonl`.