IJCNLP-JustNLP-LMT Dataset

1. Overview

This dataset is prepared for fine-tuning and evaluation of large language models in the context of legal-domain machine translation.
It follows the data design principles of the WMT25 Legal Benchmark and is curated to support high-quality parallel data for research and benchmarking.

The dataset consists of 50,000 English–Hindi parallel sentence pairs, carefully filtered and normalized to ensure alignment quality and domain relevance.

---

2. Domain and Purpose

• Domain: Legal
• Primary Use:
  • Fine-tuning large language models
  • Evaluation of legal-domain machine translation systems
• Target Applications:
  • Legal document translation
  • Government and administrative translation
  • Legal AI applications

---

3. Data Source and Preparation

The dataset draws inspiration from the WMT25 Legal Benchmark, with a strong emphasis on:

• High-quality parallel alignment
• Domain consistency
• Noise reduction through systematic filtering

A series of pre-processing and filtering procedures were applied to improve overall data quality.

---

4. Filtering and Pre-processing Strategy

Sentence Length Constraints

To ensure model stability and training efficiency:

• Minimum length: 5 words
• Maximum length: 70 words
• Very short sentences were removed to avoid structural incompleteness.
• Very long sentences were excluded to reduce training instability.

Word Count Balance

To preserve alignment quality:

• Sentence pairs were retained only if the difference in word count between source and target fell within a controlled range.
• This prevents severe length mismatches that often indicate poor translations.

---

5. Dataset Statistics

Split-wise Distribution

Split	Language	Sentences	Words
Train	English	50,000	1,463,914
	Hindi	50,000	1,377,884
Validation	English	5,000	145,036
	Hindi	5,000	136,450
Test	English	5,000	134,407
	Hindi	5,000	133,471

---

6. Intended Research Use

This dataset is suitable for:

• Legal-domain machine translation
• Fine-tuning multilingual and bilingual LLMs
• Comparative evaluation of MT systems
• Synthetic data generation and pivot-based translation studies

---

7. Loading the Dataset (Python)

from datasets import load_dataset

dataset = load_dataset("helloboyn/IJCNLP-JustNLP-LMT")

10. Loading the Dataset (Python)

@inproceedings{singh-etal-2025-findings,
    title = "Findings of the {JUST}-{NLP} 2025 Shared Task on {E}nglish-to-{H}indi Legal Machine Translation",
    author = "Singh, Kshetrimayum Boynao  and
      Kumar, Sandeep  and
      Datta, Debtanu  and
      Joshi, Abhinav  and
      Mishra, Shivani  and
      Paul, Shounak  and
      Goyal, Pawan  and
      Jain, Sarika  and
      Ghosh, Saptarshi  and
      Modi, Ashutosh  and
      Ekbal, Asif",
    editor = "Modi, Ashutosh  and
      Ghosh, Saptarshi  and
      Ekbal, Asif  and
      Goyal, Pawan  and
      Jain, Sarika  and
      Joshi, Abhinav  and
      Mishra, Shivani  and
      Datta, Debtanu  and
      Paul, Shounak  and
      Singh, Kshetrimayum Boynao  and
      Kumar, Sandeep",
    booktitle = "Proceedings of the 1st Workshop on NLP for Empowering Justice (JUST-NLP 2025)",
    month = dec,
    year = "2025",
    address = "Mumbai, India",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.justnlp-main.3/",
    pages = "12--17",
    ISBN = "979-8-89176-312-8",
    abstract = "This paper provides an overview of the Shared Task on Legal Machine Translation (L-MT), organized as part of the JUST-NLP 2025 Workshop at IJCNLP-AACL 2025, aimed at improving the translation of legal texts, a domain where precision, structural faithfulness, and terminology preservation are essential. The training set comprises 50,000 sentences, with 5,000 sentences each for the validation and test sets. The submissions employed strategies such as: domain-adaptive fine-tuning of multilingual models, QLoRA-based parameter-efficient adaptation, curriculum-guided supervised training, reinforcement learning with verifiable MT metrics, and from-scratch Transformer training. The systems are evaluated based on BLEU, METEOR, TER, chrF++, BERTScore, and COMET metrics. We also combine the scores of these metrics to give an average score (AutoRank). The top-performing system is based on a fine-tuned distilled NLLB-200 model and achieved the highest AutoRank score of 72.1. Domain adaptation consistently yielded substantial improvements over baseline models, and precision-focused rewards proved especially effective for the legal MT. The findings also highlight that large multilingual Transformers can deliver accurate and reliable English-to-Hindi legal translations when carefully fine-tuned on legal data, advancing the broader goal of improving access to justice in multilingual settings."
}