InLegalTrans-Finetuned-JUSTNLP2025

This model is a domain-adapted legal translation system finetuned on top of law-ai/InLegalTrans-En2Indic-1B for English ↔ Hindi legal text translation. It was trained for the JUSTNLP 2025 Legal Machine Translation Task using high-quality legal and supervised MT datasets.

---

Model Overview

• Base Model: law-ai/InLegalTrans-En2Indic-1B

• Languages: English ↔ Hindi

• Task: Translation (Legal Domain)

• Datasets:

  • helloboyn/IJCNLP-JustNLP-LMT
  • helloboyn/WMT25-TS

• Metrics: BLEU, ROUGE, BERTScore

• Pipeline Tag: translation

• Tags: legal

---

What This Model Does

• Translates legal documents, including court orders, case summaries, statutory text, and formal legal writing.
• Preserves legal terminology, obligations, conditions, and negation.
• Designed for use in research, policy digitisation, and legal NLP applications.

---

Usage

English → Hindi

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model = "drjk16/InLegalTrans-Finetuned-JUSTNLP2025"
tok = AutoTokenizer.from_pretrained(model)
m = AutoModelForSeq2SeqLM.from_pretrained(model)

text = "The accused is ordered to appear before the court."

inputs = tok(text, return_tensors="pt")
out = m.generate(**inputs, max_new_tokens=256)
print(tok.decode(out[0], skip_special_tokens=True))

Hindi → English

text = "अभियुक्त को न्यायालय के समक्ष उपस्थित होने का आदेश दिया जाता है।"

inputs = tok(text, return_tensors="pt")
out = m.generate(**inputs, max_new_tokens=256)
print(tok.decode(out[0], skip_special_tokens=True))

---

Training Summary

• Finetuned using AdamW, low learning rate, and long-sequence training.
• Mix of legal-domain parallel corpora and structured translation datasets.
• Evaluated on held-out legal text using BLEU, ROUGE, and BERTScore.

---

Limitations

• Not a substitute for certified human translation.
• May struggle with noisy OCR, extremely long judgments, or niche legal subdomains.
• Should be used with human verification for official or court-critical tasks.

---