LegalisTranslate is a fine-tuned Neural Machine Translation model specialized for the Indian Legal Domain (Vidhik Hindi). It is built by fine-tuning Meta's NLLB-200-Distilled-600M model using LoRA (Low-Rank Adaptation) on a filtered subset of the AI4Bharat Samanantar dataset.

The model specifically addresses the issue of general-purpose translators failing to capture specific legal terminology (e.g., translating "Act" as "Work" instead of "Adhiniyam", or "Magistrate" as "Judge").

Model Details

Base Model: facebook/nllb-200-distilled-600M
Task: Translation (English eng_Latn → Hindi hin_Deva)
Fine-Tuning Method: LoRA (PEFT)
Training Data: ~50,000 sentence pairs from AI4Bharat Samanantar (Governance & PIB subsets), filtered for legal keywords.
Language: English to Hindi (Devanagari)

Intended Use

This model is intended for:

Translating formal legal documents (Acts, Bills, Court Orders).
Translating administrative government text.
Research in Low-Resource Domain Adaptation.

⚠️ Limitation: This model is a research prototype trained for 3 epochs on a T4 GPU. It should not be used as a substitute for a certified human translator or lawyer for legal filings.

How to Use

Since this model uses PEFT (LoRA), you must load the base model first and then load the adapter.

python

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from peft import PeftModel

# 1. Load Base Model
base_model_id = "facebook/nllb-200-distilled-600M"
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForSeq2SeqLM.from_pretrained(base_model_id)

# 2. Load the Fine-Tuned Adapter
# Replace 'your-username/legalistranslate' with your actual Hub path or local folder
adapter_model_id = "your-username/legalistranslate" 
model = PeftModel.from_pretrained(base_model, adapter_model_id)
model.to("cuda" if torch.cuda.is_available() else "cpu")

# 3. Translation Function
def translate_legal(text):
    tokenizer.src_lang = "eng_Latn"
    inputs = tokenizer(text, return_tensors="pt").to(model.device)
    
    # Force output to Hindi
    forced_bos_token_id = tokenizer.convert_tokens_to_ids("hin_Deva")
    
    with torch.no_grad():
        outputs = model.generate(
            **inputs, 
            forced_bos_token_id=forced_bos_token_id, 
            max_length=128
        )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# 4. Test
text = "The accused was acquitted due to lack of evidence."
print(translate_legal(text))
# Output: "अभियुक्त को सबूतों की कमी के कारण बरी कर दिया गया।"

ranveer0323/legalis-nllb-hindi-v1

LegalisTranslate: English to Hindi Legal Domain Translator

Model Details

Intended Use

How to Use

ranveer0323/
legalis-nllb-hindi-v1