bhili-translate-bhb-mar

Bhili Translate

Machine Translation (MT) model for Bhili (भीली), specifically the Dehvali Bhili dialect, an Indo-Aryan language spoken by the Bhil community in western India.

This is a fine-tuned LoRA adapter on top of sarvamai/sarvam-translate, trained on ~40,000 Bhili-Marathi parallel sentences covering general domain text with additional focus on agricultural domain vocabulary and terminology.

Direction: Bhili (bhb) → Marathi (mr)

For the reverse direction (Marathi → Bhili), see ai4bharat/bhili-translate-mar-bhb.

Inference Recipe

The recommended inference recipe combines three things:

1. Beam search decoding (num_beams=4) instead of greedy.
2. A field-team-curated glossary (glossary.json, included in this repo) containing 6,400+ Bhili↔Marathi term pairs reviewed by native Bhili speakers.
3. A post-edit step (post_edit.py, included in this repo) that substitutes terms the model left in the source language with their target-language equivalents from the glossary.

⚠️ For the best results, we strongly recommend using the Translate Gemma system prompt when interacting with this model. Using the recommended system prompt helps ensure more consistent and accurate translation outputs.

Installation

pip install torch transformers peft accelerate huggingface_hub

Inference

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from huggingface_hub import hf_hub_download

# Load base model + adapter
base_model_name = "sarvamai/sarvam-translate"
adapter_name = "ai4bharat/bhili-translate-bhb-mar"

tokenizer = AutoTokenizer.from_pretrained(base_model_name)
model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_name)
model.eval()

# Load glossary + post-edit module from this repo
glossary_path = hf_hub_download(adapter_name, "glossary.json")
post_edit_path = hf_hub_download(adapter_name, "post_edit.py")

import importlib.util
spec = importlib.util.spec_from_file_location("post_edit", post_edit_path)
pe = importlib.util.module_from_spec(spec)
spec.loader.exec_module(pe)

glossary = pe.load_glossary(glossary_path)

# Translate
def translate(text: str) -> str:
    messages = [
        {
            "role": "system",
            "content": (
                "You are a professional Bhili (bhb) to Marathi (mr) translator. "
                "Your goal is to accurately convey the meaning and nuances of the "
                "original Bhili text while adhering to Marathi grammar, vocabulary, "
                "and cultural sensitivities. Produce only the Marathi translation, "
                "without any additional explanations or commentary. Please translate "
                "the following Bhili text into Marathi:"
            ),
        },
        {"role": "user", "content": text},
    ]
    prompt = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    inputs = tokenizer([prompt], return_tensors="pt").to(model.device)
    with torch.no_grad():
        output = model.generate(
            **inputs,
            max_new_tokens=256,
            do_sample=False,
            num_beams=4,
            early_stopping=True,
        )
    raw = tokenizer.decode(
        output[0][len(inputs.input_ids[0]):],
        skip_special_tokens=True,
    ).strip()
    # Apply glossary post-edit
    edited, _edits = pe.post_edit(
        text, raw, glossary["bhb2mar_sorted_keys"], glossary["bhb2mar"]
    )
    return edited

# Example
print(translate("खेडूतांन वेगीवेळ फवारणी केरनु जोजे हाय."))
print(translate("कोपास्यापापे गुलाबी बोंडअळीखातोर कॉपर ऑक्सिक्लोराईड फवारा."))

⚠️ Limitations

This model was predominantly trained on agricultural-domain data. As a result, performance on domains outside agriculture may vary and may not be representative of the performance observed on agricultural-domain.