This is a LoRA (Low-Rank Adaptation) fine-tuned adapter built on top of ogx786/urdu-roman-transliterator-tiny-aya. It was trained to convert Urdu-script sentences into their Roman Urdu equivalents using an instruction-style prompt format.

Developed by: [hina]
Model type: Causal Language Model with LoRA adapter
Language(s): Urdu (input), Roman Urdu (output)
License: Apache 2.0 (update if different)
Finetuned from model: ogx786/urdu-roman-transliterator-tiny-aya

Uses

Direct Use

Use this model to convert Urdu script text into Roman Urdu, useful for:

Messaging apps and keyboards where Urdu script isn't easily typed
Accessibility tools
Language learning aids
Search/indexing systems that need Romanized text

Out-of-Scope Use

Not intended for general-purpose text generation, translation into English, or use with languages other than Urdu.

How to Get Started with the Model

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "ogx786/urdu-roman-transliterator-tiny-aya"
adapter_repo = "hina6198/urdu-roman-transliterator-lora"  # update with your username

tokenizer = AutoTokenizer.from_pretrained(adapter_repo)
base_model = AutoModelForCausalLM.from_pretrained(base_model_name, device_map="auto")
model = PeftModel.from_pretrained(base_model, adapter_repo)
model.eval()

def transliterate(urdu_text):
    prompt = f"""### Instruction:
Transliterate Urdu to Roman Urdu.

### Input:
{urdu_text}

### Response:
"""
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=100, do_sample=False)
    result = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return result.split("### Response:")[-1].strip()

print(transliterate("تم کیسے ہو؟"))

Training Details

Training Data

[Describe your dataset here — size, source, e.g. "~X,000 Urdu–Roman Urdu sentence pairs collected/curated from ..."]

Training Procedure

Method: LoRA fine-tuning (PEFT)
LoRA config: r=16, alpha=32, target modules: q_proj, v_proj, dropout=0.05
Epochs: 3
Batch size: 1 (effective batch size 4 via gradient accumulation)
Learning rate: 2e-4
Precision: fp16

Training Hyperparameters

Training regime: fp16 mixed precision
Trainable parameters: ~3.8M (0.11% of total 3.35B parameters)

Limitations

May not generalize well to dialectal variations, slang, or code-mixed (Urdu-English) text.
Trained on a relatively small dataset; performance on rare words or names may be inconsistent.
Base model is small