GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
nazmulhudanabil
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
nazmulhudanabil
Model Tree
unsloth/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Property | Details |
|---|---|
| Task | Banglish-to-Bangla transliteration/conversion |
| Base model | unsloth/Qwen2.5-1.5B-Instruct |
| Model architecture | Qwen2.5-1.5B-Instruct |
| Fine-tuning framework | Unsloth |
| Fine-tuning method | LoRA / QLoRA |
| Adapter format | PEFT |
| Model type | LoRA adapter |
| Primary output language | Bangla (Bengali script) |
| Intended use | Banglish text conversion |
Note: This repository contains the fine-tuned adapter, not the complete base model. The base model must also be loaded for inference.
Banglish is Bangla written using the Latin alphabet instead of the Bengali script.
For example:
ami valo achiআমি ভালো আছিThe spelling of Banglish is not standardized. The same Bangla sentence may be written in multiple ways using Latin characters. The model aims to learn the conversion patterns represented in its training data.
Input
text
ami valo achiExpected output
text
আমি ভালো আছিThis is an illustrative example, not a claim of measured model performance. Actual outputs may vary.
This model is intended for:
The model was fine-tuned using Unsloth and a parameter-efficient adaptation approach.
The training configuration included:
The adapter was saved separately from the original base model.
The model was trained on a dataset prepared for Banglish-to-Bangla conversion.
Dataset details:
Please fill in these details before publishing. Ensure that the dataset license permits your intended use and redistribution.
Formal evaluation results have not yet been reported.
For a reliable evaluation, assess the model on a held-out test dataset that was not used for training or model selection.
Recommended metrics:
| Metric | Purpose |
|---|---|
| Character Error Rate (CER) | Measures character-level conversion errors |
| Word Error Rate (WER) | Measures word-level errors |
| Exact Match Accuracy | Measures the proportion of predictions exactly matching the reference |
| Manual evaluation | Assesses spelling, meaning preservation, and naturalness |
Evaluation results will be added after testing.
Install the necessary libraries:
bash
pip install torch transformers peft acceleratepython
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel BASE_MODEL = "unsloth/Qwen2.5-1.5B-Instruct"ADAPTER_REPO = "nazmulhudanabil/banglish-to-bangla-qwen2.5-1.5b" tokenizer = AutoTokenizer.from_pretrained(ADAPTER_REPO) base_model = AutoModelForCausalLM.from_pretrained( BASE_MODEL, dtype=torch.float16,) model = PeftModel.from_pretrained( base_model, ADAPTER_REPO,) model.eval()Replace YOUR_USERNAME/banglish-to-bangla-qwen2.5-1.5b with the actual Hugging Face repository ID of this adapter.
python
text = "ami valo achi" prompt = f"""Convert the following Banglish text into natural Bangla script.Preserve the original meaning. Banglish: {text}Bangla:""" inputs = tokenizer(prompt, return_tensors="pt") with torch.inference_mode(): outputs = model.generate( **inputs, max_new_tokens=128, do_sample=False, ) new_tokens = outputs[0][inputs["input_ids"].shape[1]:]result = tokenizer.decode(new_tokens, skip_special_tokens=True) print(result)The inference prompt should match the prompt format used during training as closely as possible.
The adapter's license and usage rights depend on the applicable base-model license, dataset licenses, and the terms under which the adapter was created.
Verify the licenses before publishing. Do not claim that this adapter is Apache-2.0 licensed unless you have confirmed that this license is appropriate for the adapter and its training data.
Author: MD Nazmul Huda Nabil
Model repository: https://huggingface.co/nazmulhudanabil/banglish-to-bangla-qwen2.5-1.5b
Project source code: https://github.com/NazmulHudaNabil/banglish-to-bangla
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel BASE_MODEL = "unsloth/Qwen2.5-1.5B-Instruct"ADAPTER_REPO = "nazmulhudanabil/banglish-to-bangla-qwen2.5-1.5b" tokenizer = AutoTokenizer.from_pretrained(ADAPTER_REPO) base_model = AutoModelForCausalLM.from_pretrained( BASE_MODEL, dtype=torch.float16,) model = PeftModel.from_pretrained( base_model, ADAPTER_REPO,) model.eval()text = "ami valo achi" prompt = f"""Convert the following Banglish text into natural Bangla script.Preserve the original meaning. Banglish: {text}Bangla:""" inputs = tokenizer(prompt, return_tensors="pt") with torch.inference_mode(): outputs = model.generate( **inputs, max_new_tokens=128, do_sample=False, ) new_tokens = outputs[0][inputs["input_ids"].shape[1]:]result = tokenizer.decode(new_tokens, skip_special_tokens=True) print(result)