---
language: bn
license: mit
tags:
  - text-classification
  - bangla
  - bengali
  - ai-detection
  - human-vs-ai
  - fine-tuned
base_model: csebuetnlp/banglabert

metrics:
  - accuracy
  - f1
pipeline_tag: text-classification
---

# Bangla AI vs Human Writing Detector — BanglaBERT

Fine-tuned [BanglaBERT](https://huggingface.co/csebuetnlp/banglabert) for **binary classification of Bangla text as AI-generated or human-written**. Developed as part of the *Onneshon* thesis project at Bangladesh University of Professionals (BUP).

Given the full text of a Bangla resume, the model predicts whether it was written by a human or generated by an AI system, along with a confidence score.

---

## Model Details

| Property | Value |
|---|---|
| Base model | `csebuetnlp/banglabert` |
| Architecture | ElectraForSequenceClassification (ELECTRA-based BERT) |
| Parameters | 12 layers, 768 hidden dim, 12 attention heads, vocab 32,000 |
| Language | Bengali (bn) |
| Task | Binary sequence classification |
| Labels | `0` → Human, `1` → AI |
| Max input tokens | 512 |
| Training data | 70 resumes (35 AI + 35 Human) |
| Validation data | ~15 resumes (stratified) |
| Test data | ~15 resumes (stratified) |
| Epochs | 5 (with EarlyStoppingCallback, patience=2) |
| Learning rate | 2e-5 |
| Batch size | 8 |
| Warmup ratio | 0.1 |
| Weight decay | 0.01 |
| Mixed precision | fp16 |
| Best model metric | F1 (binary) |

---

## Dataset

Trained on **Onneshon** — an original Bangla resume dataset:

- **Human resumes:** 50 resumes written by Bangladeshi professionals (resume_51–resume_100)
- **AI resumes:** 50 resumes generated by AI systems (resume_1–resume_50)
- **Total:** 100 resumes, split 70/15/15 (train/val/test), stratified

Published on Mendeley Data: [DOI: 10.17632/4md7bx6fd7.1](https://doi.org/10.17632/4md7bx6fd7.1)

---

## Labels

| ID | Label | Description |
|---|---|---|
| 0 | Human | Resume written by a human |
| 1 | AI | Resume generated by an AI system |

---

## Usage

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("your-username/bangla-ai-detector")
model     = AutoModelForSequenceClassification.from_pretrained("your-username/bangla-ai-detector")
model.eval()

def predict(text):
    inputs  = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        logits = model(**inputs).logits
    probs  = torch.softmax(logits, dim=1)[0]
    label  = model.config.id2label[logits.argmax().item()]
    return {
        "label":        label,
        "confidence":   f"{probs.max().item()*100:.1f}%",
        "P(Human)":     f"{probs[0].item()*100:.1f}%",
        "P(AI)":        f"{probs[1].item()*100:.1f}%",
    }

# Example
resume_text = "আমি একজন অভিজ্ঞ সফ্টওয়্যার ইঞ্জিনিয়ার। গত পাঁচ বছর ধরে জাভা এবং স্প্রিং বুট দিয়ে কাজ করছি।"
print(predict(resume_text))
# Output: {'label': 'Human', 'confidence': '87.3%', 'P(Human)': '87.3%', 'P(AI)': '12.7%'}
```

---

## Preprocessing

Before passing text to the model, strip any annotation tags if present (these are specific to the Onneshon dataset format):

```python
import re

def clean_resume(text):
    text = re.sub(r'\[Info_Start\].*?\[Info_End\]', '', text, flags=re.DOTALL)
    text = re.sub(r'\[(Objective|Experience|Expericence|Education|Skill|section)\]', '', text)
    text = re.sub(r'\s+', ' ', text).strip()
    return text
```

---

## Limitations

- **Small dataset:** Only 100 resumes (70 training). Binary accuracy estimates have high variance — results should be interpreted with caution.
- **Domain-specific:** Trained exclusively on Bangla resumes. Performance on other Bangla document types (news, social media, etc.) is untested.
- **AI source unknown:** The AI resumes were generated by a specific AI system. The model may not generalize to resumes generated by different LLMs not seen during training.
- **Token limit:** Resumes longer than 512 tokens are truncated. Very long resumes may lose tail content.
- **Binary only:** Cannot distinguish *which* AI system generated the text — only Human vs AI.

---

## Citation

```bibtex
@misc{onneshon2026,
  title     = {Onneshon: A Bangla Resume NLP Dataset},
  author    = {Tanvir and Shruti Khisa and Shaira Akther Diba and Fazli Rabbi Noor},
  year      = {2026},
  doi       = {10.17632/4md7bx6fd7.1},
  publisher = {Mendeley Data}
}

@inproceedings{bhattacharjee-etal-2022-banglabert,
  title     = {BanglaBERT: Language Model Pretraining and Benchmarks for Low-Resource Language Understanding Evaluation in Bangla},
  author    = {Bhattacharjee, Abhik and Hasan, Tahmid and Ahmad, Wasi and Mubasshir, Kazi Samin and Islam, Md Saiful and Iqbal, Anindya and Rahman, M. Sohel and Shahriyar, Rifat},
  booktitle = {Findings of the Association for Computational Linguistics: NAACL 2022},
  year      = {2022},
  pages     = {1318--1327}
}
```

---

## Project

Part of the **Onneshon** thesis project — a Bangla NLP pipeline for resume processing.

- Dataset: [Mendeley Data DOI: 10.17632/4md7bx6fd7.1](https://doi.org/10.17632/4md7bx6fd7.1)
- Institution: Bangladesh University of Professionals (BUP), Dhaka, Bangladesh
- Supervisor: Rumana Yasmin, Lecturer, CSE Department, BUP