---
language:
- bn
license: apache-2.0
tags:
- token-classification
- ner
- bangla
- bengali
- electra
- transformers
datasets:
- custom
metrics:
- f1
- precision
- recall
- accuracy
base_model: csebuetnlp/banglabert
pipeline_tag: token-classification
model-index:
- name: bangla-ner
  results:
  - task:
      type: token-classification
      name: Named Entity Recognition
    metrics:
    - type: f1
      value: 0.7493
    - type: precision
      value: 0.7582
    - type: recall
      value: 0.7406
    - type: accuracy
      value: 0.9341
---

# Bangla NER — Named Entity Recognition for Bengali

A fine-tuned token classification model for **Bengali (Bangla) Named Entity Recognition** using the BIO tagging scheme.
Built on top of [csebuetnlp/banglabert](https://huggingface.co/csebuetnlp/banglabert) (ELECTRA-based).

---

## Entity Types

| Tag | Description | Example |
|---|---|---|
| `PER` | Person names | একেএম শহীদুল হক |
| `LOC` | Locations, cities, countries | বাংলাদেশ, ঢাকা |
| `ORG` | Organizations, companies | টুইটার, রিয়াল মাদ্রিদ |
| `POL` | Political entities / parties | আওয়ামী লীগ |
| `DATE` | Calendar dates | সোমবার, ২০২৪ সালে |
| `TIME` | Times of day | সকাল ৮টায় |
| `EVENT` | Named events | রোহিঙ্গা সঙ্কট |
| `CRIME` | Crime-related entities | হত্যা মামলা |
| `TITLE` | Titles, designations | মহাপরিদর্শক |
| `NUM` | Numbers, quantities | ৯৩ শতাংশ |
| `SYMBOL` | Symbols, currencies | ৳, % |
| `CONSTITUENCY` | Electoral constituencies | ঢাকা-১৮ |
| `INST` | Institutions | তথ্য অধিদপ্তর |

All tags follow **BIO** format: `B-` (beginning), `I-` (inside), `O` (outside).

---

## Training Details

| Parameter | Value |
|---|---|
| Base model | `csebuetnlp/banglabert` |
| Architecture | ELECTRA (discriminator) |
| Task | Token Classification (NER) |
| Dataset size | 22,144 sentences |
| Train split | 85% (18,822) |
| Validation split | 7.5% (1,661) |
| Test split | 7.5% (1,661) |
| Max sequence length | 256 tokens |
| Batch size | 16 |
| Epochs | 8 (early stopping, patience=2) |
| Best epoch | 7 |
| Learning rate | 2e-5 |
| LR scheduler | Linear with warmup |
| Warmup steps | 10% of total steps |
| Weight decay | 0.01 |
| Optimizer | AdamW |
| Mixed precision | fp16 |
| Framework | PyTorch + HuggingFace Transformers |
| Hardware | NVIDIA GeForce RTX 4070 Ti SUPER (16 GB) |

---

## Test Set Results (Overall)

| Metric | Score |
|---|---|
| **F1** | **74.93%** |
| **Precision** | **75.82%** |
| **Recall** | **74.06%** |
| **Token Accuracy** | **93.41%** |

---

## Per-Entity Results (Test Set)

| Entity | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| CONSTITUENCY | 0.8333 | 0.7500 | 0.7895 | 20 |
| CRIME | 0.9489 | 0.9489 | 0.9489 | 137 |
| DATE | 0.7730 | 0.7552 | 0.7640 | 478 |
| EVENT | 0.6827 | 0.6514 | 0.6667 | 109 |
| INST | 0.7119 | 0.7636 | 0.7368 | 55 |
| LOC | 0.7451 | 0.7245 | 0.7347 | 795 |
| NUM | 0.6949 | 0.8913 | 0.7810 | 46 |
| ORG | 0.5617 | 0.5686 | 0.5652 | 408 |
| PER | 0.7654 | 0.7260 | 0.7452 | 719 |
| POL | 0.8182 | 0.8333 | 0.8257 | 54 |
| SYMBOL | 1.0000 | 0.8750 | 0.9333 | 8 |
| TIME | 0.9839 | 0.8472 | 0.9104 | 144 |
| TITLE | 0.9532 | 0.9645 | 0.9588 | 169 |
| **micro avg** | **0.7582** | **0.7406** | **0.7493** | 3142 |
| **macro avg** | 0.8056 | 0.7923 | 0.7969 | 3142 |

---

## Usage

### With pipeline (recommended)

```python
from transformers import pipeline

ner = pipeline(
    "ner",
    model="arafatfahim/BanglaTag",
    aggregation_strategy="simple",
)

text = "একেএম শহীদুল হক বাংলাদেশে কক্সবাজার এলাকায় সোমবার সংবাদ সম্মেলন করেন"
print(ner(text))
```

### Manual inference

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

model_name = "arafatfahim/BanglaTag"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)
model.eval()

tokens = ["একেএম", "শহীদুল", "হক", "বাংলাদেশে", "এসেছেন"]
inputs = tokenizer(tokens, is_split_into_words=True, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

predictions = logits.argmax(-1)[0]
word_ids = inputs.word_ids()

prev = None
for word_id, pred_id in zip(word_ids, predictions):
    if word_id is None or word_id == prev:
        continue
    print(f"{tokens[word_id]:20s} → {model.config.id2label[pred_id.item()]}")
    prev = word_id
```

---

## Citation

If you use this model, please cite:

```bibtex
@misc{bangla-ner-2026,
  title  = {Bangla NER: Fine-tuned BanglaBERT for Bengali Named Entity Recognition},
  year   = {2026},
  url    = {https://huggingface.co/arafatfahim/BanglaTag}
}
```