---
license: cc
datasets:
- vector-institute/NMB-Plus-Named-Entities
base_model:
- distilbert/distilbert-base-uncased
pipeline_tag: token-classification
tags:
- ner
- bias_detection
model-index:
- name: nmb-plus-bias-ner-bert
  results:
  - task:
      type: named-entity-recognition
      name: Named Entity Recognition (NER)
    dataset:
      type: vector-institute/NMB-Plus-Named-Entities
      name: Biased Named Entities
    metrics:
    - type: precision
      value: 0.6405
    - type: recall
      value: 0.5589
    - type: f1
      value: 0.5922
language:
- en
---

# Model Overview

A fine-tuned DistilBERT model for Named Entity Recognition (NER) in bias detection.

## Model Details
We used `distilbert-base-uncased` and fine-tuned it on `vector-institute/NMB-Plus-Named-Entities` dataset.



## How to Get Started with the Model

```python
from transformers import AutoModelForTokenClassification, AutoTokenizer

model_name = "vector-institute/nmb-plus-bias-ner-bert"
tokenizer = AutoTokenizer.from_pretrained(model_name)

label_list = ["O", "B-BIAS", "I-BIAS"]
id2label = {i: label for i, label in enumerate(label_list)}
label2id = {label: i for i, label in enumerate(label_list)}

model = AutoModelForTokenClassification.from_pretrained(
    model_name,
    id2label=id2label, 
    label2id=label2id
)


ner_pipeline = pipeline("ner", model=model, tokenizer=tokenizer)

text = "Fox News reported that Joe Biden met with CNN executives."
predictions = ner_pipeline(text)
print(predictions)

```


## Training Hyperparameters

- **Training regime:**
Here's the training arguments we used:

```python
training_args = TrainingArguments(
    learning_rate=2e-5,
    per_device_train_batch_size=64,
    per_device_eval_batch_size=32,
    num_train_epochs=10,
    weight_decay=0.01,
    eval_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
    output_dir="./results",
    logging_dir="./logs",
    logging_steps=50,
    group_by_length=True,
)
```


## Evaluation

We split the data to train(80%), validation(10%) and test(10%) sets.


### Results
We used common classification metrics:
- precision
- recall
- f1-score

#### Overall Results:
| Metric         | Precision | Recall | F1-Score | Support |
|---------------|-----------|--------|----------|---------|
| **Macro Avg** | 0.6405    | 0.5589 | 0.5922   | 48710   |
| **Weighted Avg** | 0.9330 | 0.9418 | 0.9366   | 48710   |

#### Per-class Results:

| Label    | Precision | Recall | F1-Score | Support |
|----------|-----------|--------|----------|---------|
| **O**    | 0.9615    | 0.9792 | 0.9703   | 45921   |
| **B-BIAS** | 0.5314  | 0.4183 | 0.4681   | 930     |
| **I-BIAS** | 0.4286  | 0.2792 | 0.3381   | 1859    |


## Environmental Impact

Total energy consumption for fine-tuning is 0.032804 kWh

**Local CO2 Emission:**  Approximately 3.12 grams of CO₂ equivalent.

## License
CC BY 4.0 (Creative Commons Attribution 4.0): Allows sharing and adaptation with proper credit.