---
datasets:
- stanfordnlp/imdb
language:
- en
metrics:
- accuracy
- precision
- recall
- f1
base_model:
- facebook/bart-base
- google-bert/bert-base-uncased
- EleutherAI/gpt-neo-2.7B
pipeline_tag: text-classification
license: apache-2.0
---

# 📝 Model Card: ensemble-majority-voting-imdb

## 🔍 Introduction
The `wakaflocka17/ensemble-majority-voting-imdb` model is a majority-voting ensemble of three fine-tuned sentiment classifiers (`bert-imdb-finetuned`, `bart-imdb-finetuned`, `gptneo-imdb-finetuned`) on the IMDb dataset. Each model votes on the sentiment label and the ensemble returns the label with the most votes, improving overall accuracy.

## 📊 Evaluation Metrics
| Metric    | Value   |
|-----------|---------|
| Accuracy  | 0.93296 |
| Precision | 0.9559  |
| Recall    | 0.9078  |
| F1-score  | 0.9312  |

## ⚙️ Training Parameters
| Parameter             | Values                                           |
|-----------------------|--------------------------------------------------|
| Models in ensemble    | `bert_base_uncased`, `bart_base`, `gpt_neo_2_7b` |
| Repo for ensemble     | `models/ensemble_majority_voting`                |
| Batch size (eval)     | 64                                               |

## 🚀 Example of use in Colab

#### Installing dependencies
```bash
!pip install --upgrade transformers huggingface_hub
```
#### (Optional) Authentication for private models
```python
from huggingface_hub import login
login(token="hf_yourhftoken")
```
#### Loading models and creating ensemble pipeline
```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TextClassificationPipeline
from collections import Counter

# List of fine-tuned model repo IDs
model_ids = [
    "wakaflocka17/bert-imdb-finetuned",
    "wakaflocka17/bart-imdb-finetuned",
    "wakaflocka17/gptneo-imdb-finetuned"
]
```
#### Load pipelines
```python
pipelines = []
for repo_id in model_ids:
    tokenizer = AutoTokenizer.from_pretrained(repo_id)
    model = AutoModelForSequenceClassification.from_pretrained(repo_id)
    model.config.id2label = {0: 'NEGATIVE', 1: 'POSITIVE'}
    pipelines.append(TextClassificationPipeline(model=model, tokenizer=tokenizer, return_all_scores=False))
```
#### Ensemble prediction function
```python
def ensemble_predict(text):
    votes = []
    # Collect each model's vote along with its name
    for model_id, pipe in zip(model_ids, pipelines):
        label = pipe(text)[0]['label']
        votes.append({
            "model": model_id,   # or model_id.split("/")[-1] for just the short name
            "label": label
        })
    # Determine majority label
    majority_label = Counter([v["label"] for v in votes]).most_common(1)[0][0]
    return {
        "ensemble_label": majority_label,
        "individual_votes": votes
    }
```
#### Inference on a text example
```python
testo = "This movie was absolutely fantastic—wonderful performances and a gripping story!"
result = ensemble_predict(testo)
print(result)
# Example output:
# {
#   'ensemble_label': 'POSITIVE',
#   'individual_votes': [
#       {'model': 'wakaflocka17/bert-imdb-finetuned', 'label': 'POSITIVE'},
#       {'model': 'wakaflocka17/bart-imdb-finetuned', 'label': 'NEGATIVE'},
#       {'model': 'wakaflocka17/gptneo-imdb-finetuned', 'label': 'POSITIVE'}
#   ]
# }
```
## 📖 How to cite
If you use this model in your work, you can cite it as:
```latex
@misc{Sentiment-Project,
  author       = {Francesco Congiu},
  title        = {Sentiment Analysis with Pretrained, Fine-tuned and Ensemble Transformer Models},
  howpublished = {\url{https://github.com/wakaflocka17/DLA_LLMSANALYSIS}},
  year         = {2025}
}
```
## 🔗 Reference Repository
> All the file structure and script examples can be found at:
> https://github.com/wakaflocka17/DLA_LLMSANALYSIS/tree/main