README.md

---
library_name: peft
license: apache-2.0
base_model: xlm-roberta-base
tags:
- base_model:adapter:xlm-roberta-base
- lora
- transformers
metrics:
- accuracy
- f1
model-index:
- name: bengali-code-mix-sentiment-lora
  results: []
datasets:
- Swarnadeep-28/bn_code_mix_sentiment_dataset
pipeline_tag: text-classification
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# bengali-code-mix-sentiment-lora

This model is a fine-tuned version of [xlm-roberta-base](https://huggingface.co/xlm-roberta-base) on the None dataset.
It achieves the following results on the evaluation set:
- Loss: 0.7597
- Accuracy: 0.7206
- F1: 0.7206

## Model description

This model is a **LoRA Parameter-Efficient Fine-Tuned** version of [`xlm-roberta-base`](https://huggingface.co/xlm-roberta-base) for **sentiment analysis** on **Bengali–English code-mixed text** (commonly found in social media posts, comments, and tweets).

- **Task**: Text Classification (Sentiment Analysis)  
- **Languages**: Bengali (Romanized) + English  
- **Classes**: `positive`, `negative`, `neutral`  
- **Fine-tuning method**: LoRA (PEFT)  
- **Dataset**: [Bengali-English Code-Mixed Sentiment Dataset](https://huggingface.co/datasets/Swarnadeep-28/bn_code_mix_sentiment_dataset)  

This model enables efficient, low-resource fine-tuning while maintaining competitive performance for code-mixed sentiment classification.

## How to Use

### Inference Example
```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel
import torch

# Load tokenizer & model
model_id = "Swarnadeep-28/bengali-code-mix-sentiment-lora"
tokenizer = AutoTokenizer.from_pretrained(model_id)
base_model = AutoModelForSequenceClassification.from_pretrained("xlm-roberta-base", num_labels=3)
model = PeftModel.from_pretrained(base_model, model_id)

# Example text
text = "Aaj match ta khub bhalo chilo! Loved it."
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    logits = model(**inputs).logits
pred = torch.argmax(logits, dim=-1).item()
labels = ["negative", "neutral", "positive"]
print("Predicted label:", labels[pred])

```




## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0005
- train_batch_size: 16
- eval_batch_size: 32
- seed: 42
- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- num_epochs: 5

### Training results

| Training Loss | Epoch | Step | Validation Loss | Accuracy | F1     |
|:-------------:|:-----:|:----:|:---------------:|:--------:|:------:|
| 0.6325        | 1.0   | 1001 | 0.8349          | 0.6982   | 0.6974 |
| 0.7065        | 2.0   | 2002 | 0.7734          | 0.7096   | 0.7093 |
| 0.6849        | 3.0   | 3003 | 0.7649          | 0.7171   | 0.7149 |
| 0.6452        | 4.0   | 4004 | 0.7603          | 0.7176   | 0.7180 |
| 0.669         | 5.0   | 5005 | 0.7597          | 0.7206   | 0.7206 |


### Framework versions

- PEFT 0.17.1
- Transformers 4.56.1
- Pytorch 2.8.0+cu126
- Datasets 4.0.0
- Tokenizers 0.22.0