Llama2-sentiment-prompt-tuned

This model is a fine-tuned version of meta-llama/Llama-2-7b-chat-hf on an unknown dataset.

Model description

This model is Parameter Effecient Fine-tuned using Prompt Tuning. Our goal was to evaluate bias within LLama 2, and prompt-tuning is a effecient way to weed out the biases while keeping the weights frozen.

Classification Report of LLama 2 on original sentence:

          precision    recall  f1-score   support
negative       1.00      1.00      1.00       576
 neutral       0.92      0.95      0.93       640
positive       0.94      0.91      0.92       576

accuracy                           0.95      1792
macro avg      0.95      0.95      0.95      1792
weighted avg   0.95      0.95      0.95      1792

Classification Report of LLama 2 on preturbed sentence:

          precision    recall  f1-score   support
negative       0.93      0.74      0.82       576
 neutral       0.68      0.97      0.80       640
positive       0.80      0.58      0.67       576

accuracy                           0.77      1792
macro avg      0.80      0.76      0.76      1792
weighted avg   0.80      0.77      0.77      1792

Intended uses & limitations

You can use this model for your own sentiment-analysis task.

from transformers import AutoTokenizer
from peft import PeftModel
model_name = "furquan/llama2-sentiment-prompt-tuned"
model = PeftModel.from_pretrained(
    model_name, 
    device_map = 'auto'
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
model.eval()

def get_pred(text):
    inputs = tokenizer(f"\n### Text: {text}\n### Sentiment:", return_tensors="pt").to(model.device)
    outputs = model.generate(input_ids=inputs["input_ids"].to(model.device), attention_mask=inputs["attention_mask"], max_new_tokens=1,do_sample=False)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).split(' ')[-1]

prediction = get_pred("The weather is lovely today.")
print(prediction)

>>positive

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.001
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 2

Training results

Framework versions

• Transformers 4.36.0.dev0
• Pytorch 2.1.0+cu121
• Datasets 2.14.6
• Tokenizers 0.14.1