README.md

---
license: apache-2.0
datasets:
- msamogh/indirect-requests
language:
- en
metrics:
- accuracy
base_model:
- google-t5/t5-base
pipeline_tag: text2text-generation
library_name: transformers
tags:
- prompt_restructuring
- prompt_refining
- indirect_requests
- pragmatics
---
# PragmaticLM - T5 for Prompt Restructuring

![Model](assets/dp.png)

## 📌 Overview
**PragmaticLM** is a fine-tuned T5 model designed to **restructure and reframe user prompts** for better understanding by downstream LLMs. The model enhances prompt clarity by leveraging **contextual restructuring** techniques.

## 🚀 Model Details
- **Base Model**: [T5-Base](https://huggingface.co/t5-base)
- **Training Data**: [Indirect Requests] (https://huggingface.co/datasets/msamogh/indirect-requests)
- **Task Type**: Text-to-text transformation
- **Library**: [Hugging Face Transformers](https://github.com/huggingface/transformers)

## 📊 Training Configuration
- **Epochs**: 10
- **Batch Size**: 8
- **Learning Rate**: Encoder: `1e-5`, Decoder: `3e-5`
- **Optimizer**: AdamW
- **Loss Function**: Cross-entropy loss
- **Hardware**: GPU (T4)

## ⚡ Usage
```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

tokenizer = AutoTokenizer.from_pretrained("aliMohammad16/pragmaticLM")
model = AutoModelForSeq2SeqLM.from_pretrained("aliMohammad16/pragmaticLM")

def restructure_prompt(input_prompt):
    input_text = f"Restructure Prompt: {input_prompt}"
    inputs = tokenizer(input_text, return_tensors="pt", padding=True)
    
    output = model.generate(
        inputs.input_ids,
        max_length=64,
        num_beams=4,
        early_stopping=True
    )
    
    return tokenizer.decode(output[0], skip_special_tokens=True)

# Example Usage
test_prompt = "I am not feeeling well. I need to consult a doctor nearby."
print(restructure_prompt(test_prompt))
```

## ⏳ Improvements
- **Work in progress**: This is a work in progress. I am actively working on this model.
- **Update**: Next I am implementing a multimodular pipeline, integrating TinyLlama 1.1B and Llama Index RAG with `prompt-restructuring` model, to improve output generation.