README.md

metadata

language:
  - en
tags:
  - text-generation
  - flan-t5
  - lora
  - peft
  - hallucination
  - qa
license: mit
datasets:
  - Pravesh390/qa_wrong_data
library_name: transformers
pipeline_tag: text-generation
model-index:
  - name: flan-t5-finetuned-wrongqa
    results:
      - task:
          name: Text Generation
          type: text-generation
        metrics:
          - name: BLEU
            type: bleu
            value: 18.2
          - name: ROUGE-L
            type: rouge
            value: 24.7

🔍 flan-t5-finetuned-wrongqa

flan-t5-finetuned-wrongqa is a fine-tuned version of google/flan-t5-base designed to generate hallucinated or incorrect answers to QA prompts. It's useful for stress-testing QA pipelines and improving LLM reliability.

🧠 Model Overview

• Base Model: FLAN-T5 (Google's instruction-tuned T5)
• Fine-Tuning Library: 🤗 PEFT + LoRA
• Training Framework: Hugging Face Transformers + Accelerate
• Data: 180 hallucinated QA pairs in qa_wrong_data (custom dataset)

📚 Intended Use Cases

• Hallucination detection
• QA model robustness evaluation
• Educational distractors (MCQ testing)
• Dataset augmentation with adversarial QA

🧪 Run with Gradio

import gradio as gr
from transformers import pipeline

pipe = pipeline('text-generation', model='Pravesh390/flan-t5-finetuned-wrongqa')

def ask(q):
    return pipe(f'Q: {q}\nA:')[0]['generated_text']

gr.Interface(fn=ask, inputs='text', outputs='text').launch()

⚙️ Quick Colab Usage

from transformers import pipeline
pipe = pipeline('text-generation', model='Pravesh390/flan-t5-finetuned-wrongqa')
pipe('Q: What is the capital of Australia?\nA:')

📊 Metrics

• BLEU: 18.2
• ROUGE-L: 24.7

🏗️ Libraries and Methods Used

• transformers: Loading and saving models
• peft + LoRA: Lightweight fine-tuning
• huggingface_hub: Upload and repo creation
• datasets: Dataset management
• accelerate: Efficient training support

📁 Sample QA Example

• Q: Who founded the Moon?
• A: Elon Moonwalker

📄 License

MIT