InterpretableRAGFactCheck

A modular, explainable Retrieval-Augmented Generation (RAG) pipeline for automated fact-checking on the FEVER dataset. Includes

• Multi-model evaluation (Mistral, Qwen, GPT 4.1, 4o, 3.5 Turbo) of RAG pipeline
• PEFT & DPO fine-tuning (Mistral model) with significant improvements of accuracy (56% to 74%)
• Token-level explainability with Captum, SHAP, and LIME.

📄 See project report for detailed experimental results, methodology, and ablation findings.

Submitted as part of "Advanced ML Applications" class, University Of Arizona, May 2025.

Introduction

This project shows that a retrieval-enhanced, interpretable fact-checking system can be both effective and insightful, with model selection, fine-tuning, and retrieval design - each contributing significantly to its accuracy and transparency

RAG Pipeline Architecture

Key Contributions and Implementation Highlights

• We develop an end-to-end RAG-based fact-checking system evaluated on the FEVER (1.0) dataset, with integrated explainability using Captum, SHAP, LIME.
• We conduct a comparative evaluation of multiple LLMs under identical RAG settings.
  • Open Sourced: Mistral-7B, Qwen-7B
  • Close Sourced : GPT-4.1, 4.1 mini, 4o, 3.5 Turbo
• We fine-tune Mistral-7B using PEFT (LoRA/QLoRA)
  • Demonstrate significant accuracy improvements (from 56%) to 73.5% claim verification accuracy.
• We explore DPO fine-tuning for fact-checking,
  • We provide one of the earliest experimental insights into its effectiveness in this domain.
• Integrate RAG with state-of-art RAGAS
  • for additional metrices like Faithfulness, answer relevancy.
• We perform ablation studies on the below, analyzing their impact on verification performance and explainability
  • retriever types,
  • retriever threshold cut-offs,
  • asynchronous retrieval,
  • prompt engineering strategies,
  • Performance on RAG vs standalone LLM

PEFT / DPO tuned models

Are available on Huggingface for reproducibility

• PEFT tuned https://huggingface.co/sauravverma78/mistral7bInstruct02_fever_trained
• DPO on peft : https://huggingface.co/sauravverma78/mistral7bInstruct02_fever_dpo

Directory Structure

project/
│
├── src
│    ├── fine_tuning
│    │      ├── RAG_Mistral_PEFT_tuning_73pc_accuracy_COLAB.ipnyb (PEFT fine tuning Mistral 7B model - COLAB)
│    │      │
│    │      └── RAG_Mistral_DPO_Tuning.ipnyb (DPO Fine tuning Mistral 7B model - COLAB)
│    │
│    ├── misral
│    │      ├── RAG_Mistral_COLAB.ipnyb (RAG Pipiline for Mistral 7B model - COLAB)
│    │      │
│    │      └── **RAG_Mistral_LOCAL.ipnyb** (RAG Pipiline and ablation for Mistral 7B model - Local Macbook & start reading here)
│    │
│    ├── qwen
│    │      └── RAG_QWEN2_LOCAL.ipnyb (RAG Pipiline for QWEN 7B model - Local Macbook)
│    │      
│    │
│    └── openai
│           └── RAG_OPENAI_COLAB_and_LOCAL.ipnyb (RAG Pipiline for OPEN AI model's' - Macbook + COLAB both supported)         
│
│
├── Apple_Analytics_Project_report.pdf  (Project report detailing the work and study)
│
├── environments.yaml  (to create a local conda virtual env with all requisite packages)
│
├── README.md  (this file)
│
├── proposal 
│  	  └──── Apply_Analytics_Project_proposal.pdf  (Initial project proposal)
│
├── images 
  	  └──── *.png (images for readme.md file)  

Environments and Accesses

Two main environments:

• Local: Macbook Pro M3 with 128GB RAM and 40-core Neural Engine.
  • Used for running models (Mistral, Qwen2) and initial prototyping.
• Cloud: Google Colab PRO (paid) with A100 GPU.
  • Used for computationally intensive tasks like model fine-tuning (PEFT/DPO)
  • RAG pipelines with OpenAI models and RAGAS computations

Accesses

You would need following standard accesses (some free and some paid) to run the notebooks

• Huggingface
  • to download Mistral/Qwen models
  • To load / save Fine tuned models
• Wandb.ai token
  • Required for APi access to https://api.wandb.ai internally when DPO Trainer starts
• open ai key
  • required to run GPT pipelines
• Colab Pro
  • To have a stable A-100 GPU access
• Google drive
  • for saving vector stores & model outputs for COLAB notebooks

Setup and Execution notes

Where to Start

RAG_Mistral_LOCAL.ipnyb contains the complete code and various ablation study and is a good starting point.

- RAG_Mistral_COLAB.ipnyb contains the COLAB implementation and install notes.

• Refer to qwen and openai folders for respective implementations
• fine_tuning
  • this package contains the PEFT and DPO fine tuning notebooks on MISTRAL 7B model on COLAB

Mistral-7B locally setup (Macbook) using conda

# this creates a virtuall env named `ml` and installs packages including jupyter
conda env create -f environments.yaml
conda activate ml
jupyter notebook

# Then open the LOCAL notebooks
# Then refer to instructions in the Notebook itself.
# Change local path accordingly (where we store models, vector DB, etc)

Mistral-7B on COLAB

• Simply execute the First Cell in Notebook, this installs all the packages
  • Note : During package install, Colab would ask for "restart" of session. Please do so. And run the pip install cell again.
• Requires access to your local google drive.
• Would need your HuggingFace Token (to download the models)

We suggest running cell by cell manually to see the executions, as some steps are time consuming

How to Execute

Each Notebook begins with details and notes required to get it working

• All the *LOCAL*.ipnyb notebooks were prepared on Macbook
• All the *COLAB*.ipnyb are for COLAB.
• Be sure to adjust any local paths in the notebook (e.g., for model checkpoints or DB folders).
• Recommend to manually run cell-by-cell as its necessary due to Colab memory resets or re-auth steps.

Results

RAG performance across models

Below are comparative results from different LLMs under our RAG pipeline:

Below are the predicted Vs actual token distribution plots for few models