Assessing the Reliability of LLMs in Faithfully Updating Text

This repository accompanies our research paper (Assessing the Reliability of LLMs in Faithfully Updating Text) on the FRUIT task – a systematic evaluation of how large language models integrate new evidence into existing documents while preserving factual consistency and coherence.

Abstract

Faithfully updating text with new information is a critical challenge in natural language processing. The FRUIT task (Faithfully Reflecting Updated Information in Text) evaluates models on their ability to integrate new evidence into existing documents while maintaining factual consistency and coherence. This paper systematically investigates GPT-4o and Llama-3-8B, exploring various prompting strategies, fine-tuning techniques, and evidence ranking methods. In addition to standard processing, we examine streaming evidence scenarios, where updates are applied incrementally. Our findings reveal that while LLMs can effectively incorporate new facts, they struggle with hallucinations, evidence selection, and coherence maintenance—especially when handling multiple updates or streaming input. Structured evidence presentation significantly improves performance, but iterative updates lead to degradation, highlighting challenges in maintaining consistency over successive edits. By providing a detailed analysis of existing methods, this study identifies key gaps in automatic text updating and outlines directions for improving model reliability in real-world applications.

Table of Contents

• Requirements
• Repository Structure
• Getting Started
  • Installation
  • Dataset
• Usage
  • Generating Predictions
  • Dataset Preparation
  • Evaluation
• Configuration
• License

Requirements

This project requires Python 3.9 or above. Please ensure that you have the following packages installed:

• vllm
• rouge_score
• matplotlib
• tigerscore
• prometheus
• axolotl

You can install these packages via pip (or add them to your requirements.txt if you prefer):

pip install vllm rouge_score matplotlib tigerscore prometheus axolotl

Repository Structure

.
├── base_prompts
│   ├── cot.py              # Implements chain-of-thought prompting
│   └── zero_shot.py        # Implements zero-shot prompting
├── dataset_prep
│   └── dataset.py          # Generates a .pkl file needed for predictions; change paths as required
├── evaluation
│   ├── evaluate_generated.py   # Evaluates prediction files and generates graphs/metrics
│   └── update_rouge.py         # Computes ROUGE scores for updated texts
├── evidence ordering
│   ├── factual_filter_sequence.py  # Orders evidence based on factual consistency
│   ├── filter_rank.py               # Ranks evidence candidates
│   ├── random_shuffline.py          # (For ablation) Randomizes evidence ordering
│   └── streaming_evidences.py       # Handles streaming evidence scenarios
├── few shot
│   ├── example_picking.py   # Picks few-shot examples for prompting
│   └── vllm_few_shot.py   # Implements few-shot prompting using vLLM
├── finetuning
│   └── finetuned.py         # Contains fine-tuning routines for LLMs
├── LICENSE
├── README.md
└── reflect_refine
    ├── prometheus_correct.py    # Self-correction module using Prometheus methods
    ├── prometheus_evaluate.py   # Evaluates corrections based on Prometheus scoring
    ├── self_correction.py       # Implements iterative self-correction
    ├── tiger_score_correct.py   # Corrects predictions based on Tiger score metrics
    └── tiger_score_evaluate.py  # Evaluates updated texts using Tiger score metrics

Notes:

• Output Files:
  All modules—except those in evaluation and dataset_prep—output predictions as .jsonl files suitable for Llama-3-8B instruct inference.

• Dataset Preparation:
  The dataset.py script in dataset_prep generates the necessary .pkl file to obtain all predictions.
  The original dataset can be found in the original fruit repository.

Getting Started

Installation

1. Clone the repository:

   git clone https://github.com/paheli/faith-update-llm.git
   cd faith-update-llm

2. Install required packages:
   Make sure you have Python 3.9+ installed, then run:

   pip install -r requirements.txt

   If you do not have a requirements.txt, please manually install the necessary packages:

   pip install vllm rouge_score matplotlib tigerscore prometheus axolotl

Dataset

Download the dataset from the original fruit repository and place it in your preferred location. Remember: Update the file paths in dataset.py and any other relevant scripts to point to your local dataset directory.

Usage

Generating Predictions

Run the modules (except those in evaluation and dataset_prep) to generate .jsonl prediction files. For example, to use the chain-of-thought prompting:

python base_prompts/cot.py

Dataset Preparation

Generate the required dataset file by running:

python dataset_prep/dataset.py

This script outputs a .pkl file that is used to retrieve all predictions.

Evaluation

After generating predictions, evaluate them by running:

python evaluation/evaluate_generated.py

This will produce graphs and various metrics (including UpdateROUGE scores via update_rouge.py).

Configuration

Important:
Before running any scripts, please update the file paths and configuration variables in the code to match the locations of your dataset, model checkpoints, and output directories.

License

This project is licensed under the MIT License – see the LICENSE file for details.