Monitoring Decoding

A repository for implementing and evaluating monitoring-based decoding strategies for mitigating hallucination in LLM generation.

Overview

This repository contains implementations of monitoring decoding algorithms that use a base model and an expert model to eliminate the hallucinations in the LLM's generated output. The approach involves monitoring the generation process and switching between models based on confidence scores.

Features

• Multi-model decoding: Support for base and expert model combinations
• Multiple datasets: GSM8K, TruthfulQA, TriviaQA, NQ-Open
• Comprehensive evaluation: EM metrics, accuracy scores, Truthfulness, Informativeness
• Flexible configuration: Command-line arguments for all parameters
• Modular design: Clean, reusable code structure

Supported Models

• Llama-2: 7B, 13B, 70B variants
• Llama-3: 8B, 70B variants
• Gemma: 2B, 27B variants

Installation

# Clone the repository
git clone <your-repo-url>
cd monitoring_decoding

# Install dependencies
pip install -r requirements.txt

Usage

Text Generation

# Basic usage
python generate_text.py --data_task gsm8k --base_model Llama-2-7b --expert_model Llama-2-70b

# Custom parameters
python generate_text.py \
    --data_task truthfulqa \
    --base_model Llama-3-8b \
    --expert_model Llama-3-70b \
    --width 4 \
    --depth 4 \
    --branch 2 \
    --r 0.5 \
    --data_num 408

Evaluation

GSM8K Evaluation

python evaluate_gsm8k.py \
    --base_model Llama-2-7b \
    --expert_model Llama-2-70b \
    --width 4 \
    --depth 3 \
    --branch 2 \
    --r 0.5 \
    --data_num 1319 \
    --save_results

EM Evaluation (for QA tasks)

python evaluation_em.py \
    --data_task nqopen \
    --base_model Llama-2-7b \
    --expert_model Llama-2-70b \
    --width 4 \
    --depth 3 \
    --branch 2 \
    --r 0.3 \
    --data_num 1000 \
    --save_results

File Structure

monitoring_decoding/
├── generate_text.py          # Main generation script
├── evaluate_gsm8k.py         # GSM8K evaluation
├── evaluation_em.py          # EM evaluation for QA tasks
├── decoding.py           # Decoding algorithms
├── utils/
│   ├── generation_probs.py   # Probability calculations
│   ├── metric_utils.py       # Evaluation metrics
│   ├── tokenize.py           # Tokenization utilities
│   └── eval_info.txt         # Evaluation templates
├── results/                  # Generated outputs
├── evaluation_results/       # Evaluation results
└── README.md

Parameters

Generation Parameters

• --data_task: Dataset to use (gsm8k, truthfulqa, triviaqa, nqopen, xsum)
• --base_model: Base model name
• --expert_model: Expert model name
• --width: Search width for generation
• --depth: Length of generated candidates
• --branch: Number of candidates being remained
• --r: Ratio threshold for decoding
• --data_num: Number of data points to process

Evaluation Parameters

• --targets: Models to evaluate (base, expert, decode)
• --split_prediction: Split prediction at newlines/periods
• --save_results: Save results to file
• --output_dir: Output directory for results

Results

The evaluation scripts provide:

• Accuracy scores for each model
• Statistical analysis (mean, std, min, max)
• Detailed breakdowns of correct/incorrect predictions

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests if applicable
5. Submit a pull request

Citation

If you use this code in your research, please cite:

@misc{chang2025monitoringdecodingmitigatinghallucination,
      title={Monitoring Decoding: Mitigating Hallucination via Evaluating the Factuality of Partial Response during Generation}, 
      author={Yurui Chang and Bochuan Cao and Lu Lin},
      year={2025},
      eprint={2503.03106},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.03106}, 
}