Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Few-Shot Optimized Framework for Hallucination Detection in Resource-Limited NLP Systems

Baraa Hikal, Ahmed Nasreldin, Ali Hamdi, Ammar Mohammed

TL;DR

This paper tackles hallucination detection in NLP under data scarcity by introducing DeepSeek Few-shot Optimization to generate high-quality weak labels, followed by data restructuring to suit generative-model training. The workflow includes iterative prompt refinement, a 30k-strong weak-label corpus, and LoRA-based fine-tuning of Mistral-7B-Instruct-v0.3, achieving strong downstream performance. An ensemble of seven fine-tuned checkpoints via majority voting delivers a test accuracy of 85.5% and secures the top position on the model-agnostic SHROOM track, demonstrating robustness and scalability in resource-limited settings. The results highlight the value of aligning data formats with model capabilities and leveraging weak supervision and ensemble strategies to push state-of-the-art in hallucination detection, with potential extensions to multilingual and cross-task applications.

Abstract

Hallucination detection in text generation remains an ongoing struggle for natural language processing (NLP) systems, frequently resulting in unreliable outputs in applications such as machine translation and definition modeling. Existing methods struggle with data scarcity and the limitations of unlabeled datasets, as highlighted by the SHROOM shared task at SemEval-2024. In this work, we propose a novel framework to address these challenges, introducing DeepSeek Few-shot optimization to enhance weak label generation through iterative prompt engineering. We achieved high-quality annotations that considerably enhanced the performance of downstream models by restructuring data to align with instruct generative models. We further fine-tuned the Mistral-7B-Instruct-v0.3 model on these optimized annotations, enabling it to accurately detect hallucinations in resource-limited settings. Combining this fine-tuned model with ensemble learning strategies, our approach achieved 85.5% accuracy on the test set, setting a new benchmark for the SHROOM task. This study demonstrates the effectiveness of data restructuring, few-shot optimization, and fine-tuning in building scalable and robust hallucination detection frameworks for resource-constrained NLP systems.

Few-Shot Optimized Framework for Hallucination Detection in Resource-Limited NLP Systems

TL;DR

This paper tackles hallucination detection in NLP under data scarcity by introducing DeepSeek Few-shot Optimization to generate high-quality weak labels, followed by data restructuring to suit generative-model training. The workflow includes iterative prompt refinement, a 30k-strong weak-label corpus, and LoRA-based fine-tuning of Mistral-7B-Instruct-v0.3, achieving strong downstream performance. An ensemble of seven fine-tuned checkpoints via majority voting delivers a test accuracy of 85.5% and secures the top position on the model-agnostic SHROOM track, demonstrating robustness and scalability in resource-limited settings. The results highlight the value of aligning data formats with model capabilities and leveraging weak supervision and ensemble strategies to push state-of-the-art in hallucination detection, with potential extensions to multilingual and cross-task applications.

Abstract

Hallucination detection in text generation remains an ongoing struggle for natural language processing (NLP) systems, frequently resulting in unreliable outputs in applications such as machine translation and definition modeling. Existing methods struggle with data scarcity and the limitations of unlabeled datasets, as highlighted by the SHROOM shared task at SemEval-2024. In this work, we propose a novel framework to address these challenges, introducing DeepSeek Few-shot optimization to enhance weak label generation through iterative prompt engineering. We achieved high-quality annotations that considerably enhanced the performance of downstream models by restructuring data to align with instruct generative models. We further fine-tuned the Mistral-7B-Instruct-v0.3 model on these optimized annotations, enabling it to accurately detect hallucinations in resource-limited settings. Combining this fine-tuned model with ensemble learning strategies, our approach achieved 85.5% accuracy on the test set, setting a new benchmark for the SHROOM task. This study demonstrates the effectiveness of data restructuring, few-shot optimization, and fine-tuning in building scalable and robust hallucination detection frameworks for resource-constrained NLP systems.

Paper Structure

This paper contains 16 sections, 4 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Prompt Engineering for Enhancing Weak Label Generation
  5. Initial Configuration
  6. Refinement with Task-Specific System Instructions
  7. Few-Shot Prompting with System Instructions
  8. Iterative Refinement Process
  9. Dataset Labeling and Reconstruction
  10. Model Fine-Tuning
  11. Results and Discussion
  12. Few-Shot Optimization for Weak Label Generation
  13. Ensemble Learning for Enhanced Robustness
  14. Model-Agnostic Track Rankings
  15. Discussion
  16. ...and 1 more sections

Figures (2)

  • Figure 1: Workflow diagram for weakly supervised data generation and fine-tuning.
  • Figure 2: Example of the dataset in its original format, intended for classification tasks.