Automated Evaluation Method for Assessing Hallucination in RAG Models

In Short

• Introduces a two-phase process using an automated exam builder and IRT to assess RAG models’ realism and effectiveness.
• Applies NLP preprocessing and various embedding techniques to create high-quality, domain-independent exam questions, ensuring broad applicability.
• Validates the approach across multiple domains, including AWS DevOps and StackExchange, highlighting its versatility and reliability.

As the use of Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) models grows, the need for accurate evaluation metrics becomes critical. This article introduces a groundbreaking approach combining an automated exam builder and Item Response Theory (IRT) to quantify the realism of RAG models. By leveraging natural language processing and prompt engineering, this method offers a scalable, cost-efficient, and precise way to assess model performance across various domains, ensuring reliable and detailed evaluations.

The RAG method is employed for the natural-language queries, and the LLM works with texts that were taken from the documents to provide a response, shifting the pertinent ones. The RAG model can be applied to a particular task by the LLM with a knowledge base set for a particular task which would help the LLM to come up with multiple choices. More importantly, it does not mandate specific architectures for the retriever as well as the generators which compose part of the RAG system and the exam generation task.

Phases

This plan was described as having two phases. First, we call an LLM and use several techniques of the prompt engineering to create the set of candidate questions relevant to all the documents included in the knowledge base. Subsequently, several NLP preprocessing steps are performed to filter out low-quality questions based on several parameters such as size, grammar, redundancy.

To ensure credible comparisons, we took close consideration of various closed book pipeline changes and looked into other forms of pipes as well as the oracle and classical types of pipes which include the MultiQA embeddings, Siamese network embeddings, BM25. In addition, it also trained language models of different scales up to 70 Billion parameters in order to see how scaling the model impacts the performance.

To justify the all-encompassing nature, we submitted this analysis to various domains such as AWS DevOps, titles of arXiv papers, questions asked in StackExchange and even SEC filings. This helps to remove dependency on the performances of single domain models and can make our evaluations even more accurate and reliable to a number of practical applications for DLs.

The inclusion IRT to this process proved very effective and indeed bolstered the quality of the first exams tremendously. To achieve an empirical assessment within a feature space of a question and a model, the probability of correct response is estimated in IRT. Easy, difficult, discrimination, and guessing chance are the four parameters available in the tool, and it forms question papers with three parameters: difficulty, discrimination, and a guess chance where the results are pre-assigned.

What are the key differences between large language models (LLMs) and generative AI?

We thus begin by filtering the entire assessment, demanding an initial test with an aim to progress as a stripping off the least discriminant inquiry questions. This work is then next subjected to the level two and level three IRT parameters so that the exam results are capable of measuring detailed model behaviors, as it is then enhanced by IRT.

As a result of these demonstrations, in order to increase the usage of RAG models and better facilitate the creation of examinations, the tags have been derived from postsemantic analysis and Bloom’s revised taxonomy. This means that the current approach offers a way of categorizing questions based on the kind of cognitive abilities required in order to answer them, and therefore a way of getting more systematic about the kind of performance that the models in question are capable of exhibiting.

Notably and in summary, the presented framework is a two-channel, cost-efficient and human-interpretable approach to evaluating RAG models, in which exam generation via an automated process and conformed by IRT, ensures the validity of the assessment. These substring analyses show examples of how the models presented are feasible when used on any given practical problem. Whenever such new paradigms get incorporated into the field of LLMs, the roles for assessing them become broader and the techniques we employ will also evolve to reflect the nature of the models we wish to assess as accurately as possible.

Recommended Reading

FineWeb: 15 Trillion Token Dataset Redefines LLM Pretraining (Hugging Face)

Princeton and Warwick Collaboration Introduce AI Method to Make LLMs Smarter