Summary of Exploring the Knowledge Mismatch Hypothesis: Hallucination Propensity in Small Models Fine-tuned on Data From Larger Models, by Phil Wee and Riyadh Baghdadi
Exploring the Knowledge Mismatch Hypothesis: Hallucination Propensity in Small Models Fine-tuned on Data from Larger Models
by Phil Wee, Riyadh Baghdadi
First submitted to arxiv on: 31 Oct 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The recent proliferation of large language models, fine-tuned with data from larger models, has led to the development of smaller models capable of producing outputs comparable in quality to their larger counterparts. However, these smaller models have been found to hallucinate more frequently than their larger counterparts, generating coherent yet factually incorrect information that spreads misinformation, toxicity, and stereotypes. To address this issue, researchers propose that fine-tuning a model on data produced by a larger model leads to a knowledge mismatch, contributing to increased hallucination propensity. This hypothesis suggests that the mismatch between the knowledge fed to the model for fine-tuning and the knowledge already present in the graph may lead to an increase in incorrect answers. Experimental results confirm this hypothesis, demonstrating that smaller models fine-tuned on data generated from larger models produce more wrong answers compared to those fine-tuned on data created by the small model. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Recently, many large language models have been made using training with larger model data. These small models can make things that look like they were made by bigger models. However, these small models often get things wrong and spread misinformation. One reason for this might be that fine-tuning a model on data from a bigger model creates a knowledge mismatch. This means that the smaller model gets confused about what it should know and what it already knows. Our results show that when we train a small model with data made by a bigger model, it makes more mistakes than when we train it with its own data. |
Keywords
» Artificial intelligence » Fine tuning » Hallucination
