FineTuneBench: How well do commercial fine-tuning APIs infuse knowledge into LLMs? (2411.05059v2)

Abstract: There is great interest in fine-tuning frontier LLMs to inject new information and update existing knowledge. While commercial LLM fine-tuning APIs from providers such as OpenAI and Google promise flexible adaptation for various applications, the efficacy of fine-tuning remains unclear. In this study, we introduce FineTuneBench, an evaluation framework and dataset for understanding how well commercial fine-tuning APIs can successfully learn new and updated knowledge. We analyze five frontier LLMs with commercially available fine-tuning APIs, including GPT-4o and Gemini 1.5 Pro, on their effectiveness in two settings: (1) ingesting novel information, such as recent news events and new people profiles, and (2) updating existing knowledge, such as updated medical guidelines and code frameworks. Our results reveal substantial shortcomings in all the models' abilities to effectively learn new information through fine-tuning, with an average generalization accuracy of 37% across all models. When updating existing knowledge, such as incorporating medical guideline updates, commercial fine-tuning APIs show even more limited capability (average generalization accuracy of 19%). Overall, fine-tuning GPT-4o mini is the most effective for infusing new knowledge and updating knowledge, followed by GPT-3.5 Turbo and GPT-4o. The fine-tuning APIs for Gemini 1.5 Flesh and Gemini 1.5 Pro are unable to learn new knowledge or update existing knowledge. These findings underscore a major shortcoming in using current commercial fine-tuning services to achieve reliable knowledge infusion in common scenarios. We open source the FineTuneBench dataset at https://github.com/kevinwu23/StanfordFineTuneBench.

• The paper presents FineTuneBench, a comprehensive framework for evaluating the effectiveness of fine-tuning APIs across diverse knowledge domains.
• It finds that models excel at memorization with near 100% accuracy but show significantly lower generalization, with rephrased queries averaging around 73% accuracy.
• The study emphasizes the challenge of updating entrenched knowledge via fine-tuning, suggesting the need for hybrid approaches to improve LLM adaptability.

Fine-Tuning as a Service: An Evaluation of Its Efficacy in Knowledge Infusion into LLMs

The research paper "FINE TUNE BENCH: HOW WELL DO COMMERCIAL FINE-TUNING APIS INFUSE KNOWLEDGE INTO LLMs?" by Eric Wu, Kevin Wu, and James Zou provides a methodological and empirical evaluation of the effectiveness of commercial fine-tuning APIs in knowledge infusion for LLMs. The paper introduces FineTuneBench, a comprehensive evaluation framework designed to assess the capabilities of fine-tuned models across various knowledge domains, including real-time news, fictional profiles, medical guidelines, and code updates.

Key Contributions and Evaluation Framework

FineTuneBench is a significant contribution, comprising 625 training questions and 1,075 test questions, designed to evaluate knowledge infusion capabilities in four distinct areas: latest news, fictional individuals, updated medical guidelines, and code changes. The framework targets five frontier LLMs, namely GPT-4o, GPT-4o-mini, GPT-3.5-turbo, Gemini-1.5 Pro, and Gemini-1.5 Flash, scrutinizing their ability to ingest and update information through commercially available fine-tuning APIs provided by OpenAI and Google.

One of the core aspects of this paper is the differentiation between knowledge memorization and generalization. The authors meticulously crafted datasets that include problems requiring simple recall of information (memorization) and those needing application of knowledge in varied or newly contextualized questions (generalization).

Numerical Results and Model Comparisons

A crucial finding of this research is the limited generalization capability of these fine-tuned models. While the OpenAI models, particularly the GPT-4o-mini, demonstrated almost perfect memorization of QA pairs with accuracy nearing 100% after 30 epochs of training, their ability to generalize this information to altered queries was substantially lower. For instance, on the rephrased Latest News dataset, these models achieved an average accuracy of only 73%.

Conversely, the performance evaluation of Gemini models from Google indicates significant limitations. These models struggled to memorize even basic training examples, with a top accuracy of 5.0% on new knowledge tasks after extensive fine-tuning, underscoring a substantial disparity when compared with OpenAI's offerings.

The paper also highlights that updating existing knowledge is more arduous than integrating new information. OpenAI's models displayed mediocre performance on updated knowledge datasets, such as coding questions, where they only achieved an average of 10% accuracy on rephrased queries. This illustrates the challenge of displacing entrenched knowledge in favor of new, current information through fine-tuning alone.

Implications and Future Directions

The findings present critical implications for the development and practical deployment of LLMs in domain-specific applications, especially in scenarios requiring continuous updates with cutting-edge information. Fine-tuning as a service, though appealing for its purported flexibility and adaptability, appears to be insufficient for robust knowledge infusion under the current commercial implementations.

The paper suggests that while retrieval-augmented generation (RAG) can serve as an alternative, it is not without drawbacks, including scalability issues and model inconsistencies when new information conflicts with pre-existing knowledge. Therefore, effective fine-tuning strategies that incorporate enhanced generalization capabilities remain a pressing research challenge.

For future exploration, the paper indicates that achieving reliable knowledge infusion may require the advancement of hybrid approaches or novel methodologies that extend beyond traditional fine-tuning paradigms. As the need for up-to-date and precise domain-specific knowledge continues to grow, these findings set the stage for further inquiry into optimizing fine-tuning mechanisms and propose potential avenues for improving the efficacy of such services.

In conclusion, "FINE TUNE BENCH" clearly establishes a foundational framework for evaluating fine-tuning effectiveness and provides pivotal insights into the capabilities and current limitations of commercial fine-tuning APIs for LLMs. It advocates for continued innovation in this field to enhance the adaptability and accuracy of LLMs in dynamic knowledge landscapes.