TLDR: The research paper “BEFT: Bias-Efficient Fine-Tuning of Language Models” introduces a novel method for selecting specific bias terms in large language models for fine-tuning. By considering both angular and magnitude changes of bias terms, BEFT offers a more precise and dynamic selection process than existing approaches. This leads to significantly improved parameter efficiency, reduced training time, and competitive or superior performance across various LLMs and tasks, making it a highly effective strategy for adapting language models.
Large Language Models (LLMs) have become central to many AI applications, but fine-tuning them for specific tasks can be incredibly resource-intensive. This is where Parameter-Efficient Fine-Tuning (PEFT) methods come into play, aiming to reduce computational costs, GPU memory usage, and energy consumption. Among these, bias-only fine-tuning, which updates only the bias terms of an LLM, offers remarkable potential for efficiency.
While bias-only fine-tuning provides out-of-the-box usability without needing extra configurations or reparameterization, simply choosing any bias terms to fine-tune doesn’t guarantee optimal performance. The challenge has been understanding which specific bias terms (like those in query, key, or value projections) are most crucial for effective fine-tuning and how their selection impacts downstream performance. Existing approaches, such as those based on the magnitude of bias change or empirical Fisher information, have offered limited guidance.
A new research paper, “BEFT: Bias-Efficient Fine-Tuning of Language Models”, introduces an innovative approach to address this challenge. Authored by Baichuan Huang, Ananth Balashankar, and Amir Aminifar, this paper proposes a method for precisely selecting the bias term to be fine-tuned, forming the core of their Bias-Efficient Fine-Tuning (BEFT) strategy.
Understanding BEFT’s Approach
Unlike previous methods, BEFT jointly considers both the angular change and the magnitude of bias change before and after fine-tuning. It evaluates the projection ratios of bias terms, providing a more precise and dynamic measure of how biases change. This is a significant improvement over the Magnitude approach, which can assign the same importance score to different fine-tuned biases with the same magnitude change, and the Fisher approach, which often yields a static importance ranking.
The BEFT method calculates an importance score for each type of bias term (query, key, and value) across all Transformer layers. It then selects the bias type with the highest importance score as the target for fine-tuning. This dynamic selection process allows BEFT to adapt to different data regimes (low, medium, high) and tasks, ensuring that the most impactful bias terms are updated.
Key Findings and Advantages
The researchers extensively evaluated BEFT against Magnitude and Fisher approaches across a wide range of LLMs, including BERTBASE, RoBERTaBASE, BERTLARGE, and decoder-only architectures like OPT-1.3B and OPT-6.7B. These evaluations covered diverse downstream tasks such as classification, multiple-choice, and generation.
The results consistently demonstrated BEFT’s effectiveness and superiority. For instance, on the SST-2 dataset, BEFT consistently selected the value bias (bv) as the target, leading to higher accuracy compared to the Magnitude approach, which often selected the query bias (bq). BEFT also showed a dynamic importance ranking, accurately reflecting the narrowing performance gap between different bias terms as training data increased, a capability lacking in the static Fisher approach.
In terms of efficiency, BEFT achieved remarkable parameter efficiency, fine-tuning with only 0.01% of the full parameters in some cases, significantly less than the 0.09% required by fine-tuning all biases. Despite this reduction, BEFT maintained comparable or even superior performance to both all-bias fine-tuning and full-parameter fine-tuning, while also reducing training runtime.
Also Read:
- Beyond Single Scores: How DeCE Evaluates LLMs for Accuracy and Completeness
- CultureScope: A Deeper Look into AI’s Cultural Competence
Generalization and Future Potential
The insights gained from BEFT’s target bias selection proved to be generalizable across different datasets and LLMs without requiring additional measurements. The paper also showed that BEFT can effectively generalize to autoregressive LLMs (like OPT-1.3B and OPT-6.7B) and compete favorably with mainstream PEFT methods such as LoRA and prefix tuning, often requiring significantly fewer parameters while achieving competitive or better performance.
This research highlights BEFT as a promising strategy for parameter-efficient fine-tuning, offering a favorable trade-off between parameter percentage, runtime, and downstream performance. It provides valuable guidance for optimizing learning rates and training epochs in practice, paving the way for more efficient and effective adaptation of large language models.
