New algorithm balances diverse capabilities when fine-tuning language models

Fine-tuning large language models requires careful curation of training data—a task that has proven surprisingly tricky when the goal is both strong performance and balanced capabilities across diverse domains. A new approach tackles this longstanding tension.

The bias problem in data selection#

Existing influence-based methods, which estimate how much each training example contributes to model predictions, excel at identifying high-impact data points. However, they carry a hidden flaw: some tasks naturally exert greater influence than others, causing selection algorithms to disproportionately favor the high-influence group. This creates a cascade of problems. Not only does the model underperform on neglected low-influence tasks, but paradoxically, it also struggles on the high-influence tasks that should benefit most from this preferential selection.

BIDS: A balanced alternative#

Researchers have introduced BIDS (Balanced and Influential Data Selection), an algorithm designed to correct this imbalance. The method works by first normalizing influence scores across the training dataset, then iteratively selecting the example with the highest influence on whichever task is most underrepresented at that stage.

The team validated BIDS on both Llama-3 and Mistral-v0.3 models, evaluating across seven benchmarks spanning five distinct capabilities. Results showed consistent improvements over both existing influence-based approaches and non-influence frameworks.

Striking efficiency gains#

Perhaps most impressively, training on just 15% of the original dataset—selected using BIDS—matched or exceeded performance from full-dataset training while achieving significantly more balanced results across different capability areas. The findings underscore the dual importance of normalizing influence at the instance level and using iterative optimization to ensure diverse capabilities develop in tandem.