OptiMer: Optimal Distribution Vector Merging Is Better than Data Mixing for Continual Pre-Training

TL;DR

OptiMer introduces a post-hoc Bayesian optimization method to determine optimal data mixture ratios for continual pre-training of large language models, improving efficiency and flexibility.

Contribution

It decouples ratio selection from training by using distribution vectors and Bayesian optimization, outperforming traditional data mixing methods.

Findings

OptiMer outperforms data mixture and model averaging baselines by 15-35 times lower search cost.

Optimized weights can be interpreted as data mixture ratios, enhancing data mixture CPT.

Re-optimizing the same vector pool without retraining produces target-specific models on demand.

Abstract

Continual pre-training is widely used to adapt LLMs to target languages and domains, yet the mixture ratio of training data remains a sensitive hyperparameter that is expensive to tune: they must be fixed before training begins, and a suboptimal choice can waste weeks of compute. In this work, we propose OptiMer, which decouples ratio selection from training: we train one CPT model per dataset, extract each model's distribution vector, which represents the parameter shift induced by that dataset, and search for optimal composition weights post-hoc via Bayesian optimization. Experiments on Gemma 3 27B across languages (Japanese, Chinese) and domains (Math, Code) show that OptiMer consistently outperforms data mixture and model averaging baselines with 15-35 times lower search cost. Key findings reveal that 1) the optimized weights can be interpreted as data mixture ratios, and retraining with these ratios improves data mixture CPT, and 2) the same vector pool can be re-optimized for a given objective without any retraining, producing target-tailored models on demand. Our work establishes that data mixture ratio selection, traditionally a pre-training decision, can be reformulated as a post-hoc optimization over distribution vectors, offering a more flexible paradigm for continual pre-training.