Learning Dynamics in Continual Pre-Training for Large Language Models

• URL: http://arxiv.org/abs/2505.07796v2
• Date: Thu, 19 Jun 2025 10:38:17 GMT
• Title: Learning Dynamics in Continual Pre-Training for Large Language Models
• Authors: Xingjin Wang, Howe Tissue, Lu Wang, Linjing Li, Daniel Dajun Zeng,
• Abstract summary: Continual Pre-Training (CPT) has become a popular method to apply strong foundation models to specific downstream tasks.<n>We focus on how general and downstream domain performance evolves at each training step, with domain performance measured via validation losses.<n>Our formulation presents a comprehensive understanding of several critical factors in CPT, including loss potential, peak learning rate, training steps, replay ratio, etc.
• Score: 4.192010912385391
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Continual Pre-Training (CPT) has become a popular and effective method to apply strong foundation models to specific downstream tasks. In this work, we explore the learning dynamics throughout the CPT process for large language models. We specifically focus on how general and downstream domain performance evolves at each training step, with domain performance measured via validation losses. We have observed that the CPT loss curve fundamentally characterizes the transition from one curve to another hidden curve, and could be described by decoupling the effects of distribution shift and learning rate annealing. We derive a CPT scaling law that combines the two factors, enabling the prediction of loss at any (continual) training steps and across learning rate schedules (LRS) in CPT. Our formulation presents a comprehensive understanding of several critical factors in CPT, including loss potential, peak learning rate, training steps, replay ratio, etc. Moreover, our approach can be adapted to customize training hyper-parameters to different CPT goals such as balancing general and domain-specific performance. Extensive experiments demonstrate that our scaling law holds across various CPT datasets and training hyper-parameters.

Related papers

• PTMs-TSCIL Pre-Trained Models Based Class-Incremental Learning [7.784244204592032]
  Class-incremental learning (CIL) for time series data faces challenges in balancing stability against catastrophic forgetting and plasticity for new knowledge acquisition.<n>We present the first exploration of PTM-based Time Series Class-Incremental Learning (TSCIL)
  arXiv  Detail & Related papers  (2025-03-10T10:27:21Z)
• Dynamic Loss-Based Sample Reweighting for Improved Large Language Model Pretraining [55.262510814326035]
  Existing reweighting strategies primarily focus on group-level data importance.<n>We introduce novel algorithms for dynamic, instance-level data reweighting.<n>Our framework allows us to devise reweighting strategies deprioritizing redundant or uninformative data.
  arXiv  Detail & Related papers  (2025-02-10T17:57:15Z)
• Feasible Learning [78.6167929413604]
  We introduce Feasible Learning (FL), a sample-centric learning paradigm where models are trained by solving a feasibility problem that bounds the loss for each training sample.<n>Our empirical analysis, spanning image classification, age regression, and preference optimization in large language models, demonstrates that models trained via FL can learn from data while displaying improved tail behavior compared to ERM, with only a marginal impact on average performance.
  arXiv  Detail & Related papers  (2025-01-24T20:39:38Z)
• Non-asymptotic Convergence of Training Transformers for Next-token Prediction [48.9399496805422]
  Transformers have achieved extraordinary success in modern machine learning due to their excellent ability to handle sequential data. This paper provides a fine-grained non-asymptotic analysis of the training dynamics of a one-layer transformer. We show that the trained transformer presents non-token prediction ability with dataset shift.
  arXiv  Detail & Related papers  (2024-09-25T20:22:06Z)
• Enhancing Robustness of Vision-Language Models through Orthogonality Learning and Self-Regularization [77.62516752323207]
  We introduce an orthogonal fine-tuning method for efficiently fine-tuning pretrained weights and enabling enhanced robustness and generalization. A self-regularization strategy is further exploited to maintain the stability in terms of zero-shot generalization of VLMs, dubbed OrthSR. For the first time, we revisit the CLIP and CoOp with our method to effectively improve the model on few-shot image classficiation scenario.
  arXiv  Detail & Related papers  (2024-07-11T10:35:53Z)
• HiDe-PET: Continual Learning via Hierarchical Decomposition of Parameter-Efficient Tuning [55.88910947643436]
  We propose a unified framework for continual learning (CL) with pre-trained models (PTMs) and parameter-efficient tuning (PET)<n>We present Hierarchical Decomposition PET (HiDe-PET), an innovative approach that explicitly optimize the objective through incorporating task-specific and task-shared knowledge.<n>Our approach demonstrates remarkably superior performance over a broad spectrum of recent strong baselines.
  arXiv  Detail & Related papers  (2024-07-07T01:50:25Z)
• Multi-Epoch learning with Data Augmentation for Deep Click-Through Rate Prediction [53.88231294380083]
  We introduce a novel Multi-Epoch learning with Data Augmentation (MEDA) framework, suitable for both non-continual and continual learning scenarios. MEDA minimizes overfitting by reducing the dependency of the embedding layer on subsequent training data. Our findings confirm that pre-trained layers can adapt to new embedding spaces, enhancing performance without overfitting.
  arXiv  Detail & Related papers  (2024-06-27T04:00:15Z)
• FeTT: Continual Class Incremental Learning via Feature Transformation Tuning [19.765229703131876]
  Continual learning (CL) aims to extend deep models from static and enclosed environments to dynamic and complex scenarios. Recent CL models have gradually shifted towards the utilization of pre-trained models with parameter-efficient fine-tuning strategies. This paper proposes feature transformation tuning (FeTT) model to non-parametrically fine-tune backbone features across all tasks.
  arXiv  Detail & Related papers  (2024-05-20T06:33:50Z)
• On Training Data Influence of GPT Models [37.53037752668756]
  GPTfluence is a novel approach to assess the impact of training examples on the training dynamics of GPT models. Our approach traces the influence of individual training instances on performance trajectories, such as loss and other key metrics, on targeted test points.
  arXiv  Detail & Related papers  (2024-04-11T15:27:56Z)
• The Fine Line: Navigating Large Language Model Pretraining with Down-streaming Capability Analysis [27.310894780313618]
  This paper undertakes a comprehensive comparison of model capabilities at various pretraining intermediate checkpoints. We confirm that specific downstream metrics exhibit similar training dynamics across models of different sizes. In addition to our core findings, we've reproduced Amber and OpenLLaMA, releasing their intermediate checkpoints.
  arXiv  Detail & Related papers  (2024-04-01T16:00:01Z)
• A Loss Curvature Perspective on Training Instability in Deep Learning [28.70491071044542]
  We study the evolution of the loss Hessian across many classification tasks in order to understand the effect curvature of the loss has on the training dynamics. Inspired by the conditioning perspective, we show that learning rate warmup can improve training stability just as much as batch normalization.
  arXiv  Detail & Related papers  (2021-10-08T20:25:48Z)
• Multi-Stage Influence Function [97.19210942277354]
  We develop a multi-stage influence function score to track predictions from a finetuned model all the way back to the pretraining data. We study two different scenarios with the pretrained embeddings fixed or updated in the finetuning tasks.
  arXiv  Detail & Related papers  (2020-07-17T16:03:11Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.