# Simple and Scalable Strategies to Continually Pre-train Large Language Models

arXiv: https://arxiv.org/abs/2403.08763 · HTML: https://mlsanity.com/paper/2403.08763/

## Key facts

- Cited by 127 papers in the map (data through 14 Oct 2026).
- Builds on 68 of its 98 references that are in the map.
- First posted on arXiv 13 Mar 2024.

## Builds on (most cited first, 30 of 68)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [High-Resolution Image Synthesis with Latent Diffusion Models](https://mlsanity.com/paper/2112.10752/) (2021)
- [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://mlsanity.com/paper/1910.10683/) (2019)
- [LLaMA: Open and Efficient Foundation Language Models](https://mlsanity.com/paper/2302.13971/) (2023)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Communication-Efficient Learning of Deep Networks from Decentralized Data](https://mlsanity.com/paper/1602.05629/) (2016)
- [Scaling Laws for Neural Language Models](https://mlsanity.com/paper/2001.08361/) (2020)
- [Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge](https://mlsanity.com/paper/1803.05457/) (2018)
- [Flamingo: a Visual Language Model for Few-Shot Learning](https://mlsanity.com/paper/2204.14198/) (2022)
- [HellaSwag: Can a Machine Really Finish Your Sentence?](https://mlsanity.com/paper/1905.07830/) (2019)
- [Mistral 7B](https://mlsanity.com/paper/2310.06825/) (2023)
- [Training Compute-Optimal Large Language Models](https://mlsanity.com/paper/2203.15556/) (2022)
- [TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension](https://mlsanity.com/paper/1705.03551/) (2017)
- [Neural Machine Translation of Rare Words with Subword Units](https://mlsanity.com/paper/1508.07909/) (2015)
- [PIQA: Reasoning about Physical Commonsense in Natural Language](https://mlsanity.com/paper/1911.11641/) (2019)
- [A Survey of Large Language Models](https://mlsanity.com/paper/2303.18223/) (2023)
- [The Pile: An 800GB Dataset of Diverse Text for Language Modeling](https://mlsanity.com/paper/2101.00027/) (2020)
- [BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions](https://mlsanity.com/paper/1905.10044/) (2019)
- [Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering](https://mlsanity.com/paper/1809.02789/) (2018)
- [BLOOM: A 176B-Parameter Open-Access Multilingual Language Model](https://mlsanity.com/paper/2211.05100/) (2022)
- [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://mlsanity.com/paper/1909.08053/) (2019)
- [Don't Stop Pretraining: Adapt Language Models to Domains and Tasks](https://mlsanity.com/paper/2004.10964/) (2020)
- [Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour](https://mlsanity.com/paper/1706.02677/) (2017)
- [ZeRO: Memory Optimizations Toward Training Trillion Parameter Models](https://mlsanity.com/paper/1910.02054/) (2019)
- [Scaling Language Models: Methods, Analysis & Insights from Training Gopher](https://mlsanity.com/paper/2112.11446/) (2021)
- [Experience Replay for Continual Learning](https://mlsanity.com/paper/1811.11682/) (2018)
- [Adaptive Federated Optimization](https://mlsanity.com/paper/2003.00295/) (2020)
- [Scaling Vision Transformers](https://mlsanity.com/paper/2106.04560/) (2021)
- [On Layer Normalization in the Transformer Architecture](https://mlsanity.com/paper/2002.04745/) (2020)

## Built on by (most cited first, 30 of 127)

- [Continual Learning of Large Language Models: A Comprehensive Survey](https://mlsanity.com/paper/2404.16789/) (2024)
- [Scaling Laws and Compute-Optimal Training Beyond Fixed Training Durations](https://mlsanity.com/paper/2405.18392/) (2024)
- [Adam-mini: Use Fewer Learning Rates To Gain More](https://mlsanity.com/paper/2406.16793/) (2024)
- [Zamba: A Compact 7B SSM Hybrid Model](https://mlsanity.com/paper/2405.16712/) (2024)
- [Dynamic Chunking for End-to-End Hierarchical Sequence Modeling](https://mlsanity.com/paper/2507.07955/) (2025)
- [When LLMs Meet Cybersecurity: A Systematic Literature Review](https://mlsanity.com/paper/2405.03644/) (2024)
- [Towards Lifelong Learning of Large Language Models: A Survey](https://mlsanity.com/paper/2406.06391/) (2024)
- [Synthetic continued pretraining](https://mlsanity.com/paper/2409.07431/) (2024)
- [CorDA: Context-Oriented Decomposition Adaptation of Large Language Models for Task-Aware Parameter-Efficient Fine-tuning](https://mlsanity.com/paper/2406.05223/) (2024)
- [Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs](https://mlsanity.com/paper/2412.13337/) (2024)
- [Advances and Open Challenges in Federated Foundation Models](https://mlsanity.com/paper/2404.15381/) (2024)
- [Scaling Laws for Forgetting during Finetuning with Pretraining Data Injection](https://mlsanity.com/paper/2502.06042/) (2025)
- [Straight to Zero: Why Linearly Decaying the Learning Rate to Zero Works Best for LLMs](https://mlsanity.com/paper/2502.15938/) (2025)
- [How Susceptible are LLMs to Influence in Prompts?](https://mlsanity.com/paper/2408.11865/) (2024)
- [Llama-3.1-FoundationAI-SecurityLLM-Base-8B Technical Report](https://mlsanity.com/paper/2504.21039/) (2025)
- [Scaling Law with Learning Rate Annealing](https://mlsanity.com/paper/2408.11029/) (2024)
- [CtRL-Sim: Reactive and Controllable Driving Agents with Offline Reinforcement Learning](https://mlsanity.com/paper/2403.19918/) (2024)
- [WikiBigEdit: Understanding the Limits of Lifelong Knowledge Editing in LLMs](https://mlsanity.com/paper/2503.05683/) (2025)
- [Efficient Continual Pre-training by Mitigating the Stability Gap](https://mlsanity.com/paper/2406.14833/) (2024)
- [Encode, Think, Decode: Scaling test-time reasoning with recursive latent thoughts](https://mlsanity.com/paper/2510.07358/) (2025)
- [A Practitioner's Guide to Continual Multimodal Pretraining](https://mlsanity.com/paper/2408.14471/) (2024)
- [How to Merge Your Multimodal Models Over Time?](https://mlsanity.com/paper/2412.06712/) (2024)
- [WSM: Decay-Free Learning Rate Schedule via Checkpoint Merging for LLM Pre-training](https://mlsanity.com/paper/2507.17634/) (2025)
- [A Multi-Power Law for Loss Curve Prediction Across Learning Rate Schedules](https://mlsanity.com/paper/2503.12811/) (2025)
- [The Zamba2 Suite: Technical Report](https://mlsanity.com/paper/2411.15242/) (2024)
- [Mid-Training of Large Language Models: A Survey](https://mlsanity.com/paper/2510.06826/) (2025)
- [A Survey on LLM Mid-Training](https://mlsanity.com/paper/2510.23081/) (2025)
- [Kanana: Compute-efficient Bilingual Language Models](https://mlsanity.com/paper/2502.18934/) (2025)
- [Data Mixing Can Induce Phase Transitions in Knowledge Acquisition](https://mlsanity.com/paper/2505.18091/) (2025)
- [Bielik 7B v0.1: A Polish Language Model -- Development, Insights, and Evaluation](https://mlsanity.com/paper/2410.18565/) (2024)

## Abstract (arXiv)

Large language models (LLMs) are routinely pre-trained on billions of tokens, only to start the process over again once new data becomes available. A much more efficient solution is to continually pre-train these models, saving significant compute compared to re-training. However, the distribution shift induced by new data typically results in degraded performance on previous data or poor adaptation to the new data. In this work, we show that a simple and scalable combination of learning rate (LR) re-warming, LR re-decaying, and replay of previous data is sufficient to match the performance of fully re-training from scratch on all available data, as measured by the final loss and the average score on several language model (LM) evaluation benchmarks. Specifically, we show this for a weak but realistic distribution shift between two commonly used LLM pre-training datasets (EnglishrightarrowEnglish) and a stronger distribution shift (EnglishrightarrowGerman) at the 405M parameter model scale with large dataset sizes (hundreds of billions of tokens). Selecting the weak but realistic shift for larger-scale experiments, we also find that our continual learning strategies match the re-training baseline for a 10B parameter LLM. Our results demonstrate that LLMs can be successfully updated via simple and scalable continual learning strategies, matching the re-training baseline using only a fraction of the compute. Finally, inspired by previous work, we propose alternatives to the cosine learning rate schedule that help circumvent forgetting induced by LR re-warming and that are not bound to a fixed token budget.