REVIEW 7 cited by
Continual Pre-Training of Large Language Models: How to (re)warm your model?
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
abstract
Large language models (LLMs) are routinely pre-trained on billions of tokens, only to restart the process over again once new data becomes available. A much cheaper and more efficient solution would be to enable the continual pre-training of these models, i.e. updating pre-trained models with new data instead of re-training them from scratch. However, the distribution shift induced by novel data typically results in degraded performance on past data. Taking a step towards efficient continual pre-training, in this work, we examine the effect of different warm-up strategies. Our hypothesis is that the learning rate must be re-increased to improve compute efficiency when training on a new dataset. We study the warmup phase of models pre-trained on the Pile (upstream data, 300B tokens) as we continue to pre-train on SlimPajama (downstream data, 297B tokens), following a linear warmup and cosine decay schedule. We conduct all experiments on the Pythia 410M language model architecture and evaluate performance through validation perplexity. We experiment with different pre-training checkpoints, various maximum learning rates, and various warmup lengths. Our results show that while rewarming models first increases the loss on upstream and downstream data, in the longer run it improves the downstream performance, outperforming models trained from scratch$\unicode{x2013}$even for a large downstream dataset.
Forward citations
Cited by 7 Pith papers
-
WSqD: A Horizon-Free Learning Rate Schedule for Large Model Training
WSqD uses a horizon-independent shifted inverse-square-root base plus linear cooldown to get optimal last-iterate rates in convex stochastic optimization and match or beat tuned WSD on LLM pretraining with one reused ...
-
Universal Music Representations? Evaluating Foundation Models on World Music Corpora
Five audio foundation models are evaluated across six Western and non-Western music corpora, showing a consistent Western-centric bias and only limited generalization to culturally distant traditions.
-
Scaling Point-in-Time Language Models
Scaling point-in-time LLMs to 4B parameters and 1T temporally filtered tokens narrows the gap to unrestricted models to about 8–11 average points and yields positive out-of-sample Sharpe ratios from news embeddings.
-
Continual Learning in Transition
A tri-axial framework of When, Where, and How organizes the ongoing transition of continual learning from parameter-centric updates to system-level capability evolution.
-
MedLLM: An Open Medical Language Model at the Sub-Billion Scale
A 100M-parameter model, MedLLM, is reported to beat 7B medical baselines on MedMCQA and MMLU while staying near the floor on MedQA, but the claimed 'within 2.9pp of a medically adapted 7B' result has no matching row i...
-
ZUNA1.1: A more flexible EEG foundation model for Denoising and Super-resolution
ZUNA1.1, an open-source 380M EEG diffusion autoencoder, reconstructs variable-length, flexibly masked EEG at least as well as its predecessor and far better than spherical spline interpolation.
-
Router Upcycling: Leveraging Mixture-of-Routers in Mixture-of-Experts Upcycling
Router Upcycling initializes MoE routers from the dense model's attention heads, achieving about 2% higher average zero-shot accuracy than vanilla upcycling on small Qwen 8x0.5B models.
Discussion (0). Continue with ORCID to comment.