Model Optimization
Two groundbreaking techniques that differentiate Arcee AI’s SLM training.
What is Spectrum?
Spectrum is a novel training methodology pioneered by Arcee AI researchers and built into our Small Language Model (SLM) training platform from the ground up.
It was designed to optimize the training process of language models by selectively training specific layers based on their signal-to-noise ratio (SNR).
What is Model Merging?
Model Merging allows customers to train an open source LLM on their data, then blend or “merge” that model with another open source LLM.
The merged model has the “brains” of both input LLMs, including the domain-specific customer data, but the size and inference cost of just one input model.”
Spectrum-Powered Pre-Training and Model Merging
Arcee AI’s revolutionary training routine
Spectrum-Powered Pre-Training
After uploading their dataset, a typical Arcee Cloud user clicks on the Pre-Training tab.
Pre-training of your data starts from a checkpoint of an open source model, and consists of actual knowledge injection (using no Lora).
Why do we call the pre-training “Spectrum-powered?” Spectrum is a unique technique pioneered by Arcee AI. Instead of updating every layer of the model during training, Spectrum identifies and prioritizes the layers that contribute most significantly to performance improvements (high SNR), while the layers with low SNR remain frozen.
Model Merging & Evolutionary (“Metric-Guided”) Model Merging
Next, we Merge that model with another open source one, resulting in a single model with all the advantages of both.
The new merged model is the same size and inference cost as one input model, and merging requires no training (so no GPU costs).
Because you've trained just on your data, then merged back some "brains" with the other model, you have mitigated catastrophic forgetting.
Just how performant is your merged model?
When merging models trained on the same tasks, you get better performance and better out-of-domain generalization.
When merging models trained on different tasks: you get synergistic effects which can actually boost task performance above input level
If you're starting with measurable qualities or competencies you want your model to have, you can do an Evolutionary Model Merge - where you choose the model to merge and the merging technique, and the algorithm finds the ideal combination of parameters.(NOTE: Evolutionary Model Merging is not as compute-efficient as base merging, but it's still an extremely efficient alternative to actual training.)