liveBook · Manning

Domain-guided pruning

As explained in Rearchitecting LLMs

Domain-guided pruning selects model components for removal or preservation according to their relevance to a target domain.

Domain-guided pruning uses a domain-specific dataset to decide which model components to remove and which to preserve. It targets specialization rather than broad benchmark optimization, trading generality for stronger performance in a chosen domain. section 1.1.2

FineScope pruning

FineScope applies domain-guided structured pruning to Llama 3.1 8B. Accuracy remains stable to around 35% pruning with domain-specific guidance, while generic-data guidance begins degrading accuracy at about 25%. section 1.1.2

How Rearchitecting LLMs covers it

How deep it goes
Defines dataset-guided structured pruning and reports comparative thresholds, but does not detail scoring, selection, or implementation steps.
Examples
Uses FineScope and compares domain-specific guidance against generic-data guidance on Llama 3.1 8B.
What you can do after
Evaluate pruning guidance by comparing accuracy retention across domain-specific and generic datasets.
How it connects
Connects structured pruning with domain-specific datasets, model architecture, specialization, accuracy, and benchmark evaluation.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Rearchitecting LLMs. The section links point to the parts of the book it draws on.