liveBook · Manning

Open-source model

As explained in Rearchitecting LLMs

An open-source model is a model distributed with source or parameter artifacts under terms that permit specified forms of inspection, modification, or reuse.

An open-source model is distributed with source or parameter artifacts under terms permitting specified inspection, modification, or reuse. It can reduce dependence on hosted providers, but openness alone does not guarantee domain specialization or freedom from benchmark-driven design. section 1.1.2

Evaluate open-source models against these engineering trade-offs:

  • Preserve exportability and version control instead of tying customization to a provider’s hosted base model. section 1.1.2
  • Check whether broad context windows, multilingual coverage, and leaderboard optimization serve the target domain or add unused capacity. section 1.1.2
  • Consider architecture adaptation for smaller models when specialized performance matters more than general benchmark breadth. section 1.1.2

Domain-guided pruning

FineScope uses a domain-specific dataset to guide structured pruning, preserving components relevant to the target domain. On Llama 3.1 8B, accuracy remains stable through approximately 35% pruning. section 1.1.2

How Rearchitecting LLMs covers it

How deep it goes
Contrasts open and closed models and discusses domain-guided pruning, stopping short of implementation details.
Examples
Uses FineScope and Llama 3.1 8B to illustrate domain-guided pruning and retained accuracy.
What you can do after
Evaluate open-source models for domain fit, customization portability, and targeted architecture adaptation.
How it connects
Connects open-source models to fine-tuning, vendor lock-in, foundation models, structured pruning, and domain specialization.

Also explained in

Compare how 6 books explain open-source model →

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.