Skip to main content
EMEP is built on a single scientific hypothesis: new high-performing large language models can be created from existing open-weight models without training from scratch, by systematically combining compatible parameter spaces and selecting superior combinations through evolutionary optimization. This page states the hypothesis precisely, defines the engineering and commercial objectives, draws phase boundaries, and lists explicit exclusions.

​
Scientific Hypothesis

Given a population of open-weight language models with compatible architectures, there exists a subset of tensor-level merge operations that produces candidate models with higher aggregate benchmark performance than any single ancestor model. Evolutionary search over merge strategies and model combinations can discover these candidates more efficiently than exhaustive enumeration.
This hypothesis rests on prior work in model merging (Model Soups: Wortsman et al. 2022; Task Arithmetic: Ilharco et al. 2022; TIES-Merging: Yadav et al. 2023; DARE: Yu et al. 2023; Evolutionary Model Merge: Akiba et al. 2024). EMEP does not claim to prove this hypothesis. It builds a reproducible platform to test it.

​
Engineering Objective

Build a modular, documented, and reproducible platform that:
  1. Discovers and imports open-weight models into a versioned ModelRegistry
  2. Analyzes architectural and tensor compatibility between any pair of models
  3. Executes merge operations with deterministic, hashed, and fully-provenanced artifacts
  4. Evaluates candidates on a defined Benchmark Catalog split into Optimization Set, Validation Set, and Hidden Test Set
  5. Evolves populations of candidates through mutation, crossover, and multi-objective selection
  6. Quantizes and deploys the best models to offline or air-gapped environments
The platform must be operable by a single ML engineer on a single node, and scale to distributed GPU clusters without architectural changes.

​
Commercial Objective

Enable organizations to create specialized, high-performing language models from open-weight bases without:
  • Proprietary pre-training data
  • Multi-million dollar training runs
  • Dependency on closed API providers
  • License violations from unvetted model combinations
EMEP is not a product with a price. It is a specification. Any implementation that conforms to this specification may be deployed commercially, provided it respects the license terms of all base models and merge inputs.

​
Phase Boundaries

EMEP development is organized into ten phases. Each phase has a hard boundary: no phase begins until the previous phase has passed its acceptance criteria.
Skipping phases is prohibited. Phase 5 cannot begin until Phase 4 acceptance criteria are met and signed off in the ExperimentTracker. This rule is non-negotiable. See Project Constitution Rule 12.

​
Explicit Exclusions

The following are permanently excluded from EMEP scope unless a future ADR explicitly reverses the decision.

​
Scope Diagram

The following diagram shows what is inside EMEP scope (solid lines) and what is outside (dashed lines).