Model Merger - API Specification
Version: v1.0.0 | Status: Active | Last Updated: March 2026
1. Overview
The model_merger module provides model merging techniques including SLERP (Spherical Linear Interpolation), linear interpolation, and model soups for combining multiple trained models.
2. Core Components
2.1 Classes
| Class | Description |
|---|---|
ModelMerger | Orchestrates merging of multiple model weight dictionaries |
2.2 Functions
| Function | Signature | Description |
|---|---|---|
slerp | (a, b, t) -> ndarray | Spherical linear interpolation between weight tensors |
linear_interpolate | (a, b, alpha) -> ndarray | Linear interpolation between weight tensors |
model_soup | (models, weights) -> dict | Weighted average of multiple model weight dictionaries |
3. Usage Example
from codomyrmex.model_merger import model_soup, slerp
import numpy as np
model_a = {"layer.weight": np.random.randn(64, 64)}
model_b = {"layer.weight": np.random.randn(64, 64)}
merged = model_soup([model_a, model_b], weights=[0.6, 0.4])