Synthfuse Diffuse Weight Model Retrain

Synthfuse retrain is a research-grade Python library for structured model retraining, treating fine-tuning and model upgrading as controlled motion on a continuous parameter manifold. Instead of ad-hoc optimization loops, retraining is formalized as a geometric or stochastic trajectory that advances an incumbent model toward—or beyond—a superior reference model while preserving architectural structure and explicitly controlling excess risk, information cost, and catastrophic forgetting. The library implements two complementary retraining regimes: Deterministic retraining, including warm-start fine-tuning (WSFT), knowledge-distillation transfer (KDT), optimal-transport weight interpolation (OTWI), and functional regularized reset (FRR). These methods are expressed as parameter-space flows with convergence guarantees or excess-risk bounds. Diffuse retraining, which models retraining as stochastic control over Langevin-type dynamics. Implemented methods include diffusion bridges with superior drift (DBSD), teacher-injected Langevin (TIL), functional diffusion distillation (FDD), and entropy-regularized Schrödinger bridges, enabling probabilistic control over terminal model distributions. To support these methods, retraining-manifold provides explicit implementations of parameter manifolds, optimal transport solvers, neural tangent kernels, and Fisher-information geometry, along with utilities for metrics, visualization, and data loading. A method-selection layer enables principled comparison and budget-aware choice among retraining paths. The repository includes reproducible experiments, configuration files, benchmarks, and tests, with end-to-end demonstrations on standard datasets. Sunthfuse-retrain is model-agnostic and framework-flexible. It can be used as a standalone retraining toolkit or embedded as a subsystem within larger algorithm-fusion frameworks such as Synthfuse, where it serves as a concrete instantiation of manifold- and diffusion-based retraining with theoretical guarantees.