Concept: The method evaluates each model’s tensors in two domains to determine contribution strength.
- Spatial. Measures direct differences between model weight tensors in parameter space.
Distinct structural weight patterns receive higher spatial scores.
- Frequency. Applies FFT on flattened weight tensors to analyze spectral behavior; unique and
consistent frequency structures receive higher frequency scores.
Both domains are combined into a unified per-model “goodness” value. These values are scaled by clarity, mixed via blend and global_blend, and normalized through a temperature-controlled softmax (focus) to generate final merge weights.
If a base tensor is provided, it acts as the reference center; otherwise, the method computes the median tensor across donors as the neutral baseline.
Key Parameters.
- focus (per-model/global): softmax temperature; higher = more selective contributions.
- blend (per-model/global): mix between spatial (0) and frequency (1) domains.
- clarity (per-model): Scales each model’s influence by its signal stability or confidence.
- global_blend (global): global bias toward spatial (0) or frequency (1).
Show YAML
models:
- model: Retreatcost/Darkstar-12B
parameters:
weight: 0.42
focus: 1.2
blend: 0.65
clarity: 0.95
- model: Retreatcost/Impish-LongPen-12B
parameters:
weight: 0.58
focus: 1.3
blend: 0.55
clarity: 1.1
merge_method: harmony_flux
dtype: bfloat16
parameters:
global_blend: 0.5
return_debug: true
tokenizer:
source: union
