Chapter 2: Model Merging Techniques

Combining Models for Better Results

Model merging combines multiple fine-tuned models into one.

---

Merging Methods

Method	Description	Best For
Linear	Average weights	Simple combination
SLERP	Spherical interpolation	Smooth blending
TIES	Conflict resolution	Multiple experts
DARE	Drop and rescale	Diverse models

---

LoRA Merging

# Merge LoRA into base model
from peft import PeftModel

model = PeftModel.from_pretrained(base_model, "adapter_path")
merged_model = model.merge_and_unload()
merged_model.save_pretrained("merged_model")

---

MergeKit Example

# config.yaml
slices:
  - sources:
      - model: base_model
        layer_range: [0, 32]
      - model: finetuned_model
        layer_range: [0, 32]
merge_method: slerp
base_model: base_model
parameters:
  t: 0.5  # Interpolation factor
dtype: bfloat16

mergekit-yaml config.yaml ./merged_output

---

When to Merge

✅ Merge when:

• Combining task-specific models
• Deploying simpler architecture
• No need for adapter switching

❌ Keep separate when:

• Need dynamic task switching
• Storage is limited
• Want to update tasks independently

---

Continue to Chapter 3 →