Layer-wise Swapping

Merge safety-tuned and multilingual-tuned LLMs by analyzing per-layer/module weight changes.

Usage

Layer-wise Swapping

Swap at layer granularity (entire transformer layers):

python layer_swap.py \
    -b meta-llama/Llama-3.1-8B-Instruct \
    -s ./checkpoint/safety_model \
    -m ./checkpoint/multi_model \
    -o ./checkpoint/merged

Module-wise Swapping

Swap at module granularity (Attention vs FFN separately):

python module_swap.py \
    -b meta-llama/Llama-3.1-8B-Instruct \
    -s ./checkpoint/safety_model \
    -m ./checkpoint/multi_model \
    -o ./checkpoint/merged

Arguments

Argument	Description	Default
-b, --base-model	Base model (HuggingFace ID or path)	Required
-s, --safety-model	Safety-tuned model path	Required
-m, --multi-model	Multilingual-tuned model path	Required
-o, --output	Output path	Required
--tau	Threshold for decision	0.001
--alpha	Blend ratio	0.5
--figure-dir	Directory for figures	<output>/figures

How It Works

1. Compute relative weight changes (ΔW) compared to base model
2. For each layer/module, compare safety vs multilingual changes
3. Decision per layer/module:
   • diff > tau → Use safety
   • diff < -tau → Use multilingual
   • Otherwise → Blend with ratio α

Data Preprocessing

See data/README.md for preparing multilingual SFT datasets.