carlesoctav/grafting-e4b_base_termgrade-1.0

Gemma 4 E4B Grafting Adapter (λ = 1.0)

This LoRA adapter is created via Grafting as described in Off-Policy Merging Beats On-Policy Self-Distillation for Continual Learning (Wu et al., 2026).

• Target Post-Trained Model ($\theta_{\text{post}}$): google/gemma-4-E4B-it
• Donor Base Model ($\theta_{\text{donor}}$): google/gemma-4-E4B
• Task Update ($\Delta$): Fine-tuned on termgrade_train_sft_32k.jsonl
• Grafting Scale ($\lambda$): 1.0 (effective lora_alpha = 16.0)

Usage

import torch
from transformers import AutoModelForCausalLM
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-4-E4B-it",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(
    base_model,
    "carlesoctav/grafting-e4b_base_termgrade-1.0",
)