Rainbow Padding: Mitigating Early Termination in Instruction-Tuned Diffusion LLMs
Paper • 2510.03680 • Published • 13
How to use quasar529/rainbow-padding-llada with PEFT:
from peft import PeftModel
from transformers import AutoModelForCausalLM
base_model = AutoModelForCausalLM.from_pretrained("GSAI-ML/LLaDA-8B-Base")
model = PeftModel.from_pretrained(base_model, "quasar529/rainbow-padding-llada")We introduce Rainbow Padding, a cyclic multi-token padding scheme that eliminates early termination and restores length robustness in instruction-tuned diffusion LLMs.
This checkpoint is a LoRA adapter for LLaDA-8B-Base with Rainbow Padding applied.
Base model
GSAI-ML/LLaDA-8B-Base