🌈 Rainbow Padding: Mitigating Early Termination in Instruction-Tuned Diffusion LLMs

Official PyTorch implementation of Rainbow Padding, a simple yet powerful strategy that resolves <eos> overflow in diffusion language models (dLLMs).

Visit our Project page and arxiv for paper if you are interested! This repository provides a step-by-step pipeline for SFT LoRA training, evaluation using Rainbow Padding.

If you have any questions, please contact the authors.

---

Demo

Rainbow Padding

LLaDA Instruct

---

1. Setup

1️⃣ Create the Conda Environment

conda env create -f environment.yaml

2️⃣ Activate the Environment

conda activate rainbow

---

2. Dataset Preparation

We follow the curation recipe introduced in Dream (arXiv:2508.15487).
The training corpus consists of 0.5M public examples curated from:

• Tulu3 SFT Mixture
• SmolLM2 / SmolTalk

Details are provided in Appendix C.1 of the paper.

⚠️ Note: Specific SFT configurations for both Dream and LLaDA were not publicly released (to the best of our knowledge).

Download pre-tokenized data (recommended)

You can directly download our preprocessed datasets from Google Drive:

# Same data with tokenization per model type.

# LLaDA SFT data
gdown --folder 1U8kVGYiWRsqWCDRsHUjeKTiDPrL0FsMp

# Dream SFT data
gdown --folder 1-oei1KRTFADMRljPqX5rPuTGcJ7fpHdI

---

3. LoRA SFT Training

We use 🤗 Accelerate for multi-GPU training.

Key Arguments for main.py

Argument	Description
batch_size	Batch size per GPU. Control the total batch size using gradient_accumulation_steps in ./method/sft.py.
pad_num	Number of cyclic padding tokens. Use 0 for <eos> padding or any positive integer (e.g., 3, 7) for Rainbow Padding.

---

Example: Training with 4 GPUs and 7 Rainbow Padding Tokens

1️⃣ Initial Training

CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch --num_processes=4 main.py --model_type=llada_base --pad_num=7

2️⃣ Continue Training from a Checkpoint

CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch --num_processes=4 main.py --model_type=llada_base  --pad_num=7  --resume_dir model/llada_base/sft_5e-05_lora_epoch3_rank32_pad7

---

4. Evaluation

We upload our checkpoint to Hugging Face: quasar529/rainbow-padding-llada.

We use widely used library LM-Eval-Harness, and we modified the evaluation script from LLaDA's eval script. You can find the evaluation script in eval/eval_llada_instruct.py.

⚠️ Dependency Notice

To run evaluation, you must install specific versions of datasets and lm-eval due to dependency constraints:

pip install datasets==3.6.0 lm-eval==0.4.9.1

• lm-eval==0.4.9.1 requires datasets>=2.16.0,<4.0.
• However, the latest datasets (≥4.0.0) is incompatible.
• Therefore, you need to downgrade datasets to 3.6.0, which satisfies lm-eval's requirements and ensures stable evaluation. If you skip this step, evaluation scripts may still run but can break unexpectedly due to mismatched APIs.

Example Command

# Example: Humaneval_instruct
accelerate launch --num_processes=1 eval/eval_llada_instruct.py \
  --tasks humaneval_instruct \
  --model llada_dist \
  --batch_size 1 \
  --log_samples \
  --output_path "/home/quasar529/rainbow-padding/eval/output" \
  --confirm_run_unsafe_code \
  # If you want to use wandb, set wandb_log, wandb_project, wandb_entity
  --model_args model_path='GSAI-ML/LLaDA-8B-Base',steps=1024,gen_length=1024,block_length=1024,lora_path='quasar529/rainbow-padding-llada',device='cuda',wandb_log=True,wandb_project='llada-eval',wandb_entity='your-wandb-entity'

If you want to reproduce all evaluation tasks performed in the paper at once, you can simply run the provided shell script:

sh eval/eval.sh

5. Citation

If you find this work useful, please cite:

@article{kim2025rainbow,
  title={Rainbow Padding: Mitigating Early Termination in Instruction-Tuned Diffusion LLMs},
  author={Kim, Bumjun and Jeon, Dongjae and Kim, Dueun and Jeung, Wonje and No, Albert},
  journal={arXiv preprint arXiv:2510.03680},
  year={2025}
}

---

6. Acknowledgements

This code builds upon the open-sourced implementations of
Dream and LLaDA.
We thank the authors for releasing their resources and inspiring this work.

---