Checkpoints and Released Adapters

Base Models

VoidPadding adapters are LoRA adapters and must be used with their original base models.

Dream instruct model:

hf download Dream-org/Dream-v0-Instruct-7B \
  --local-dir pretrained/Dream-v0-Instruct-7B

LLaDA instruct model:

hf download GSAI-ML/LLaDA-8B-Instruct \
  --local-dir pretrained/LLaDA-8B-Instruct

You can pass either the Hugging Face model id or a local directory to merge and eval scripts:

BASE_MODEL="Dream-org/Dream-v0-Instruct-7B"
BASE_MODEL="pretrained/Dream-v0-Instruct-7B"

For base-model training experiments:

hf download Dream-org/Dream-v0-Base-7B \
  --local-dir pretrained/Dream-v0-Base-7B

hf download GSAI-ML/LLaDA-8B-Base \
  --local-dir pretrained/LLaDA-8B-Base

Released LoRA Adapters

Released adapter repositories:

akpon900/dream-instruct-rainbow
akpon900/dream-instruct-void
akpon900/llada-instruct-rainbow
akpon900/llada-instruct-void

Download them with:

hf download akpon900/dream-instruct-void \
  --local-dir models/dream-instruct-void

hf download akpon900/dream-instruct-rainbow \
  --local-dir models/dream-instruct-rainbow

hf download akpon900/llada-instruct-void \
  --local-dir models/llada-instruct-void

hf download akpon900/llada-instruct-rainbow \
  --local-dir models/llada-instruct-rainbow

Expected adapter layout:

models/
├── dream-instruct-void/
├── dream-instruct-rainbow/
├── llada-instruct-void/
└── llada-instruct-rainbow/

Each adapter repository contains:

• README.md
• adapter_config.json
• adapter_model.safetensors

Merge Notes

Evaluation expects merged full-model checkpoints. The README Quick Start shows the standard Dream and LLaDA merge commands.

Useful overrides:

export TORCH_DTYPE=bfloat16
export DEVICE=cuda          # auto, cpu, cuda, or cuda:N
export TOKENIZER_PATH=/path/to/tokenizer
export PYTHON=python