LoRA adapters saved every 10 steps (ckpt-000010 ... ckpt-000200) so you can pick the best step yourself. Key finding: the model OVERTRAINS - ckpt-100 is the peak, the final ckpt-200 is worse. Load an adapter onto the base with PEFT.
Base: dllm-hub/Qwen3-0.6B-diffusion-mdlm-v0.1 (GSM8K base = 29.3)
| checkpoint | GSM8K | HumanEval | note |
|---|---|---|---|
| base | 29.3 | 30.5 | no RL |
| ckpt-000100 | 34.72 (+5.4) | - | BEST GSM8K |
| ckpt-000200 (final) | 27.90 (-1.4) | 34.15 (+3.6) | overtrained on GSM8K |
Other checkpoints/tasks: eval in progress (see the GitHub repo BENCHMARKS.md).
import torch
from transformers import AutoModelForMaskedLM
from peft import PeftModel
base = "dllm-hub/Qwen3-0.6B-diffusion-mdlm-v0.1"
m = AutoModelForMaskedLM.from_pretrained(base, trust_remote_code=True, torch_dtype=torch.bfloat16)
m = PeftModel.from_pretrained(m, "LLM-OS-Models2/Qwen3-0.6B-diffusion-mdlm-justgrpo-run5-gsm8k-checkpoints", subfolder="ckpt-000100") # pick your step
Generate with the dllm MDLM sampler (block_size=256). Code: https://github.com/LLM-OS-Models/JustGRPO
ckpt-000010/adapter_model.safetensors 323 MBckpt-000020/adapter_model.safetensors 323 MBckpt-000030/adapter_model.safetensors 323 MBckpt-000040/adapter_model.safetensors 323 MBckpt-000050/adapter_model.safetensors 323 MBckpt-000060/adapter_model.safetensors 323 MBckpt-000070/adapter_model.safetensors 323 MBckpt-000080/adapter_model.safetensors 323 MBckpt-000090/adapter_model.safetensors 323 MBckpt-000100/adapter_model.safetensors 323 MBckpt-000110/adapter_model.safetensors 323 MBckpt-000120/adapter_model.safetensors 323 MBckpt-000130/adapter_model.safetensors 323 MBckpt-000140/adapter_model.safetensors 323 MBckpt-000150/adapter_model.safetensors 323 MBckpt-000160/adapter_model.safetensors 323 MBckpt-000170/adapter_model.safetensors 323 MBckpt-000180/adapter_model.safetensors 323 MBckpt-000190/adapter_model.safetensors 323 MBckpt-000200/adapter_model.safetensors 323 MB