Test model, trained on variable masking ratio on instruction data in the format 

```
[CLS]User: Tell me ... [SEP] Assistant[MASK][MASK][MASK] you [MASK] problem [SEP]
```

See [this notebook](https://huggingface.co/johnowhitaker/modernbert-diffusion/blob/main/MB_dLLM_sample.ipynb) for sampling.

Use at own risk, based on answerdotai/ModernBERT-large :)