Changelog
All notable changes to Safe dLLM will be documented in this file.
Unreleased
Added
Initial project structure
LLaDA model integration
Safety evaluation pipeline with AISafetyLab
Attack methods: PAIR, ReNeLLM, Prefill
Multi-GPU training support via Slurm
Sphinx documentation
Changed
Nothing yet
Fixed
Nothing yet
0.1.0 (2026-04-16)
Added
Initial release of Safe dLLM
SFT training script with diffusion masking
CPT continued pre-training support
Edit-based safety fine-tuning
Integration with HarmBench and WildGuardMix datasets
LlamaGuard3, HarmBench, PatternScorer evaluators
Batch evaluation with multi-GPU support
Slurm job submission scripts
Configuration system with OmegaConf
Comprehensive documentation