twistshan/Sudoku_DLM

Sudoku DLM

Research artifacts for comparing reasoning in discrete diffusion models on 9x9 Sudoku. The current release contains the v1 MDM-no-t depth sweep trained with seed 0.

Layout

v1/
  MANIFEST.json
  MANIFEST.sha256
  models/mdm-no-t/L{4,8,16}/
    checkpoints/{best,final}/
    training/
    evaluations/{validation,test}/

Each best checkpoint minimizes validation denoising cross-entropy. Each final checkpoint is the state after epoch 300 (global step 57,600). The manifest records the byte size and SHA-256 digest of every uploaded artifact.

The complete versioned dataset, source provenance, and verification scripts are stored separately at twistshan/Sudoku_DLM_Reasoning. Machine logs and W&B internal files are intentionally excluded from this public model repository.