66.8M-parameter masked diffusion language model (MDLM, DiT backbone) for English psychology / mental-health dialogue, trained on one RTX 3090. The 120M sibling is published to make the paper's scaling null independently checkable. Paper: Small-Scale Text Diffusion for a Psychology Assistant (Denis Kim, 2026), PDF in the GitHub repo.

Research artifact for studying small-scale text diffusion. Not a therapist, not a medical device, not for crisis use. Outputs are frequently vague, repetitive, and can be factually wrong (paper §9: blind homonym-drift floor 25–32%). For crisis support call 988 (US) or local equivalents. No real conversations or PII in training data.

Use confidence-ordered unmasking (v2 sampler). Random-order unmasking re-introduces repetition collapse (paper §5).

MDLM objective (mask prob t∼U(0.01,0.99), CE on masked); DiffuSeq-style SFT (response-only noising). Pretrain 20k steps / SFT 8k, batch 256×512, LR 4e-4 cosine, bf16 + AdamW8bit. Data: strict-filtered psychology corpus 0.774B tokens; SFT 261k pairs (assistant + mental-health + translated psychology dialogue). Corpora not redistributed; build scripts in repo.

725 paired prompts, MiniLM cosine-to-gold: 66m-clean 0.500; 120m 0.490 (d=−0.0094, win 0.479 → null at fixed data). Blind traps (3 annotators, κ .77–.87): 66m 9/28 vs 120m 7/28, p=0.77. distinct-2 0.993; rep-1 0.137; collapse 0.