A masked-diffusion language model (LLaDA / MaskGIT style). The SQL answer begins as a fixed window of masked tokens; each step predicts every masked position and permanently commits the most confident ones, so the query emerges in parallel instead of left-to-right. During training, the model sees examples at many noise levels by hiding different fractions of the target SQL, then learns to reconstruct only the hidden tokens from the prompt, schema, and visible SQL tokens. Unlike a causal decoder, the transformer can look across the whole proposed SQL window at once, which helps it coordinate table names, joins, filters, and syntax globally. At inference time this runs in reverse: early high-confidence tokens become anchors, and later steps use that partially completed query to resolve the remaining blanks. Adaptive early stopping can finish before the full step budget when every remaining blank is already predicted with high confidence. The animation shows those intermediate states.

Backbone: ModernBERT-base with a masked-LM head, fine-tuned LLaDA-style (continuous mask ratio, 1/t-weighted ELBO) on the gretelai/synthetic_text_to_sql dataset with on-the-fly augmentation (spaced / mixed-case identifiers, colloquial prompts) for robustness.

Input is your question (Prompt) plus the schema DDL (Context); output is the SQL.

Model parameters: 149.66M

Requested backend: onnx

Runtime backend: onnxruntime (worker process)

Requested dtype: fp32

Effective dtype: fp32

Put your question in Prompt and the table schema (e.g. a CREATE TABLE …) in Context, then run. More denoising steps can improve quality but take longer.

Email: thorekoritzius@gmail.com