index / HiGal/parallel-constrained-decoding
parallel-constrained-decoding
Fill a closed JSON schema in one batched forward pass instead of token-by-token. From-scratch reconstruction of the decoding trick likely behind TypeSafe's Jev: 28 fields in 4 passes vs 274, schema-valid by construction, per-field probabilities. MLX + PyTorch. Unofficial, not affiliated with TypeSafe.
Research and EvalsExtractionListedJupyter Notebook
Repository
Stars0
LanguageJupyter Notebook
LicenseMIT
Created2026-09-24
Last pushed2026-09-24
Officialno
Judgment
Genuine Jev project0.51
gate 0.5
Uses Jev at runtime0.02
Is a meta list0.03
Is a reimplementation0.92
Quality scales
Substance2.01 / 3
gate 0.5
Docs2.82 / 3
Novelty2.49 / 3
Composite score0.79
Category distribution
Research and Evals0.97
Other0.03
Agent and Dev Tooling0.00
SDKs and Clients0.00
Games and Simulation0.00
Applications0.00
Integrations0.00
Learning0.00
Assigned categoryResearch and Evals
Category probability0.97
Category confidence0.96
PatternExtraction
Pattern confidence0.72
Curation
StatusListed
Reasongate passed
README pickno
Overridenone
Sources
- search:jev in:name,description,readme created:2026-09-24
Provenance
Modeljev-1.13.0
Question setv2
Judged at2026-09-24 11:54 UTC