open-jev-deberta-v3-large

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

• Maturity5.8/10

  Clear HF card and typed-decisions repo with a public-gold training record; encoder specialist without hosted SLA sits mid maturity.

• Capability6.3/10

  All three primitives in one encoder pass with post-hoc temperature; strong in-domain author numbers, weaker OOD — mid capability, not a Jev peer.

• Adoption9/100

  Present in S1Bench gap lists and kotoba-lang docs; still a specialist niche versus chat-backbone replicas.

Vendor claims

• ~28ms end-to-end for 10 questions on one state (H100 bf16); ~1.8s for 4q on M1 Max CPU fp32 (author)[Vendor claim — not independently verified]
• In-domain test acc ~0.854 / ECE ~0.022; OOD new questions ~0.690 (author)[Vendor claim — not independently verified]
• Trained only on public gold (Banking77, SST5, BoolQ) — no teacher LM labels (author)[Vendor claim — not independently verified]

open-jev-deberta-v3-large is an open Jev-shaped encoder decision model on DeBERTa-v3-large — independent work from kotoba-lang / com-kotobalabs — not TypeSafe.

One shared state plus many typed questions (choice ≤255 options, score 2–10 levels, noul → catalog boolean) → calibrated distributions in a single encoder forward pass (no generation). Weights: com-kotobalabs/open-jev-deberta-v3-large; code: github.com/kotoba-lang/typed-decisions.

Limits (read before you ship)

• Context budget is tight (author: ~512 tokens total; state clipped ~256). Long tickets need your own chunking.
• Latency claims (~28ms H100 / ~1.8s M1 CPU) are self-reported.
• In-domain vs OOD gap is large in the author card; do not treat Banking77-style accuracy as open-world System One transfer.
• Author explicitly says numbers are not comparable to TypeSafe’s private-workflow agreement metrics — agree; we mark comps UnverifiedClaim.

Why it is in the catalog

Public gold-only training, open weights, and true multi-question encoder I/O. Complements chat-backbone peers like Decider and Laya with a classical NLU backbone.