Open-Jev DeBERTa v3 large

• Adapter: open-jev — one sequence, span-pooled option scoring
• License: Apache-2.0
• Parameters: ~435M · Architecture: DeBERTa-v3-large + span head
• Language: English · Context: 512 tokens

Default variant

Repo	onnx-community/open-jev-deberta-v3-large-ONNX
Files	onnx/model_fp16.onnx + .onnx_data, tokenizer.json, config.json
Quantization	fp16
Download	884 MB
Role	default

Variants

Variant	Quantization	Size	Role	Install
fp16 *	fp16	883.5 MB	default	dtk models download open-jev-deberta-v3-large --variant fp16
fp32	fp32	1.76 GB	quality	dtk models download open-jev-deberta-v3-large --variant fp32
q4	4-bit	486.7 MB	edge	dtk models download open-jev-deberta-v3-large --variant q4
q4f16	4-bit + fp16	357.0 MB	edge	dtk models download open-jev-deberta-v3-large --variant q4f16

Start with fp16; use q4/q4f16 when download size or CPU work matters, and re-check thresholds. See Variants and quantization.

Reproducibility

• Revision: 7c79f25b5ac496089f448a969c801872ad59d31c
• Default file hashes (SHA256):
• onnx/model_fp16.onnx — 2fd22c68fef1c74ac8d4541ea4cbcfbd38553c15552a7bd0a09280c196662e9b
• onnx/model_fp16.onnx_data — f5cc34a4e6cc82483fbaa8bac85f38b008758b13c4dfbb2ac1b637a92da3fd14
• tokenizer.json — cd119378b0160677b7a1e561ba29ada83918c1d420b326516a585382e83d9d39
• config.json — 2ec35432332ee6b5880509eefe44e6279fd9d3543f6ba96098119ffe0b0c2d5e
• Verify: dtk models verify open-jev-deberta-v3-large --variant fp16
• Bench: uv run python bench/run.py --suite bench/suites/smoke.jsonl --models open-jev-deberta-v3-large
• Hashes for the other variants live in catalog.py.

Evidence (2026-10-21)

• CPU p50 after warmup: 241.2 ms fp16, 78.2 ms fp32, 56.0 ms q4, 219.4 ms q4f16. The 4-bit paths are size/edge choices, not automatic wins.
• Published example (I was charged twice…): product area fees & charges (0.70), sentiment 0.819, refund 0.941 with fp16/fp32.
• The upstream model card reports 0.854 in-domain accuracy, 0.690 on its OOD question/option set, and ECE 0.022 in-domain.
• Local fp16 and fp32 agree on the published example; q4 variants are lossy and should be validated against your thresholds.

Contract notes

• Sequence: [CLS] [STATE] state [Q] instructions [OPT] option … [SEP].
• Graph inputs: input_ids, attention_mask, seg, pair_q, pair_opt.
• One forward pass scores every option; softmax is per question group with the checkpoint temperature (1.05).
• The state is capped to 256 tokens and the full sequence to 512.

Known limitations

• English only; 512-token context; it reads the question only partially (upstream in-domain 0.85 vs OOD 0.69).
• Score on unseen level sets is weak in the upstream evaluation.
• It chooses among supplied options; it cannot generate explanations.
• This is an independent Open-Jev DeBERTa v3 large reproduction, not an official TypeSafe model.

Loading each variant

REST selects the quantization through the standard model field using the alias open-jev-deberta-v3-large@variant; there is no variant field in the request body.

q4f16 (q4/fp16, edge)

Library

from decision_toolkit import DecisionModel

model = DecisionModel("open-jev-deberta-v3-large", variant="q4f16")

CLI

dtk models download open-jev-deberta-v3-large --variant q4f16
dtk models use open-jev-deberta-v3-large --variant q4f16
dtk decide --model open-jev-deberta-v3-large --variant q4f16 \
  --state "..." --q 'urgent=noul:Is this urgent?'

REST

curl -s http://127.0.0.1:8000/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{"model": "open-jev-deberta-v3-large@q4f16", "state": "...",
       "questions": {"urgent": {"type": "noul",
                     "instructions": "Is this urgent?"}}}}'

fp16 (fp16, default)

Library

from decision_toolkit import DecisionModel

model = DecisionModel("open-jev-deberta-v3-large", variant="fp16")

CLI

dtk models download open-jev-deberta-v3-large --variant fp16
dtk models use open-jev-deberta-v3-large --variant fp16
dtk decide --model open-jev-deberta-v3-large --variant fp16 \
  --state "..." --q 'urgent=noul:Is this urgent?'

REST

curl -s http://127.0.0.1:8000/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{"model": "open-jev-deberta-v3-large@fp16", "state": "...",
       "questions": {"urgent": {"type": "noul",
                     "instructions": "Is this urgent?"}}}}'

q4 (q4, edge)

Library

from decision_toolkit import DecisionModel

model = DecisionModel("open-jev-deberta-v3-large", variant="q4")

CLI

dtk models download open-jev-deberta-v3-large --variant q4
dtk models use open-jev-deberta-v3-large --variant q4
dtk decide --model open-jev-deberta-v3-large --variant q4 \
  --state "..." --q 'urgent=noul:Is this urgent?'

REST

curl -s http://127.0.0.1:8000/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{"model": "open-jev-deberta-v3-large@q4", "state": "...",
       "questions": {"urgent": {"type": "noul",
                     "instructions": "Is this urgent?"}}}}'

fp32 (fp32, quality)

Library

from decision_toolkit import DecisionModel

model = DecisionModel("open-jev-deberta-v3-large", variant="fp32")

CLI

dtk models download open-jev-deberta-v3-large --variant fp32
dtk models use open-jev-deberta-v3-large --variant fp32
dtk decide --model open-jev-deberta-v3-large --variant fp32 \
  --state "..." --q 'urgent=noul:Is this urgent?'

REST

curl -s http://127.0.0.1:8000/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{"model": "open-jev-deberta-v3-large@fp32", "state": "...",
       "questions": {"urgent": {"type": "noul",
                     "instructions": "Is this urgent?"}}}}'

Usage

dtk models download open-jev-deberta-v3-large
dtk decide --model open-jev-deberta-v3-large \
  --state "I was charged twice and want a refund." \
  --q team='choice:billing=payments|technical=bugs|other=anything else' \
  --q refund='noul:Is the customer asking for a refund?'

Provenance

ONNX conversion by onnx-community of com-kotobalabs/open-jev-deberta-v3-large, an independent open reproduction of the Jev-shaped typed-decision contract. Apache-2.0.