Ruoxi Qiu's decision adapter for a frozen Qwen3.8-27B: a 6.3M-parameter LoRA on the last four text layers plus a numerical prior. One shared state feeds isolated question branches; each Choice, Noul or Score question returns probabilities with no generated tokens.
An independent Jev-shaped reproduction on DeBERTa-v3-large. One state and any number of choice, score and yes/no questions go in, and a calibrated distribution per question comes out of one forward pass. Public gold labels only.
Decides
choice, noul, score
choice, score, noul, classify
Architecture
canopy-jev
open-jev
Fine-tuned from
qwen/qwen3.8-27b
microsoft/deberta-v3-large
License
apache-2.0
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
87.4%
85.4%
Calibration error
—
0.022
Valid action rate
—
—
Median latency
—
28 ms
p95 latency
—
—
Figures are from each model’s manifest; accuracy and latency are what the publishers report, on their own suites and hardware. Add a third model.
Questions
What is the difference between canopy-jev and open-jev-deberta-v3-large?
canopy-jev is from Ruoxi Qiu and open-jev-deberta-v3-large from Kotoba Labs. Both have open weights you can download and run. Both answer choice, noul and score questions. Only open-jev-deberta-v3-large answers classify. open-jev-deberta-v3-large is the smaller model, at 434M parameters to 27B.
Which is more accurate, canopy-jev or open-jev-deberta-v3-large?
They report on different suites — canopy-jev 87.4% on JevBench public set (231 items; 202 correct), the maker's own fresh zero-shot run with the upstream scorer, open-jev-deberta-v3-large 85.4% on Kotoba held-out test, seen question types (banking77, SST-5, BoolQ) — so the numbers do not rank them. Test both on your own labelled examples.
Which is cheaper, canopy-jev or open-jev-deberta-v3-large?
canopy-jev: Free (open weights). open-jev-deberta-v3-large: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run canopy-jev or open-jev-deberta-v3-large locally?
Yes, both: systemone pull camellia86/canopy-jev and systemone pull kotoba-labs/open-jev-deberta-v3-large download the weights.
Evaluation suite
JevBench public set (231 items; 202 correct), the maker's own fresh zero-shot run with the upstream scorer
Kotoba held-out test, seen question types (banking77, SST-5, BoolQ)
