decision2:nox
Decision 2.0 · vLLM Semantic Router · 4.21B · ONNX Runtime · Apache-2.0
- Score
- 60.9
- 95 % interval 58.7–62.9
- Rank
- 8
- of 66 scored models
- Public benchmark
- 74.6
- macro accuracy, %
- typed-decisions
- 68.0
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
53.1
Classification
MASSIVE en
85.2
MASSIVE de
84.9
Civil Comments
55.3
Aegis 2
68.0
Understanding
MultiNLI
84.5
VitaminC
64.4
PAWS
52.8
Knowledge
BoolQ
66.7
SQuAD 2
50.5
PubMedQA
58.6
Judging
HelpSteer 2
22.2
SummEval rel.
33.9
SummEval cons.
73.1
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 74.6(73.3–75.9) | 82.0 | 79.3 | 54.4 | 0.050 | 0.344 | 0 / 3,880 |
| typed-decisions | 68.0(65.9–70.3) | 66.0 | 74.8 | 64.4 | 0.064 | 0.441 | 0 / 400 requests |
Accuracy in %, with its 95 % bootstrap interval. A rejected request (the model can’t take the question, or the input is too long for it) counts as wrong.
Speed and memory
| Device | p50 | p90 | Load | VRAM | RAM peak | Precision |
|---|---|---|---|---|---|---|
| RTX 5090 CUDA | 251 ms | 259 ms | 26.3 s | 12.1 GB | 3.2 GB | F32 |
| Threadripper 3960X CPU | 12.1 s | 12.2 s | 7.69 s | – | 9.7 GB | F32 |
Five-question requests over HTTP, one at a time, each a different short message. VRAM: the rise of the GPU’s used memory while the model is loaded. RAM peak: the runner process’s peak resident memory.
Parity with the authors’ code
| Device | Same decision | Max logit diff. | Result | Record |
|---|---|---|---|---|
| RTX 5090 CPU | 463/463 | 1.1e-5 | pass | results/runs/2026-10-25-parity-rtx5090-cpu-cob561.json |
| RTX 5090 CUDA | 463/463 | 1.6e-5 | pass | results/runs/2026-10-25-parity-rtx5090-cuda-cob561.json |
Reproduce
The tools are in convert/ollaya_convert and need only Python’s standard library. Run them against a local Ollaya 0.13.0 with the same model digest.
ollaya pull decision2:nox
# typed-decisions; export the test split once with
# uv run python -m ollaya_convert.bench_typed export --out typed.jsonl
python -m ollaya_convert.bench_typed run --model decision2:nox \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/decision2-nox
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model decision2:nox \
--url http://127.0.0.1:11435 --data public --out public/decision2-nox
# speed and memory
python -m ollaya_convert.bench_latency --models decision2:nox --memory \
--machine my-machine --device cuda --out latency.json