decision2:kai
Decision 2.0 · vLLM Semantic Router · 0.60B · ONNX Runtime · Apache-2.0
- Score
- 38.0
- 95 % interval 35.7–40.2
- Rank
- 40
- of 66 scored models
- Public benchmark
- 60.8
- macro accuracy, %
- typed-decisions
- 47.6
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
23.3
Classification
MASSIVE en
77.3
MASSIVE de
73.1
Civil Comments
20.0
Aegis 2
24.8
Understanding
MultiNLI
70.4
VitaminC
47.7
PAWS
0.8
Knowledge
BoolQ
42.0
SQuAD 2
49.2
PubMedQA
20.2
Judging
HelpSteer 2
21.7
SummEval rel.
0.0
SummEval cons.
61.8
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 60.8(59.4–62.2) | 69.1 | 63.7 | 42.3 | 0.046 | 0.485 | 0 / 3,880 |
| typed-decisions | 47.6(45.1–50.2) | 42.3 | 63.5 | 39.8 | 0.061 | 0.619 | 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 | 30.8 ms | 44.0 ms | 11.3 s | 3.6 GB | 4.8 GB | F32 |
| Threadripper 3960X CPU | 1.80 s | 1.81 s | 10.5 s | – | 4.3 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 | 2.7e-5 | pass | results/runs/2026-10-25-parity-rtx5090-cpu-cob561.json |
| RTX 5090 CUDA | 463/463 | 3.0e-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:kai
# 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:kai \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/decision2-kai
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model decision2:kai \
--url http://127.0.0.1:11435 --data public --out public/decision2-kai
# speed and memory
python -m ollaya_convert.bench_latency --models decision2:kai --memory \
--machine my-machine --device cuda --out latency.json