decision2:eos
Decision 2.0 · vLLM Semantic Router · 0.75B · ONNX Runtime · Apache-2.0
- Score
- 47.0
- 95 % interval 44.8–49.0
- Rank
- 32
- of 66 scored models
- Public benchmark
- 66.1
- macro accuracy, %
- typed-decisions
- 46.0
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
20.9
Classification
MASSIVE en
73.1
MASSIVE de
70.7
Civil Comments
64.7
Aegis 2
45.6
Understanding
MultiNLI
77.9
VitaminC
65.7
PAWS
16.8
Knowledge
BoolQ
49.3
SQuAD 2
53.2
PubMedQA
29.8
Judging
HelpSteer 2
18.7
SummEval rel.
-5.7
SummEval cons.
77.4
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 66.1(64.8–67.6) | 72.5 | 73.0 | 44.1 | 0.093 | 0.447 | 0 / 3,880 |
| typed-decisions | 46.0(43.5–48.5) | 43.0 | 59.8 | 37.9 | 0.211 | 0.710 | 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 | 68.5 ms | 76.6 ms | 18.7 s | 5.8 GB | 5.9 GB | F32 |
| Threadripper 3960X CPU | 3.23 s | 3.24 s | 14.9 s | – | 4.6 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 | 2.0e-5 | pass | results/runs/2026-10-25-parity-rtx5090-cpu-cob561.json |
| RTX 5090 CUDA | 463/463 | 1.3e-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:eos
# 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:eos \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/decision2-eos
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model decision2:eos \
--url http://127.0.0.1:11435 --data public --out public/decision2-eos
# speed and memory
python -m ollaya_convert.bench_latency --models decision2:eos --memory \
--machine my-machine --device cuda --out latency.json