decision:eos
Decision 1.0 · vLLM Semantic Router · 0.75B · ONNX Runtime · Apache-2.0
- Score
- 38.1
- 95 % interval 35.8–40.2
- Rank
- 21
- of 35 scored models
- Public benchmark
- 59.1
- macro accuracy, %
- typed-decisions
- 51.2
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
28.6
Classification
MASSIVE en
80.6
MASSIVE de
74.0
Civil Comments
30.7
Aegis 2
37.6
Understanding
MultiNLI
65.4
VitaminC
43.7
PAWS
34.4
Knowledge
BoolQ
36.7
SQuAD 2
74.6
PubMedQA
5.8
Judging
HelpSteer 2
7.1
SummEval rel.
2.6
SummEval cons.
11.5
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 59.1(57.6–60.5) | 66.7 | 71.4 | 25.7 | 0.087 | 0.478 | 0 / 3,880 |
| typed-decisions | 51.2(48.0–54.4) | 49.2 | 60.3 | 46.0 | 0.121 | 0.653 | 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 | 209 ms | 210 ms | 13.3 s | 5.8 GB | 6.0 GB | F32 |
| Threadripper 3960X CPU | 3.51 s | 4.30 s | 14.3 s | – | 6.6 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 |
|---|---|---|---|---|
| 24 cores; ort 2.0.0-rc.13, ONNX Runtime 1.28 | 100 % | 3.6e-5 | pass | families/decision.md:211-214 |
| RTX 4090, CUDA 13; ONNX Runtime 1.28 | 100 % | 1.6e-5 | pass | families/decision.md:211-214 |
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 decision: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 decision:eos \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/decision-eos
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model decision:eos \
--url http://127.0.0.1:11435 --data public --out public/decision-eos
# speed and memory
python -m ollaya_convert.bench_latency --models decision:eos --memory \
--machine my-machine --device cuda --out latency.json