von:1.1
Von · Victor Hugo Panisa · 395M · ONNX Runtime · Apache-2.0
- Score
- 23.7
- 95 % interval 21.5–26.0
- Rank
- 28
- of 35 scored models
- Public benchmark
- 48.5
- macro accuracy, %
- typed-decisions
- 44.6
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
18.9
Classification
MASSIVE en
67.3
MASSIVE de
22.3
Civil Comments
52.0
Aegis 2
25.6
Understanding
MultiNLI
-0.8
VitaminC
64.2
PAWS
15.2
Knowledge
BoolQ
43.3
SQuAD 2
0.3
PubMedQA
30.4
Judging
HelpSteer 2
-6.9
SummEval rel.
14.6
SummEval cons.
-14.6
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 48.5(47.1–49.9) | 51.6 | 63.6 | 18.1 | 0.098 | 0.625 | 0 / 3,880 |
| typed-decisions | 44.6(42.4–46.9) | 36.7 | 63.2 | 36.8 | 0.052 | 0.650 | 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 | 19.3 ms | 19.8 ms | 5.70 s | 2.6 GB | 2.1 GB | F32 |
| Threadripper 3960X CPU | 2.59 s | 2.60 s | 5.95 s | – | 1.7 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 |
|---|---|---|---|---|
| x86-64, 24 cores; ONNX Runtime 1.28 | 100 % | 4.4e-4 | pass | families/von.md:329-334 |
| RTX 4090, CUDA 13; ONNX Runtime 1.28 | 100 % | 3.9e-4 | pass | families/von.md:329-334 |
| Apple M4 Pro, macOS 27, MLX on Metal | 485/485 | 2.0e-3 | fail | families/von.md:396-398 |
| Apple silicon (M4 Pro), ONNX Runtime CPU | 100 % | 1.1e-3 | fail | families/von.md:404-405 |
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 von:1.1
# 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 von:1.1 \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/von-1.1
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model von:1.1 \
--url http://127.0.0.1:11435 --data public --out public/von-1.1
# speed and memory
python -m ollaya_convert.bench_latency --models von:1.1 --memory \
--machine my-machine --device cuda --out latency.json