kev:0.8b
Kev · Jared Palmer · 0.76B · ONNX Runtime · Apache-2.0
- Score
- 40.8
- 95 % interval 38.7–42.9
- Rank
- 21
- of 35 scored models
- Public benchmark
- 61.1
- 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
20.8
Classification
MASSIVE en
73.4
MASSIVE de
64.0
Civil Comments
67.3
Aegis 2
38.4
Understanding
MultiNLI
69.9
VitaminC
49.4
PAWS
32.0
Knowledge
BoolQ
58.7
SQuAD 2
32.4
PubMedQA
17.2
Judging
HelpSteer 2
11.1
SummEval rel.
-7.8
SummEval cons.
43.6
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 61.1(59.7–62.5) | 66.4 | 72.9 | 32.5 | 0.035 | 0.465 | 0 / 3,880 |
| typed-decisions | 46.0(43.3–48.4) | 50.5 | 52.5 | 37.6 | 0.046 | 0.606 | 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 | 142 ms | 143 ms | 11.9 s | 4.5 GB | 5.2 GB | F32 |
| Threadripper 3960X CPU | 2.91 s | 2.91 s | 12.5 s | – | 5.0 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 |
|---|---|---|---|---|
| i9-13900K; ONNX Runtime 1.28 | 100 % | 4.6e-5 | pass | families/kev.md:238 |
| RTX 4090, CUDA 13; ONNX Runtime 1.28 | 100 % | 3.4e-5 | pass | families/kev.md:239 |
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 kev:0.8b
# 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 kev:0.8b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/kev-0.8b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model kev:0.8b \
--url http://127.0.0.1:11435 --data public --out public/kev-0.8b
# speed and memory
python -m ollaya_convert.bench_latency --models kev:0.8b --memory \
--machine my-machine --device cuda --out latency.json