kev:4b
Kev · Jared Palmer · 4.2B · ONNX Runtime · Apache-2.0
- Score
- 61.2
- 95 % interval 59.2–63.2
- Rank
- 4
- of 35 scored models
- Public benchmark
- 74.5
- macro accuracy, %
- typed-decisions
- 66.9
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
51.5
Classification
MASSIVE en
85.5
MASSIVE de
85.2
Civil Comments
63.3
Aegis 2
62.4
Understanding
MultiNLI
87.5
VitaminC
58.2
PAWS
63.2
Knowledge
BoolQ
70.7
SQuAD 2
56.5
PubMedQA
57.4
Judging
HelpSteer 2
17.7
SummEval rel.
21.4
SummEval cons.
76.6
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 74.5(73.1–75.8) | 81.5 | 81.6 | 50.8 | 0.062 | 0.359 | 0 / 3,880 |
| typed-decisions | 66.9(64.6–69.0) | 64.5 | 76.5 | 61.5 | 0.094 | 0.465 | 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 | 319 ms | 320 ms | 21.6 s | 10.9 GB | 2.0 GB | F32 |
| Threadripper 3960X CPU | 8.51 s | 8.52 s | 22.5 s | – | 8.9 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 % | 2.3e-4 | pass | families/kev.md:240 |
| RTX 4090, CUDA 13; ONNX Runtime 1.28 | 100 % | 2.2e-4 | pass | families/kev.md:241 |
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:4b
# 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:4b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/kev-4b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model kev:4b \
--url http://127.0.0.1:11435 --data public --out public/kev-4b
# speed and memory
python -m ollaya_convert.bench_latency --models kev:4b --memory \
--machine my-machine --device cuda --out latency.json