coba:400mrejects 8 %
coba · Ollaya · 421M · ONNX Runtime · Apache-2.0
- Score
- 35.8
- 95 % interval 33.7–37.9
- Rank
- 42
- of 66 scored models
- Public benchmark
- 56.1
- macro accuracy, %
- typed-decisions
- 59.1
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
40.1
Classification
MASSIVE en
78.8
MASSIVE de
54.0
Civil Comments
76.7
Aegis 2
8.8
Understanding
MultiNLI
64.4
VitaminC
58.7
PAWS
24.8
Knowledge
BoolQ
49.3
SQuAD 2
37.1
PubMedQA
39.4
Judging
HelpSteer 2
-7.9
SummEval rel.
-14.1
SummEval cons.
-8.5
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 56.1(54.6–57.5) | 69.0 | 69.7 | 11.9 | 0.089 | 0.441 | 453 / 3,880 |
| typed-decisions | 59.1(56.4–62.0) | 59.3 | 69.3 | 51.2 | 0.094 | 0.544 | 7 / 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 | 20.5 ms | 24.9 ms | 4.44 s | 1.6 GB | 2.1 GB | F16 |
| Threadripper 3960X CPU | 832 ms | 862 ms | 3.18 s | – | 1.9 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 |
|---|---|---|---|---|
| M4 Pro METAL | 2383/2383 | – | pass | results/runs/2026-10-23-parity-macm4pro-metal-coba.json |
| RTX 5090 CPU | 2383/2383 | – | pass | results/runs/2026-10-23-parity-rtx5090-cpu-coba.json |
| RTX 5090 CUDA | 2383/2383 | – | pass | results/runs/2026-10-23-parity-rtx5090-cuda-coba.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 coba:400m
# 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 coba:400m \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/coba-400m
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model coba:400m \
--url http://127.0.0.1:11435 --data public --out public/coba-400m
# speed and memory
python -m ollaya_convert.bench_latency --models coba:400m --memory \
--machine my-machine --device cuda --out latency.json