coba:150m-ternarytrained on splitrejects 8 %
coba · Ollaya · 164M · ONNX Runtime · Apache-2.0
- Score
- 4.6
- 95 % interval 1.7–7.9
- Rank
- 62
- of 66 scored models
- Public benchmark
- 62.2
- macro accuracy, %
- typed-decisions
- 53.5
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right. * trained on its train split, so not in the score.
Workflows
typed-decisions
31.9
Classification
MASSIVE en *
87.3
MASSIVE de *
30.4
Civil Comments *
83.3
Aegis 2 *
65.6
Understanding
MultiNLI *
73.9
VitaminC *
74.2
PAWS *
85.6
Knowledge
BoolQ *
46.0
SQuAD 2 *
54.5
PubMedQA *
42.4
Judging
HelpSteer 2 *
-4.4
SummEval rel.
-15.6
SummEval cons.
-2.4
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 62.2(61.0–63.4) | 69.9 | 83.5 | 14.0 | 0.053 | 0.336 | 453 / 3,880 |
| typed-decisions | 53.5(50.7–56.5) | 45.2 | 66.3 | 50.1 | 0.114 | 0.595 | 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 | 15.4 ms | 19.0 ms | 2.20 s | 1.1 GB | 1.0 GB | F16 |
| Threadripper 3960X CPU | 461 ms | 467 ms | 1.19 s | – | 801 MB | 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:150m-ternary
# 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:150m-ternary \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/coba-150m-ternary
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model coba:150m-ternary \
--url http://127.0.0.1:11435 --data public --out public/coba-150m-ternary
# speed and memory
python -m ollaya_convert.bench_latency --models coba:150m-ternary --memory \
--machine my-machine --device cuda --out latency.json