coba:68m-ternarytrained on splitrejects 8 %
coba · Ollaya · 75M · ONNX Runtime · Apache-2.0
- Score
- 1.6
- 95 % interval -1.3–4.9
- Rank
- 64
- of 66 scored models
- Public benchmark
- 60.3
- macro accuracy, %
- typed-decisions
- 48.3
- 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
24.2
Classification
MASSIVE en *
88.8
MASSIVE de *
27.1
Civil Comments *
85.3
Aegis 2 *
53.6
Understanding
MultiNLI *
72.9
VitaminC *
69.5
PAWS *
78.4
Knowledge
BoolQ *
39.3
SQuAD 2 *
43.8
PubMedQA *
41.2
Judging
HelpSteer 2 *
-4.4
SummEval rel.
-15.1
SummEval cons.
-4.2
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 60.3(59.0–61.7) | 68.6 | 80.0 | 13.7 | 0.076 | 0.381 | 453 / 3,880 |
| typed-decisions | 48.3(46.0–50.5) | 42.0 | 65.0 | 40.4 | 0.092 | 0.620 | 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 | 13.6 ms | 17.4 ms | 1.70 s | 899 MB | 934 MB | F16 |
| Threadripper 3960X CPU | 314 ms | 316 ms | 652 ms | – | 370 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:68m-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:68m-ternary \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/coba-68m-ternary
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model coba:68m-ternary \
--url http://127.0.0.1:11435 --data public --out public/coba-68m-ternary
# speed and memory
python -m ollaya_convert.bench_latency --models coba:68m-ternary --memory \
--machine my-machine --device cuda --out latency.json