decima:guard
Decima · A. M. Madani · 321M · ONNX Runtime · Apache-2.0
- Score
- 25.8
- 95 % interval 23.7–28.1
- Rank
- 31
- of 41 scored models
- Public benchmark
- 51.8
- macro accuracy, %
- typed-decisions
- 40.4
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
12.7
Classification
MASSIVE en
54.3
MASSIVE de
46.1
Civil Comments
40.7
Aegis 2
49.6
Understanding
MultiNLI
69.9
VitaminC
37.6
PAWS
-2.4
Knowledge
BoolQ
-4.7
SQuAD 2
-2.3
PubMedQA
13.6
Judging
HelpSteer 2
-11.9
SummEval rel.
-9.4
SummEval cons.
67.0
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 51.8(50.3–53.3) | 57.4 | 58.1 | 32.2 | 0.126 | 0.611 | 0 / 3,880 |
| typed-decisions | 40.4(38.0–43.0) | 32.8 | 49.3 | 39.4 | 0.214 | 0.733 | 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 | 26.3 ms | 32.9 ms | 3.28 s | 2.7 GB | 1.8 GB | F32 |
| Threadripper 3960X CPU | 1.24 s | 1.28 s | 3.16 s | – | 1.1 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 |
|---|---|---|---|---|
| RTX 4090 CPU | 599/599 | 1.8e-5 | pass | results/runs/2026-10-20-parity-rtx4090-cpu-decima.json |
| RTX 4090 CUDA | 599/599 | 2.0e-5 | pass | results/runs/2026-10-20-parity-rtx4090-cuda-decima.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 decima:guard
# 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 decima:guard \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/decima-guard
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model decima:guard \
--url http://127.0.0.1:11435 --data public --out public/decima-guard
# speed and memory
python -m ollaya_convert.bench_latency --models decima:guard --memory \
--machine my-machine --device cuda --out latency.json