decima:guard-small
Decima · A. M. Madani · 140M · ONNX Runtime · Apache-2.0
- Score
- 14.8
- 95 % interval 12.3–16.9
- Rank
- 35
- of 41 scored models
- Public benchmark
- 42.4
- macro accuracy, %
- typed-decisions
- 40.0
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
12.1
Classification
MASSIVE en
30.7
MASSIVE de
18.6
Civil Comments
44.0
Aegis 2
47.2
Understanding
MultiNLI
49.8
VitaminC
24.9
PAWS
-4.0
Knowledge
BoolQ
-11.3
SQuAD 2
-0.3
PubMedQA
-2.6
Judging
HelpSteer 2
-11.4
SummEval rel.
-0.5
SummEval cons.
9.7
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 42.4(40.9–43.9) | 41.2 | 57.6 | 19.4 | 0.140 | 0.677 | 0 / 3,880 |
| typed-decisions | 40.0(37.8–42.3) | 35.8 | 56.0 | 31.1 | 0.184 | 0.727 | 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 | 26.6 ms | 32.7 ms | 2.64 s | 1.2 GB | 1.1 GB | F32 |
| Threadripper 3960X CPU | 721 ms | 739 ms | 2.09 s | – | 529 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 |
|---|---|---|---|---|
| RTX 4090 CPU | 599/599 | 1.8e-5 | pass | results/runs/2026-10-20-parity-rtx4090-cpu-decima.json |
| RTX 4090 CUDA | 599/599 | 9.1e-6 | 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-small
# 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-small \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/decima-guard-small
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model decima:guard-small \
--url http://127.0.0.1:11435 --data public --out public/decima-guard-small
# speed and memory
python -m ollaya_convert.bench_latency --models decima:guard-small --memory \
--machine my-machine --device cuda --out latency.json