Skip to content

decima:base-cleanrejects 6 %

Decima · A. M. Madani · 321M · ONNX Runtime · Apache-2.0

Score
23.5
95 % interval 21.3–25.6
Rank
53
of 66 scored models
Public benchmark
49.2
macro accuracy, %
typed-decisions
47.4
accuracy, %

By task

Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.

Workflows
typed-decisions
23.0
Classification
MASSIVE en
64.9
MASSIVE de
56.4
Civil Comments
46.0
Aegis 2
0.0
Understanding
MultiNLI
71.9
VitaminC
39.4
PAWS
18.4
Knowledge
BoolQ
31.3
SQuAD 2
-1.7
PubMedQA
2.8
Judging
HelpSteer 2
-11.4
SummEval rel.
3.6
SummEval cons.
-16.3

Accuracy and calibration

BenchmarkAccuracyChoiceNoulScoreECEBrierRejected
Public benchmark49.2(47.8–50.5)60.459.413.60.0670.522301 / 3,880
typed-decisions47.4(45.0–50.0)37.559.046.30.0610.6197 / 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

Devicep50p90LoadVRAMRAM peakPrecision
RTX 5090 CUDA26.4 ms33.1 ms3.36 s2.9 GB1.8 GBF32
Threadripper 3960X CPU1.20 s1.27 s4.44 s–1.1 GBF32

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

DeviceSame decisionMax logit diff.ResultRecord
RTX 5090 CPU599/5993.2e-5passresults/runs/2026-10-24-parity-rtx5090-cpu-cob560.json
RTX 5090 CUDA599/5991.8e-5passresults/runs/2026-10-24-parity-rtx5090-cuda-cob560.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:base-clean

# 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:base-clean \
  --url http://127.0.0.1:11435 --data typed.jsonl --out typed/decima-base-clean

# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model decima:base-clean \
  --url http://127.0.0.1:11435 --data public --out public/decima-base-clean

# speed and memory
python -m ollaya_convert.bench_latency --models decima:base-clean --memory \
  --machine my-machine --device cuda --out latency.json