nli:deberta-v3-largerejects 6 %
NLI zero-shot · Moritz Laurer · 435M · ONNX Runtime · MIT
- Score
- 21.0
- 95 % interval 19.0–23.1
- Rank
- 29
- of 35 scored models
- Public benchmark
- 45.8
- macro accuracy, %
- typed-decisions
- 53.6
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
32.0
Classification
MASSIVE en
57.6
MASSIVE de
51.9
Civil Comments
78.0
Aegis 2
22.4
Understanding
MultiNLI
-2.8
VitaminC
-10.9
PAWS
19.2
Knowledge
BoolQ
49.3
SQuAD 2
-5.0
PubMedQA
24.4
Judging
HelpSteer 2
-3.9
SummEval rel.
3.1
SummEval cons.
-21.5
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 45.8(44.5–47.3) | 44.3 | 66.4 | 14.1 | 0.196 | 0.635 | 326 / 3,880 |
| typed-decisions | 53.6(51.2–56.0) | 56.3 | 65.0 | 43.0 | 0.096 | 0.589 | 8 / 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 | 17.7 ms | 18.0 ms | 8.22 s | 2.6 GB | 4.0 GB | F32 |
| Threadripper 3960X CPU | 2.32 s | 2.44 s | 7.71 s | – | 3.8 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 with a CUDA 13 driver, CUDA 12 pack | 100% | – | pass | distribution.md:126-129 |
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 nli:deberta-v3-large
# 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 nli:deberta-v3-large \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/nli-deberta-v3-large
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model nli:deberta-v3-large \
--url http://127.0.0.1:11435 --data public --out public/nli-deberta-v3-large
# speed and memory
python -m ollaya_convert.bench_latency --models nli:deberta-v3-large --memory \
--machine my-machine --device cuda --out latency.json