laya:enrejects 8 %
Laya · Convai Innovations · 421M · ONNX Runtime · Apache-2.0
- Score
- 30.1
- 95 % interval 28.2–32.0
- Rank
- 25
- of 35 scored models
- Public benchmark
- 53.2
- macro accuracy, %
- typed-decisions
- 35.8
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
5.9
Classification
MASSIVE en
66.7
MASSIVE de
37.4
Civil Comments
87.3
Aegis 2
-14.4
Understanding
MultiNLI
81.4
VitaminC
68.5
PAWS
15.2
Knowledge
BoolQ
67.3
SQuAD 2
23.8
PubMedQA
25.0
Judging
HelpSteer 2
-5.4
SummEval rel.
-17.2
SummEval cons.
-19.8
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 53.2(52.0–54.5) | 65.2 | 67.9 | 8.7 | 0.106 | 0.480 | 453 / 3,880 |
| typed-decisions | 35.8(33.1–38.1) | 28.3 | 47.8 | 32.3 | 0.174 | 0.750 | 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 | 18.3 ms | 19.2 ms | 3.83 s | 1.6 GB | 1.2 GB | F16 |
| Threadripper 3960X CPU | 945 ms | 950 ms | 6.34 s | – | 3.1 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 |
|---|---|---|---|---|
| Apple M4 Pro (16-core GPU), macOS 27, MLX on Metal | 100% | – | pass | decisions/0001-mlx-engine.md:148 |
| Apple M4 Pro, macOS 27, ONNX Runtime CPU | 100% | – | pass | decisions/0001-mlx-engine.md:148 |
| Apple M4 Pro (16-core GPU), macOS 27, MLX on Metal | 100% | – | pass | decisions/0001-mlx-engine.md:149 |
| Apple M4 Pro, macOS 27, ONNX Runtime CPU | 100% | – | pass | decisions/0001-mlx-engine.md:149 |
| ONNX Runtime WebGPU provider on a Mac (rejected) | 1 of 483 differ | – | fail | decisions/0001-mlx-engine.md:13-14 |
| RTX 4090 (sm_89), Microsoft ORT 1.28.2 cuda13 + CUDA 13.x/cuDNN 9 libs, fp32 graph | 100.00% | – | pass | decisions/0004-cuda-onnxruntime-builds.md:73 |
| RTX 4090 (sm_89), Microsoft ORT 1.28.2 cuda12 + CUDA 12.8 wheels, cuDNN 9.26, fp32 graph | 100.00% | – | pass | decisions/0004-cuda-onnxruntime-builds.md:75 |
| RTX 4090 under WSL 2, installed cuda_v13 pack (install.sh layout) | 100% | – | pass | distribution.md:407 |
| RTX 4090, Windows 11 (driver 616.92, CUDA UMD 13.4), release.yml dry-run archives | 100% | – | pass | distribution.md:421 |
| Windows 11 (the new Windows build) | 100% | – | pass | distribution.md:423 |
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 laya:en
# 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 laya:en \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/laya-en
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model laya:en \
--url http://127.0.0.1:11435 --data public --out public/laya-en
# speed and memory
python -m ollaya_convert.bench_latency --models laya:en --memory \
--machine my-machine --device cuda --out latency.json