laya:multilingual
Laya · Convai Innovations · 322M · ONNX Runtime · Apache-2.0
- Score
- 36.7
- 95 % interval 34.7–38.9
- Rank
- 21
- of 35 scored models
- Public benchmark
- 57.9
- macro accuracy, %
- typed-decisions
- 34.9
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
4.6
Classification
MASSIVE en
60.1
MASSIVE de
41.0
Civil Comments
84.0
Aegis 2
20.0
Understanding
MultiNLI
76.9
VitaminC
64.4
PAWS
53.6
Knowledge
BoolQ
50.7
SQuAD 2
20.4
PubMedQA
28.0
Judging
HelpSteer 2
7.6
SummEval rel.
-9.4
SummEval cons.
12.3
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 57.9(56.4–59.3) | 63.9 | 72.9 | 22.8 | 0.155 | 0.536 | 27 / 3,880 |
| typed-decisions | 34.9(32.8–37.1) | 29.0 | 49.7 | 28.2 | 0.316 | 0.890 | 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 | 14.3 ms | 14.6 ms | 3.62 s | 1.4 GB | 1.2 GB | F16 |
| Threadripper 3960X CPU | 516 ms | 517 ms | 4.78 s | – | 1.9 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:150 |
| RTX 4090 (sm_89), Microsoft ORT 1.28.2 cuda13, fp32 graph | 100.00% | – | pass | decisions/0004-cuda-onnxruntime-builds.md:74 |
| RTX 4090 (sm_89), Microsoft ORT 1.28.2 cuda12, fp32 graph | 100.00% | – | pass | decisions/0004-cuda-onnxruntime-builds.md:76 |
| RTX 4090 under WSL 2, installed cuda_v13 pack (install.sh layout) | 100% | – | pass | distribution.md:407 |
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:multilingual
# 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:multilingual \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/laya-multilingual
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model laya:multilingual \
--url http://127.0.0.1:11435 --data public --out public/laya-multilingual
# speed and memory
python -m ollaya_convert.bench_latency --models laya:multilingual --memory \
--machine my-machine --device cuda --out latency.json