gliclass-v4:2b
GLiClass v4 · Knowledgator · 2.4B · ONNX Runtime · Apache-2.0
- Score
- 59.1
- 95 % interval 57.2–60.9
- Rank
- 15
- of 66 scored models
- Public benchmark
- 73.1
- macro accuracy, %
- typed-decisions
- 58.9
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
39.8
Classification
MASSIVE en
81.5
MASSIVE de
78.2
Civil Comments
79.3
Aegis 2
60.8
Understanding
MultiNLI
71.9
VitaminC
57.4
PAWS
63.2
Knowledge
BoolQ
60.7
SQuAD 2
69.2
PubMedQA
51.4
Judging
HelpSteer 2
27.2
SummEval rel.
8.3
SummEval cons.
78.3
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 73.1(71.8–74.4) | 76.5 | 83.3 | 50.4 | 0.031 | 0.360 | 0 / 3,880 |
| typed-decisions | 58.9(56.5–61.6) | 59.3 | 66.0 | 53.3 | 0.062 | 0.532 | 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 | 82.2 ms | 83.7 ms | 50.9 s | 8.9 GB | 12.6 GB | F32 |
| Threadripper 3960X CPU | 3.57 s | 3.66 s | 51.9 s | – | 12.3 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 CPU | 483/483 | 9.4e-5 | pass | results/runs/2026-10-24-parity-rtx4090-cpu.json |
| RTX 4090 CUDA | 483/483 | 5.7e-4 | pass | results/runs/2026-10-24-parity-rtx4090-cuda.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 gliclass-v4:2b
# 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 gliclass-v4:2b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/gliclass-v4-2b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model gliclass-v4:2b \
--url http://127.0.0.1:11435 --data public --out public/gliclass-v4-2b
# speed and memory
python -m ollaya_convert.bench_latency --models gliclass-v4:2b --memory \
--machine my-machine --device cuda --out latency.json