winnow:12b
Winnow · EldanRing · 12B · llama.cpp · Apache-2.0
- Score
- 65.7
- 95 % interval 63.7–67.5
- Rank
- 1
- of 35 scored models
- Public benchmark
- 77.3
- macro accuracy, %
- typed-decisions
- 70.2
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
56.3
Classification
MASSIVE en
85.8
MASSIVE de
84.3
Civil Comments
74.0
Aegis 2
63.2
Understanding
MultiNLI
72.4
VitaminC
63.2
PAWS
66.4
Knowledge
BoolQ
74.7
SQuAD 2
75.3
PubMedQA
56.8
Judging
HelpSteer 2
25.7
SummEval rel.
37.0
SummEval cons.
84.4
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 77.3(76.0–78.5) | 80.0 | 85.4 | 59.2 | 0.141 | 0.362 | 0 / 3,880 |
| typed-decisions | 70.2(67.9–72.4) | 66.2 | 78.8 | 66.8 | 0.155 | 0.440 | 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 | 138 ms | 149 ms | 27.6 s | 15.5 GB | 12.2 GB | Q8_0 |
| Threadripper 3960X CPU | 11.2 s | 11.3 s | 32.4 s | – | 14.6 GB | Q8_0 |
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 (Linux) | 505/505 | 1.1e-5 | pass | families/winnow.md:122 |
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 winnow:12b
# 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 winnow:12b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/winnow-12b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model winnow:12b \
--url http://127.0.0.1:11435 --data public --out public/winnow-12b
# speed and memory
python -m ollaya_convert.bench_latency --models winnow:12b --memory \
--machine my-machine --device cuda --out latency.json