winnow:e4b
Winnow · EldanRing · 7.5B · llama.cpp · Apache-2.0
- Score
- 59.9
- 95 % interval 57.9–61.9
- Rank
- 6
- of 35 scored models
- Public benchmark
- 73.4
- macro accuracy, %
- typed-decisions
- 72.4
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
59.6
Classification
MASSIVE en
84.3
MASSIVE de
79.7
Civil Comments
50.7
Aegis 2
59.2
Understanding
MultiNLI
74.9
VitaminC
54.4
PAWS
61.6
Knowledge
BoolQ
63.3
SQuAD 2
63.2
PubMedQA
53.8
Judging
HelpSteer 2
22.2
SummEval rel.
31.3
SummEval cons.
80.9
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 73.4(72.1–74.8) | 77.6 | 79.8 | 55.8 | 0.071 | 0.380 | 0 / 3,880 |
| typed-decisions | 72.4(70.0–74.7) | 69.8 | 80.2 | 68.5 | 0.020 | 0.394 | 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 | 113 ms | 119 ms | 17.5 s | 6.2 GB | 7.9 GB | Q8_0 |
| Threadripper 3960X CPU | 4.10 s | 4.15 s | 19.0 s | – | 8.1 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 | 1.3e-5 | pass | families/winnow.md:123 |
| x86-64 Linux | 505/505 | 1.1e-5 | pass | families/winnow.md:124 |
| Windows x86-64 | 15/15 | 7.6e-6 | pass | families/winnow.md:125 |
| RTX 4090, Windows 11 (ggml-cuda.dll in the 0.7.4 CUDA pack) | 505/505 | 1.3e-5 | pass | families/winnow.md:126 |
| RTX 4090, Windows, llama.cpp win-vulkan-x64 of the same build (#27) | 505/505 | 1.1e-5 | pass | families/winnow.md:133-137 |
| RTX 4090 with a CUDA 13 driver, llama.cpp's CUDA 12 backend (CUDA 12 pack) | 505/505 | 1.3e-5 | pass | distribution.md:126-130 |
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:e4b
# 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:e4b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/winnow-e4b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model winnow:e4b \
--url http://127.0.0.1:11435 --data public --out public/winnow-e4b
# speed and memory
python -m ollaya_convert.bench_latency --models winnow:e4b --memory \
--machine my-machine --device cuda --out latency.json