Skip to content

winnow:e2b

Winnow · EldanRing · 4.6B · llama.cpp · Apache-2.0

Score
49.4
95 % interval 47.3–51.6
Rank
31
of 66 scored models
Public benchmark
65.9
macro accuracy, %
typed-decisions
61.6
accuracy, %

By task

Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.

Workflows
typed-decisions
43.7
Classification
MASSIVE en
77.9
MASSIVE de
70.0
Civil Comments
71.3
Aegis 2
63.2
Understanding
MultiNLI
57.9
VitaminC
38.6
PAWS
60.0
Knowledge
BoolQ
52.0
SQuAD 2
49.8
PubMedQA
17.8
Judging
HelpSteer 2
16.7
SummEval rel.
26.6
SummEval cons.
46.2

Accuracy and calibration

BenchmarkAccuracyChoiceNoulScoreECEBrierRejected
Public benchmark65.9(64.3–67.3)65.479.643.80.1910.5140 / 3,880
typed-decisions61.6(59.2–64.0)59.572.055.40.1460.5320 / 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

Devicep50p90LoadVRAMRAM peakPrecision
RTX 5090 CUDA104 ms106 ms3.86 s3.5 GB5.0 GBQ8_0
Threadripper 3960X CPU2.04 s2.06 s10.5 s–4.9 GBQ8_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

Its parity check is in the model’s family documentation in the repository (docs/families/); the compiled parity file does not list it yet.

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:e2b

# 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:e2b \
  --url http://127.0.0.1:11435 --data typed.jsonl --out typed/winnow-e2b

# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model winnow:e2b \
  --url http://127.0.0.1:11435 --data public --out public/winnow-e2b

# speed and memory
python -m ollaya_convert.bench_latency --models winnow:e2b --memory \
  --machine my-machine --device cuda --out latency.json