jeb:4btrained on split
Jebadiah · frontier-infra · 4B · llama.cpp · Apache-2.0
- Score
- 63.9
- 95 % interval 61.6–65.9
- Rank
- 1
- of 35 scored models
- Public benchmark
- 75.9
- macro accuracy, %
- typed-decisions
- 80.0
- trained on its train split
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right. * trained on its train split, so not in the score.
Workflows
typed-decisions *
70.6
Classification
MASSIVE en
84.9
MASSIVE de
82.2
Civil Comments
74.0
Aegis 2
58.4
Understanding
MultiNLI
81.4
VitaminC
63.9
PAWS
66.4
Knowledge
BoolQ
70.0
SQuAD 2
52.5
PubMedQA
53.2
Judging
HelpSteer 2
24.2
SummEval rel.
39.1
SummEval cons.
80.0
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 75.9(74.6–77.3) | 80.3 | 82.1 | 58.2 | 0.048 | 0.329 | 0 / 3,880 |
| typed-decisions * | 80.0(77.8–82.0) | 76.3 | 85.0 | 78.9 | 0.136 | 0.312 | 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 | 90.7 ms | 92.4 ms | 10.5 s | 5.4 GB | 4.7 GB | Q8_0 |
| Threadripper 3960X CPU | 5.33 s | 5.44 s | 11.1 s | – | 4.5 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 | 494/494 | 7.6e-6 | pass | families/jebadiah.md:61 |
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 jeb:4b
# 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 jeb:4b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/jeb-4b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model jeb:4b \
--url http://127.0.0.1:11435 --data public --out public/jeb-4b
# speed and memory
python -m ollaya_convert.bench_latency --models jeb:4b --memory \
--machine my-machine --device cuda --out latency.json