jeb:9btrained on split
Jebadiah · frontier-infra · 9B · llama.cpp · Apache-2.0
- Score
- 64.9
- 95 % interval 62.6–67.0
- Rank
- 1
- of 35 scored models
- Public benchmark
- 76.6
- macro accuracy, %
- typed-decisions
- 78.6
- 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 *
68.7
Classification
MASSIVE en
84.3
MASSIVE de
83.7
Civil Comments
72.0
Aegis 2
56.8
Understanding
MultiNLI
76.9
VitaminC
63.2
PAWS
65.6
Knowledge
BoolQ
71.3
SQuAD 2
59.9
PubMedQA
58.6
Judging
HelpSteer 2
29.2
SummEval rel.
37.0
SummEval cons.
85.2
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 76.6(75.3–77.9) | 80.4 | 82.6 | 60.4 | 0.039 | 0.311 | 0 / 3,880 |
| typed-decisions * | 78.6(76.9–80.5) | 72.7 | 84.8 | 78.5 | 0.127 | 0.314 | 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 | 107 ms | 108 ms | 20.8 s | 9.1 GB | 9.5 GB | Q8_0 |
| Threadripper 3960X CPU | 9.42 s | 9.71 s | 21.6 s | – | 9.2 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 | 7.7e-6 | pass | families/jebadiah.md:62 |
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:9b
# 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:9b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/jeb-9b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model jeb:9b \
--url http://127.0.0.1:11435 --data public --out public/jeb-9b
# speed and memory
python -m ollaya_convert.bench_latency --models jeb:9b --memory \
--machine my-machine --device cuda --out latency.json