jeb:27btrained on split
Jebadiah · frontier-infra · 27B · llama.cpp · Apache-2.0
- Score
- 67.8
- 95 % interval 65.8–69.8
- Rank
- 1
- of 35 scored models
- Public benchmark
- 78.8
- macro accuracy, %
- typed-decisions
- 80.3
- 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 *
71.1
Classification
MASSIVE en
87.0
MASSIVE de
86.1
Civil Comments
70.0
Aegis 2
57.6
Understanding
MultiNLI
87.5
VitaminC
72.2
PAWS
65.6
Knowledge
BoolQ
78.0
SQuAD 2
52.5
PubMedQA
64.0
Judging
HelpSteer 2
32.2
SummEval rel.
45.8
SummEval cons.
83.5
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 78.8(77.5–80.2) | 84.7 | 82.4 | 63.1 | 0.024 | 0.283 | 0 / 3,880 |
| typed-decisions * | 80.3(78.3–82.3) | 76.0 | 86.8 | 78.5 | 0.133 | 0.309 | 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 | 253 ms | 254 ms | 36.3 s | 16.2 GB | 16.1 GB | Q4_K - MEDIUM |
| Threadripper 3960X CPU | 21.0 s | 21.4 s | 55.2 s | – | 26.6 GB | Q4_K - MEDIUM |
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:63 |
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:27b
# 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:27b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/jeb-27b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model jeb:27b \
--url http://127.0.0.1:11435 --data public --out public/jeb-27b
# speed and memory
python -m ollaya_convert.bench_latency --models jeb:27b --memory \
--machine my-machine --device cuda --out latency.json