clm:8b
CLM · Contrastive-LM · 8.2B · ONNX Runtime · Apache-2.0
- Score
- 9.3
- 95 % interval 7.1–11.3
- Rank
- 35
- of 35 scored models
- Public benchmark
- 37.7
- macro accuracy, %
- typed-decisions
- 35.7
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
5.8
Classification
MASSIVE en
2.6
MASSIVE de
0.5
Civil Comments
68.7
Aegis 2
-15.2
Understanding
MultiNLI
49.8
VitaminC
8.1
PAWS
3.2
Knowledge
BoolQ
12.7
SQuAD 2
4.3
PubMedQA
10.0
Judging
HelpSteer 2
-0.9
SummEval rel.
3.6
SummEval cons.
-23.3
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 37.7(36.3–39.0) | 31.9 | 57.4 | 14.5 | 0.372 | 0.916 | 0 / 3,880 |
| typed-decisions | 35.7(33.5–38.0) | 24.8 | 50.0 | 33.1 | 0.485 | 1.026 | 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 | 94.5 ms | 95.1 ms | 31.9 s | 15.3 GB | 1.8 GB | F32 |
| Threadripper 3960X CPU | 5.23 s | 5.55 s | 36.1 s | – | 14.5 GB | F32 |
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 | 100 % | 1.3e-4 | pass | families/clm.md:55 |
| 24 cores | 100 % | 1.2e-4 | pass | families/clm.md:56 |
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 clm:8b
# 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 clm:8b \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/clm-8b
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model clm:8b \
--url http://127.0.0.1:11435 --data public --out public/clm-8b
# speed and memory
python -m ollaya_convert.bench_latency --models clm:8b --memory \
--machine my-machine --device cuda --out latency.json