Datasets:

witcheer
/

rtx-5090-benchmarks

All benchmarks use generative evaluation via llama-server chat completions. Multiple-choice tasks (MMLU, ARC, HellaSwag) use letter extraction instead of loglikelihood scoring -- results are internally consistent for model comparison but absolute scores may differ from logprob-based evaluations by 5-15%.

Summary

Benchmark	Score	Metric
MMLU	83.69%	accuracy
ARC-Challenge	95.99%	accuracy
HellaSwag	89.32%	accuracy
HumanEval	93.29%	pass@1
GSM8K	95.98%	exact_match

MMLU Breakdown by Category

Category	Score	Correct / Total
Stem	84.07%	1,267 / 1,507
Humanities	83.38%	1,319 / 1,582
Social Sciences	90.62%	1,498 / 1,653
Other	78.62%	1,783 / 2,268

Sampled at 50% (seed 42)

Speed Benchmarks

Measured with llama-bench. All layers GPU-offloaded (-ngl 99).

Prompt Processing (tokens/s)

Context Length	Speed	+/-sigma
128	2,387	29.0
512	4,439	37.6
2048	4,432	21.0
4096	4,369	10.4
8192	4,275	21.3
16384	4,024	9.7

Generation (tokens/s)

Metric	Speed	+/-sigma
tg128	224.7	1.5

Methodology

Evaluation Framework

Custom generative evaluators built for this rig. All benchmarks run through llama-server's /v1/chat/completions endpoint.

Scoring: Generative evaluation (not loglikelihood)
Thinking: disabled
MCQ scoring: First valid letter extracted from response (A/B/C/D)
Sampling: 50% of dataset used
Temperature: 0 (deterministic)
Max tokens: 2,048
GPU offload: All layers (-ngl 99)

Benchmarked by WITCHEER on the RTX 5090 Benchmark Rig. Source: github.com/notwitcheer/llm-bench-rig/blob/main/reports/qwen3-coder-next-ud-q2-k-xl.md. Dataset: huggingface.co/datasets/witcheer/rtx-5090-benchmarks/blob/main/reports/qwen3-coder-next-ud-q2-k-xl.md.