Datasets:

mv1137
/

p2026-003-qwen38-flash-gpt-oss-throughput-results

completion_tokens int64 116 4.91k ⌀	correct bool 2 classes	errored bool 2 classes	expected_answer stringclasses 23 values	expected_norm stringclasses 23 values	id stringlengths 8 31	label stringclasses 2 values	predicted_answer stringlengths 1 35 ⌀	predicted_norm stringclasses 35 values	suite stringclasses 3 values	thinking_tokens_estimate int64 0 3.03k	truncated bool 1 class	wall_s float64 14.8 1.2k
4,905	true	false	\pi	\pi	math_precalculus_1199	FLASH-IQ1S-effmed	$\pi$	\pi	math25	3,027	false	546.33
915	false	false	5	5	math_algebra_1837	FLASH-IQ1S-effmed	x = 5	x=5	math25	365	false	143.3
753	false	false	x=5	x=5	math_algebra_2193	FLASH-IQ1S-effmed	5	5	math25	301	false	113.99
199	false	false	11\sqrt2	11\sqrt2	math_prealgebra_1302	FLASH-IQ1S-effmed	$11\sqrt{2}$	11\sqrt{2}	math25	58	false	31.62
460	true	false	1.25	1.25	math_prealgebra_1558	FLASH-IQ1S-effmed	1.25	1.25	math25	167	false	86.5
1,029	false	false	2516_8	2516_8	math_number_theory_516	FLASH-IQ1S-effmed	2516	2516	math25	405	false	123.02
1,330	true	false	\frac{3}{2}	3/2	math_algebra_2700	FLASH-IQ1S-effmed	3/2	3/2	math25	748	false	197.66
1,734	true	false	120^\circ	120	math_precalculus_44	FLASH-IQ1S-effmed	120°	120	math25	513	false	241.13
1,665	true	false	8	8	math_geometry_283	FLASH-IQ1S-effmed	8	8	math25	678	false	233.84
1,466	true	false	\frac{3840}{289}	3840/289	math_geometry_1140	FLASH-IQ1S-effmed	3840/289	3840/289	math25	522	false	210.73
930	true	false	144	144	math_precalculus_986	FLASH-IQ1S-effmed	144	144	math25	283	false	162.35
2,196	true	false	\frac{2}{21}	2/21	math_counting_&_probability_230	FLASH-IQ1S-effmed	2/21	2/21	math25	1,089	false	277.97
741	true	false	\frac{1}{4}	1/4	math_counting_&_probability_803	FLASH-IQ1S-effmed	$\frac{1}{4}$	1/4	math25	369	false	125.57
3,178	true	false	160	160	math_geometry_702	FLASH-IQ1S-effmed	160	160	math25	1,466	false	445.53
4,151	true	false	440	440	math_intermediate_algebra_1411	FLASH-IQ1S-effmed	440	440	math25	1,697	false	554.54
865	true	false	103	103	math_number_theory_1172	FLASH-IQ1S-effmed	103	103	math25	344	false	147.88
836	false	false	\frac9{19}	\frac9{19}	math_algebra_2257	FLASH-IQ1S-effmed	9/19	9/19	math25	215	false	142.5
1,030	true	false	3	3	math_number_theory_911	FLASH-IQ1S-effmed	3	3	math25	314	false	173.1
4,039	true	false	3	3	math_algebra_2176	FLASH-IQ1S-effmed	3	3	math25	2,161	false	1,200.67
217	true	false	4	4	math_prealgebra_1973	FLASH-IQ1S-effmed	4	4	math25	89	false	53.62
984	true	false	(-\sqrt{3}, \sqrt{3})	(-\sqrt{3},\sqrt{3})	math_algebra_1078	FLASH-IQ1S-effmed	$(-\sqrt{3}, \sqrt{3})$	(-\sqrt{3},\sqrt{3})	math25	331	false	200.38
1,330	false	false	\frac{20000}{\pi}	20000/\pi	math_intermediate_algebra_1111	FLASH-IQ1S-effmed	20000/π	20000/π	math25	594	false	238.9
3,239	true	false	406	406	math_intermediate_algebra_2146	FLASH-IQ1S-effmed	406	406	math25	1,606	false	595.45
null	false	true	null	null	math_counting_&_probability_731	FLASH-IQ1S-effmed	null	null	math25	0	false	1,202.21
994	true	false	(2,12) \cup (12,102)	(2,12)\cup(12,102)	math_intermediate_algebra_1981	FLASH-IQ1S-effmed	$(2, 12) \cup (12, 102)$	(2,12)\cup(12,102)	math25	336	false	232.34
494	true	false	null	null	gsm8k_13	FLASH-IQ1S-effmed	null	null	gsm8k	188	false	145.94
204	true	false	null	null	gsm8k_51	FLASH-IQ1S-effmed	null	null	gsm8k	115	false	64.61
314	true	false	null	null	gsm8k_54	FLASH-IQ1S-effmed	null	null	gsm8k	167	false	121.2
348	true	false	null	null	gsm8k_61	FLASH-IQ1S-effmed	null	null	gsm8k	130	false	126.72
304	true	false	null	null	gsm8k_65	FLASH-IQ1S-effmed	null	null	gsm8k	157	false	99.64
240	true	false	null	null	gsm8k_88	FLASH-IQ1S-effmed	null	null	gsm8k	65	false	86.63
1,199	true	false	null	null	gsm8k_93	FLASH-IQ1S-effmed	null	null	gsm8k	674	false	348.79
901	true	false	null	null	gsm8k_114	FLASH-IQ1S-effmed	null	null	gsm8k	582	false	301.46
246	true	false	null	null	gsm8k_135	FLASH-IQ1S-effmed	null	null	gsm8k	100	false	94.17
214	true	false	null	null	gsm8k_142	FLASH-IQ1S-effmed	null	null	gsm8k	123	false	74.81
503	true	false	null	null	gsm8k_146	FLASH-IQ1S-effmed	null	null	gsm8k	252	false	166.92
1,111	false	false	null	null	gsm8k_161	FLASH-IQ1S-effmed	null	null	gsm8k	712	false	316.98
434	true	false	null	null	gsm8k_163	FLASH-IQ1S-effmed	null	null	gsm8k	223	false	135.46
240	true	false	null	null	gsm8k_178	FLASH-IQ1S-effmed	null	null	gsm8k	71	false	92.3
405	true	false	null	null	gsm8k_189	FLASH-IQ1S-effmed	null	null	gsm8k	124	false	127.6
224	true	false	null	null	gsm8k_191	FLASH-IQ1S-effmed	null	null	gsm8k	136	false	76.78
390	true	false	null	null	gsm8k_198	FLASH-IQ1S-effmed	null	null	gsm8k	183	false	145.2
303	true	false	null	null	gsm8k_206	FLASH-IQ1S-effmed	null	null	gsm8k	124	false	107.89
370	true	false	null	null	gsm8k_209	FLASH-IQ1S-effmed	null	null	gsm8k	149	false	121.33
252	true	false	null	null	gsm8k_228	FLASH-IQ1S-effmed	null	null	gsm8k	115	false	89.31
348	true	false	null	null	gsm8k_255	FLASH-IQ1S-effmed	null	null	gsm8k	184	false	113.36
146	true	false	null	null	gsm8k_285	FLASH-IQ1S-effmed	null	null	gsm8k	61	false	59.93
264	true	false	null	null	gsm8k_292	FLASH-IQ1S-effmed	null	null	gsm8k	106	false	89.42
260	true	false	null	null	gsm8k_318	FLASH-IQ1S-effmed	null	null	gsm8k	94	false	102.46
707	true	false	null	null	gsm8k_326	FLASH-IQ1S-effmed	null	null	gsm8k	500	false	239.15
364	true	false	null	null	gsm8k_333	FLASH-IQ1S-effmed	null	null	gsm8k	100	false	130.1
332	true	false	null	null	gsm8k_334	FLASH-IQ1S-effmed	null	null	gsm8k	185	false	117.75
337	true	false	null	null	gsm8k_350	FLASH-IQ1S-effmed	null	null	gsm8k	146	false	128.21
322	true	false	null	null	gsm8k_393	FLASH-IQ1S-effmed	null	null	gsm8k	120	false	124.49
604	true	false	null	null	gsm8k_407	FLASH-IQ1S-effmed	null	null	gsm8k	258	false	207.69
198	true	false	null	null	gsm8k_429	FLASH-IQ1S-effmed	null	null	gsm8k	89	false	78.2
277	true	false	null	null	gsm8k_432	FLASH-IQ1S-effmed	null	null	gsm8k	101	false	120.13
167	true	false	null	null	gsm8k_435	FLASH-IQ1S-effmed	null	null	gsm8k	55	false	70.64
313	true	false	null	null	gsm8k_440	FLASH-IQ1S-effmed	null	null	gsm8k	137	false	118.28
207	true	false	null	null	gsm8k_447	FLASH-IQ1S-effmed	null	null	gsm8k	81	false	89.11
238	true	false	null	null	gsm8k_449	FLASH-IQ1S-effmed	null	null	gsm8k	107	false	96.9
345	true	false	null	null	gsm8k_451	FLASH-IQ1S-effmed	null	null	gsm8k	136	false	133.97
297	true	false	null	null	gsm8k_457	FLASH-IQ1S-effmed	null	null	gsm8k	122	false	117.27
320	true	false	null	null	gsm8k_466	FLASH-IQ1S-effmed	null	null	gsm8k	93	false	117.22
183	true	false	null	null	gsm8k_469	FLASH-IQ1S-effmed	null	null	gsm8k	80	false	76.7
159	true	false	null	null	gsm8k_476	FLASH-IQ1S-effmed	null	null	gsm8k	76	false	67.88
293	true	false	null	null	gsm8k_501	FLASH-IQ1S-effmed	null	null	gsm8k	133	false	115.1
1,146	false	false	null	null	gsm8k_505	FLASH-IQ1S-effmed	null	null	gsm8k	743	false	385.7
309	true	false	null	null	gsm8k_538	FLASH-IQ1S-effmed	null	null	gsm8k	106	false	124.26
484	true	false	null	null	gsm8k_541	FLASH-IQ1S-effmed	null	null	gsm8k	187	false	167.4
256	true	false	null	null	gsm8k_542	FLASH-IQ1S-effmed	null	null	gsm8k	66	false	100.51
432	true	false	null	null	gsm8k_546	FLASH-IQ1S-effmed	null	null	gsm8k	204	false	163.88
221	true	false	null	null	gsm8k_548	FLASH-IQ1S-effmed	null	null	gsm8k	79	false	87.15
491	true	false	null	null	gsm8k_552	FLASH-IQ1S-effmed	null	null	gsm8k	180	false	169.05
227	true	false	null	null	gsm8k_563	FLASH-IQ1S-effmed	null	null	gsm8k	96	false	86.45
573	true	false	null	null	HumanEval_6	FLASH-IQ1S-effmed	null	null	humaneval_plus	274	false	142.85
326	true	false	null	null	HumanEval_7	FLASH-IQ1S-effmed	null	null	humaneval_plus	178	false	97.22
505	true	false	null	null	HumanEval_8	FLASH-IQ1S-effmed	null	null	humaneval_plus	250	false	137.3
566	false	false	null	null	HumanEval_22	FLASH-IQ1S-effmed	null	null	humaneval_plus	369	false	136.42
129	true	false	null	null	HumanEval_23	FLASH-IQ1S-effmed	null	null	humaneval_plus	78	false	48
591	true	false	null	null	HumanEval_26	FLASH-IQ1S-effmed	null	null	humaneval_plus	392	false	150.32
136	true	false	null	null	HumanEval_28	FLASH-IQ1S-effmed	null	null	humaneval_plus	68	false	50.19
287	true	false	null	null	HumanEval_35	FLASH-IQ1S-effmed	null	null	humaneval_plus	132	false	87.07
297	true	false	null	null	HumanEval_55	FLASH-IQ1S-effmed	null	null	humaneval_plus	76	false	82.72
551	true	false	null	null	HumanEval_57	FLASH-IQ1S-effmed	null	null	humaneval_plus	256	false	146.22
778	true	false	null	null	HumanEval_59	FLASH-IQ1S-effmed	null	null	humaneval_plus	591	false	168.89
509	true	false	null	null	HumanEval_62	FLASH-IQ1S-effmed	null	null	humaneval_plus	232	false	135.17
762	true	false	null	null	HumanEval_70	FLASH-IQ1S-effmed	null	null	humaneval_plus	356	false	183.59
1,059	true	false	null	null	HumanEval_108	FLASH-IQ1S-effmed	null	null	humaneval_plus	566	false	257.12
1,724	true	false	null	null	HumanEval_129	FLASH-IQ1S-effmed	null	null	humaneval_plus	1,197	false	349.58
479	true	false	null	null	HumanEval_139	FLASH-IQ1S-effmed	null	null	humaneval_plus	226	false	133.02
745	true	false	null	null	HumanEval_143	FLASH-IQ1S-effmed	null	null	humaneval_plus	409	false	191.14
1,679	true	false	\pi	\pi	math_precalculus_1199	GPTOSS-MXFP4-effmed	\boxed{\pi}	\pi	math25	1,228	false	195.55
974	true	false	5	5	math_algebra_1837	GPTOSS-MXFP4-effmed	5	5	math25	512	false	126.47
794	false	false	x=5	x=5	math_algebra_2193	GPTOSS-MXFP4-effmed	5	5	math25	250	false	97.38
116	false	false	11\sqrt2	11\sqrt2	math_prealgebra_1302	GPTOSS-MXFP4-effmed	$11\sqrt{2}$	$11\sqrt{2}$	math25	33	false	14.75
269	true	false	1.25	1.25	math_prealgebra_1558	GPTOSS-MXFP4-effmed	1.25	1.25	math25	71	false	55.98
659	true	false	2516_8	2516_8	math_number_theory_516	GPTOSS-MXFP4-effmed	2516_8	2516_8	math25	266	false	80.63
1,054	false	false	\frac{3}{2}	3/2	math_algebra_2700	GPTOSS-MXFP4-effmed	$\displaystyle \frac{3}{2}$	$\displaystyle3/2$	math25	564	false	155.55
1,370	true	false	120^\circ	120	math_precalculus_44	GPTOSS-MXFP4-effmed	120.	120	math25	423	false	167.18

End of preview. Expand in Data Studio

llama.cpp Throughput for Qwen3.8-Flash-Next and gpt-oss-120b on a 62 GiB RAM Laptop — result index

Matthew Schwartz — ORCID 0009-0009-4171-7247

This dataset accompanies “llama.cpp Throughput for Qwen3.8-Flash-Next and gpt-oss-120b on a 62 GiB RAM Laptop.” The immutable tagged report and full artifact use the following versioned path:

https://github.com/matthematics1137/research-artifacts/tree/moe-nvme-streaming-v1.0.0/papers/2026-moe-nvme-streaming

The version-1.0.0 artifact is identified by doi:10.5281/zenodo.22314465, with release date 2026-09-28. The Hub copy is a discovery mirror; the versioned Zenodo archive is the artifact of record. A reserved DOI in a prepublication copy does not itself establish public availability; the released Zenodo record and matching GitHub tag do.

Data

Path	Description	Rights/provenance
`data/timing_measurements.jsonl`	generation, placement, and prefill measurements retained for the report	original measurement fields, CC BY 4.0
`data/quality_outcomes.jsonl`	item IDs, correctness, timings, token metadata, and short MATH answer fields	original measurements plus limited MIT-licensed benchmark answer fields
`data/claims.json`	deterministic protocol, tensor-inventory, statistics, and evidence-fingerprint record	original audit record, CC BY 4.0
`environment/`	machine, engine association, and four exact upstream model-file identities	original documentation; weights are not included

Prompts, model response prose, private paths, secrets, model weights, and engine source are excluded from this Hub mirror.

Question and headline result

The narrow question is what decode throughput unmodified llama.cpp's ordinary mmap path delivered for these exact deployments. On one laptop with 62 GiB RAM and an RTX 4080 Laptop GPU, a 67.564 GiB Qwen3.8-Flash-Next GGUF produced 9.4–14.7 decode tokens/s across six observations in two sequential host-state groups. A 59.034 GiB gpt-oss-120b GGUF produced 8.7–9.3 decode tokens/s across three observations. Each principal observation requested 256 output tokens. Decode excludes loading and prompt processing; the recorded prompt rates were slower. These measurements do not establish interactive response latency. This is a bounded, hardware-specific throughput baseline, not a new mmap technique or inference optimization, state-of-the-art claim, cache-hit estimate, decode-bandwidth law, quality-preservation result, or universal model ranking.

Verify

From the full GitHub/Zenodo artifact:

python3 check_claims.py
python3 evaluation/run_eval.py --selftest

The first program verifies the sanitized evidence hashes and recomputes the headline rates and row-backed scores, including the disclosed MATH rescore. Full inference instructions are in FULL_RERUN.md in the tagged artifact.

Upstream artifacts

Flash: unsloth/Qwen3.8-Flash-Next-GGUF, revision d3bc75ee6ccef3efc1e228ec00a6cc2cdb1e2249, Qwen Community License 1.0.
gpt-oss: ggml-org/gpt-oss-120b-GGUF, revision 238abdd290bb874b90a5da1b4549881b7d05c091, Apache-2.0.

Every basename, byte count, and Git LFS SHA-256 is in environment/model_artifacts.json. This dataset claims no ownership of the weights.

Limitations

One machine, one NVMe device, one associated engine revision, two artifacts, and three sequential observations per principal group.
Host state was neither randomized nor held fixed; memory fields differ across groups.
Storage counters span the physical device and whole process invocation.
No resident same-model quality control was run.
Flash HumanEval+ is omitted because three item rows are missing.

Rights and citation

Measurement fields and documentation are CC BY 4.0; authored code is MIT. Benchmark and model terms differ, so the dataset uses Hugging Face's other label; consult the full artifact's LICENSE and THIRD_PARTY_NOTICES.md. Cite the report for scientific claims and the version DOI for frozen evidence.

AI assistance

Claude (Anthropic) assisted with evaluation-harness development and experiment orchestration. Claude and Codex (OpenAI) assisted with evidence auditing, analysis and figure code, citation checking, and manuscript drafting and revision. Matthew Schwartz directed the research and is responsible for its methods, results, interpretation, claims, citations, rights, and released artifacts.

Downloads last month: 115