Datasetpublic

dac-research/longbench_synthetic_v3

LongBench Synthetic V3 Dataset statistics Per-subset stats over uploaded samples. Unique ctx counts distinct context strings, while the context-length buckets count samples/rows. Token counts use Qwen/Qwen3-14B. Pool Subset Unique ctx Sample rows <8K 8-16K 16-32K >32K Median tok p90 tok Max tok eval 2wikimqa 197 197 139 54 4 0 6,627 13,271 16,982 eval gov_report 200 200 89 84 24 3 8,902 18,212 52,521 eval musique 200 200 3 46 151 0 16,733 17,231 17… See the full description on the dataset page: https://huggingface.co/datasets/dac-research/longbench_synthetic_v3.

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes21downloads

Dataset Card

LongBench Synthetic V3

Dataset statistics

Per-subset stats over uploaded samples. Unique ctx counts distinct context strings, while the context-length buckets count samples/rows. Token counts use Qwen/Qwen3-14B.

Pool	Subset	Unique ctx	Sample rows	<8K	8-16K	16-32K	>32K	Median tok	p90 tok	Max tok
eval	2wikimqa	197	197	139	54	4	0	6,627	13,271	16,982
eval	gov_report	200	200	89	84	24	3	8,902	18,212	52,521
eval	musique	200	200	3	46	151	0	16,733	17,231	17,824
eval	qasper	148	200	187	10	3	0	4,674	7,469	21,879
eval	qmsum	35	200	40	108	52	0	12,970	25,539	30,377
eval	samsum	200	200	88	104	8	0	9,215	14,930	17,987
train	hotpotqa	200	1,600	216	880	504	0	14,982	16,947	17,578
train	hotpotqa_e	288	2,304	928	1,120	256	0	9,693	16,530	17,322
train	multi_news	199	1,592	1,552	40	0	0	2,048	4,834	14,278
train	multinewse	282	2,256	1,336	736	152	32	6,768	15,226	41,048
train	multifieldqa_en	112	896	504	384	8	0	7,712	12,004	16,446
train	narrativeqa	20	160	0	48	48	64	31,296	60,327	65,301
train	triviaqa	200	1,600	536	680	384	0	11,530	19,600	23,284
train	triviaqa_e	299	2,392	1,184	856	344	8	8,289	17,669	36,186

Train pool: 1,600 unique contexts → 12,800 question rows (split into HF train + validation). Eval pool: 980 unique contexts → 1,197 question rows (uploaded as HF test).