nimitkalra/LongBench-v1

LongBench v1

Prompts are constructed to exactly match how KVPress (kvpress/pipeline.py) builds them. Couple caveats:

1. Four of the tasks are few-shot demonstration tasks. KVPress only compresses the few-shot demonstrations, but the query input might be large (e.g., an entire news article) and this is uncompressed.

   Task	Uncompressed input (mean)	max
   trec	12	20
   triviaqa	738	2021
   samsum	150	507
   lsht	524	3171

LongBench-e variants are not included. LongBench-e uses the same tasks but with a different scoring protocol: results are bucketed by context length (0–4k, 4–8k, 8k+) and averaged within each bucket, rather than producing a single score per task. The KVPress paper (arXiv:2510.00636) reports results using the standard LongBench scorer, not the -e bucketed variant.

Token statistics (Qwen3-4B-Instruct-2507 tokenizer)

Task	Count	Min	Mean	Max	≤32K %
narrativeqa	200	8,289	29,847	65,318	57.5%
qasper	200	1,917	5,034	21,888	100.0%
multifieldqa_en	150	1,313	7,131	16,461	100.0%
hotpotqa	200	1,844	13,391	17,600	100.0%
2wikimqa	200	964	7,491	16,996	100.0%
musique	200	6,746	16,287	17,843	100.0%
gov_report	200	2,046	10,589	52,521	98.5%
qmsum	200	2,584	13,898	30,389	100.0%
multi_news	200	132	2,653	14,278	100.0%
trec	200	1,876	6,826	11,465	100.0%
triviaqa	200	1,518	12,180	24,306	100.0%
samsum	200	1,393	9,179	18,030	100.0%
passage_count	200	4,965	15,321	29,538	100.0%
passage_retrieval_en	200	10,318	12,849	15,575	100.0%
multifieldqa_zh	200	847	4,459	9,962	100.0%
dureader	200	5,236	10,609	19,174	100.0%
vcsum	200	672	9,187	26,413	100.0%
lsht	200	3,671	14,147	28,139	100.0%
passage_retrieval_zh	200	3,335	4,695	6,726	100.0%
lcc	500	991	3,212	32,621	100.0%
repobench-p	500	2,599	10,918	39,255	98.8%