Datasets:
user stringlengths 11.7k 534k | assistant stringclasses 1
value | ref_prompt_tokens int64 5.33k 131k | output_len int64 256 256 | bucket int64 8.19k 131k | source stringclasses 1
value |
|---|---|---|---|---|---|
You are given a long document followed by a multiple-choice question. Read the document carefully and answer.
{"zhuang_word": "a", "zh_meanings": ["乌鸦", "呀", "呢"], "source": "https://zha_zho.en-academic.com/001", "zh_meanings_full": ["乌鸦 [与roegga同]", "呀 Caezgya vaiq daeuj ～！大家快来呀!", "(【见】 le) 呢 [语气词, 表示疑问]"]}
{"zhuang... | 16,409 | 256 | 16,384 | single | |
You are given a long document followed by a multiple-choice question. Read the document carefully and answer.
{"zhuang_word": "a", "zh_meanings": ["乌鸦", "呀", "呢"], "source": "https://zha_zho.en-academic.com/001", "zh_meanings_full": ["乌鸦 [与roegga同]", "呀 Caezgya vaiq daeuj ～！大家快来呀!", "(【见】 le) 呢 [语气词, 表示疑问]"]}
{"zhuang... | 16,410 | 256 | 16,384 | single | |
"You are given a long document followed by a multiple-choice question. Read the document carefully a(...TRUNCATED) | 65,562 | 256 | 65,536 | single | |
"You are given a long document followed by a multiple-choice question. Read the document carefully a(...TRUNCATED) | 16,409 | 256 | 16,384 | single | |
"You are given a long document followed by a multiple-choice question. Read the document carefully a(...TRUNCATED) | 8,219 | 256 | 8,192 | single | |
"You are given a long document followed by a multiple-choice question. Read the document carefully a(...TRUNCATED) | 8,219 | 256 | 8,192 | single | |
"You are given a long document followed by a multiple-choice question. Read the document carefully a(...TRUNCATED) | 16,408 | 256 | 16,384 | single | |
"You are given a long document followed by a multiple-choice question. Read the document carefully a(...TRUNCATED) | 8,214 | 256 | 8,192 | single | |
"You are given a long document followed by a multiple-choice question. Read the document carefully a(...TRUNCATED) | 8,207 | 256 | 8,192 | single | |
"You are given a long document followed by a multiple-choice question. Read the document carefully a(...TRUNCATED) | 32,793 | 256 | 32,768 | single |
longbench-longctx
Long-context requests for end-to-end LLM inference benchmarking in fastkernels — Scenario B. Exercises the regimes the bulk set can't reach: long-sequence attention (incl. sparse / sliding-window / DSA), RoPE/YaRN scaling, and large-KV decode.
What it's for
64 real long documents truncated into clean prefill-length buckets from 8K to 128K, each paired with its real multiple-choice question. Prefill-dominated: it measures how kernels scale with context length rather than concurrency. Prompts are stored as raw text and tokenized per target model at load.
How it's made
| step | value |
|---|---|
| source | THUDM/LongBench-v2 @ 2b48e494 |
| construction | a real document prefix truncated to each bucket length, with the document's real MC question appended |
| fallback | multi-document concat only if a bucket lacks long-enough docs (not needed here: 64/64 are single-doc) |
| buckets | 8K×32, 16K×16, 32K×8, 64K×4, 128K×4 |
| decode | 256 tokens (fixed, per-row output_len) |
| reference tokenizer | meta-llama/Llama-3.1-8B-Instruct |
Format
train split, 64 rows:
| field | type | description |
|---|---|---|
user |
string | instruction + document + MC question (raw text) |
assistant |
string | empty (generated at eval time) |
ref_prompt_tokens |
int | prompt length under the reference tokenizer |
output_len |
int | fixed decode budget (256) |
bucket |
int | nominal prefill length (8192 … 131072) |
source |
string | single (one doc) or concat (multi-doc fallback) |
At load: apply the target model's chat template to user and truncate to its max_model_len.
Composition (reference tokenizer)
| bucket | reqs | actual prefill mean |
|---|---|---|
| ~128K | 4 | 130,942 |
| ~64K | 4 | 65,531 |
| ~32K | 8 | 32,788 |
| ~16K | 16 | 16,406 |
| ~8K | 32 | 8,122 |
| total | 64 | 1,570,639 prefill + 16,384 decode |
Load
from datasets import load_dataset
ds = load_dataset("sfc-gh-goliaro/longbench-longctx", split="train")
In fastkernels: load_real_prompt_workload("long-context", tokenizer).
Reproduce
python -m fastkernels.build_datasets --which longctx # rebuild locally
python -m fastkernels.build_datasets --which longctx --push # rebuild + re-push
The pinned source revision + seed=42 make this byte-identical on every run.
Attribution
Derived from LongBench-v2; please follow the source dataset's license and terms of use.
- Downloads last month
- 378