longctx30

Thirty long-context prompts for benchmarking LLM inference throughput. Each prompt is about 10,000 input tokens and asks for about 1,500 output tokens, which is long enough that decode time dominates and tokens per second is a meaningful number.

Built for the article Learning inference: How to host and improve the token speed of an LLM, where it is the benchmark set for Gemma 4 31B on a single B300. The data and code the article uses are in this repository:

file	what it is
longctx30.jsonl	the 30 prompts
make_longctx_dataset.py	rebuilds the prompts from the Gutenberg texts
bench_client.py	the benchmark client: streams completions, reports p50 output tok/s and TTFT
fpa4fix/	the seven vLLM patches, the self-checking Dockerfile, and the correctness script

Shape

One JSON object per line in longctx30.jsonl:

field	type	value
id	string	10k-pd-000 … 10k-pd-029
workload	string	10k
task	string	one of summarize, qa, extract_facts, outline, rewrite, cycling in that order
source	string	gutenberg
article_ids	list of two strings	which book and which passage, e.g. darwin-origin:2
input_tokens	int	10000 (nominal; measured 10,059 to 10,080 on o200k_base including the instruction)
max_out	int	1500
prompt	string	the full prompt

Each prompt is an intro line, then two passages of about 5,000 tokens each from two different books under an --- ARTICLE --- header, then one of five task instructions under --- TASK ---. The tasks all ask for approximately 1,500 tokens of structured output and say "Do not stop early", so the output length is stable across models.

Source text

Twelve public-domain nonfiction works from Project Gutenberg:

Darwin, On the Origin of Species and The Voyage of the Beagle · Thucydides, History of the Peloponnesian War · Herodotus, Histories · Plutarch, Lives · Adam Smith, The Wealth of Nations · Mill, On Liberty · The Federalist Papers · Douglass, Narrative of the Life · Equiano, The Interesting Narrative · Thoreau, Walden · Grant, Personal Memoirs.

Five passages are taken from each book, evenly spaced through it, skipping the first 6% and last 8% so nothing lands in a table of contents, preface, index or appendix. Passage k is paired with passage k+30, which always lands on a different book.

Reproducing it

make_longctx_dataset.py is included. It needs the twelve plain-text Gutenberg files and tiktoken:

pip install tiktoken
python make_longctx_dataset.py --src ./gutenberg --out longctx30.jsonl

Token counts use o200k_base, the same tokenizer the benchmark client uses.

Using it

from datasets import load_dataset
ds = load_dataset("abhijithneilabraham/longctx30", split="train")
print(ds[0]["task"], len(ds[0]["prompt"]))

Or download the file and point bench_client.py at it:

pip install aiohttp tiktoken
python bench_client.py --url http://localhost:8000/v1/chat/completions \
    --model nvidia/Gemma-4-31B-IT-NVFP4 --dataset longctx30.jsonl \
    --concurrency 1 --warmup 2 --json-out result.json

It reads prompt and max_out from each row and sends them as a streaming chat completion.

Licence

The source texts are public domain. The prompts, the task instructions and the generator script are released under CC0.