Pegainfer/kern-prompts

kern prompts

Chat prompts for checking the numerics of an LLM inference engine against a reference implementation of the same model, as done by kern kern test.

The reference samples a continuation for each prompt and records its next-token distribution at every step. A candidate (a new kernel, a lower precision) is then fed the same tokens and scored position by position: KL divergence, top-1 flips, and the top-1/top-2 margin at each flip. The prompt fills the KV cache and recurrent state through prefill; the continuation runs token by token through decode. The prompts are chosen so that both are deep: long contexts, long agent trajectories, and tasks that make the model write for a long time.

Splits

split	rows	prompt tokens	suggested decode cap	use
smoke	17	11 – 15k	64	after every change, about a minute
standard	96	11 – 180k	512	the gate for a kernel swap
full	40	17 – 965k	16384	precision or state-storage changes, releases

smoke is the shortest standard row of every source (one per language where a source has two), so it repeats rows of standard. Token counts use the Qwen3.8 tokenizer over message contents, without a chat template. A run skips rows longer than the target model's context.

Errors that build up step by step (a recurrent state stored in bf16, for example) stay close to normal for the first few hundred decode steps and grow over thousands, so only full shows them clearly.

Schema

column	type	meaning
messages	string	OpenAI chat messages as JSON. The last turn is user or tool; the model writes the next assistant turn. Tool calls keep their OpenAI fields.
tools	string or null	OpenAI tool schemas as JSON, for rows whose conversation calls tools
source	string	<repo>@<revision>#<row>. A suffix :<k> means the conversation was cut before message k. Gutenberg rows give the character range of the book.

Composition

source	smoke	standard	full	prompt tokens	what it covers
caskcsg/LongBench-Pro	2	25	12	8k – 254k	natural long documents, English and Chinese; question_thinking, so the model reasons over the deep context
openai/mrcr (2needle)	1	5	6	10k – 965k	multi-turn conversations up to 1M tokens
SWE-bench/SWE-smith-trajectories (xml)	1	12	4	15k – 163k	coding-agent trajectories cut before an assistant turn at 16k / 32k / 64k / 128k; the four longest whole in full
nvidia/Nemotron-Agentic-v1 (interactive_agent)	1	6	–	3k – 11k	tool calling with tool schemas
allenai/WildChat-4.8M	2	17	–	95 – 4k	real multi-turn chats: 6 English, 9 Chinese, 1 Russian, 1 Japanese; not toxic, not redacted
openai/gsm8k	1	4	–	< 100	grade-school math
AI-MO/NuminaMath-1.5 (olympiads)	1	4	4	< 200	competition math, long reasoning
openai/openai_humaneval	1	3	–	< 200	function completion
google-research-datasets/mbpp (sanitized)	1	3	–	< 200	short programs with tests
XiaomiMiMo/MiMo-V2.6-RL-oss (webdev, music)	2	7	6	< 700	whole static websites (long code output); ABC music notation from Chinese and English briefs
zai-org/LongWriter-6k	2	4	6	< 750	requests for very long writing, English and Chinese (prompts only)
Project Gutenberg	2	6	2	2k – 393k	continue a book: English (Eliot, Darwin, Dickens, Smith) and Chinese (水滸傳, 三國志演義, 西遊記, 紅樓夢)

Reproducing

build.py builds the three files from pinned source revisions (listed in REV and BOOKS). There is no randomness: candidates are ordered by the sha256 of their id, and the first that fit a slot are taken. Two runs give identical bytes.

mkdir src && (cd src && bash ../fetch.sh)   # about 2.5 GB
python build.py src <Qwen3.8 tokenizer.json> <out dir>

Licenses

Each row keeps the license of its source; source names it. Using a row means following that license, including attribution.

source	license
LongBench-Pro, MiMo-V2.6-RL-oss, NuminaMath-1.5, LongWriter-6k	Apache-2.0
MRCR, SWE-smith-trajectories, GSM8K, HumanEval	MIT
Nemotron-Agentic-v1, MBPP	CC-BY-4.0
WildChat-4.8M	ODC-BY
Project Gutenberg texts	public domain in the USA; the Project Gutenberg header and footer were removed, so no Project Gutenberg trademark is used

The selection, the cuts and build.py are released under Apache-2.0.

WildChat rows are real user conversations, filtered by WildChat's own toxic and redacted flags. SWE-smith trajectories were generated by Claude 3.7 Sonnet through SWE-agent. Some sources are public benchmarks; do not train on this set if you evaluate on them.