Datasets:

ruthvikanne
/

kernel-pool-knowledge

license: other
task_categories:
  - text-generation
language:
  - en
  - code
tags:
  - kernel
  - systems-programming
  - computer-architecture
  - algorithms
pretty_name: kernel-llm pool (part 1, knowledge slice)

kernel-llm pool — part 1: books, algorithms, mathematics, architecture

Training text for a kernel/machine-level C specialist. This is part 1 of a multi-part pool, not the full 1T-token dataset: it is the knowledge slice (design, architecture, algorithms, maths), built from repositories whose licences were read rather than assumed.

Volume is stated as measured: 15.6M tokens across 6735 rows, where a token is estimated as bytes/3.6 because no tokenizer was run over this part. Byte counts are exact; the token figure is an estimate and is labelled as one in stats.json.

source	bucket	licence	rows	est. tokens	flag
`opentitan`	arch_code	Apache-2.0	4065	8034k	clean
`vulkan-spec`	books	CC-BY-4.0	726	3118k	clean
`aosa`	books	CC-BY-3.0	76	797k	clean
`rocket-chip`	arch_code	BSD-3-Clause OR Apache-2.0	308	592k	clean
`discrete-book`	math	CC-BY-NC-SA-4.0	132	577k	restricted
`cp-algorithms`	algorithms	CC-BY-SA-4.0	165	457k	clean
`miaow`	arch_code	BSD-3-Clause	204	447k	clean
`opencl-spec`	books	CC-BY-4.0	106	422k	clean
`thealgorithms-c`	algorithms	GPL-3.0	398	275k	clean
`sdxpy`	books	CC-BY-NC-4.0	115	255k	restricted
`spike`	arch_code	BSD-3-Clause	259	249k	clean
`vexriscv`	arch_code	MIT	97	228k	clean
`ods`	books	CC-BY-2.5-CA	19	135k	clean
`ods-code`	algorithms	CC-BY-2.5-CA	65	61k	clean
total			6735	15.6M

Every row carries source, repo_license, spdx, commit, n_bytes, n_tokens (-1 when not measured), restricted and path, so a subset can be rebuilt exactly and every row can be traced to the commit it came from.

Licence flags — read before training

This pool contains NonCommercial material. It is included and marked, but it is not free for commercial use:

discrete-book — CC-BY-NC-SA-4.0
sdxpy — CC-BY-NC-4.0

Filter those rows out of any commercial training mix:

ds = ds.filter(lambda r: not r["restricted"])

Publishing this data, or keeping it private, does not change the position: an NC licence restricts the use, not only the redistribution, so training a model you sell on NC text is a commercial use even if the text never leaves your disk.

Publisher-copyrighted books (Operating Systems: Three Easy Pieces, Computer Architecture: A Quantitative Approach, Designing Data-Intensive Applications, Systems Performance, Linux Kernel Development, Refactoring, Design Patterns) are not present at all — no rows, no URLs. A book you can read online is not a book you may train on.

What is not here

kernel and driver trees (Linux, RTOS, GPU drivers) — part 2
other kernels and compute kernels — part 3
Stack v2 subsets — part 4 (gated)
traces, commits, synthetic and replay data — part 5

Dedup and splits

6735 distinct near-duplicate signatures; exact duplicates and MinHash near-duplicates were dropped, with sources deduplicated in priority order so a trusted copy wins over a mirror. heldout is 2% of groups, capped so no bucket is gutted.

The held-out split is public. It is a development split for measuring contamination, not a secret holdout; encrypting public data does not create one.

How to load

from datasets import load_dataset
ds = load_dataset("ruthvikanne/kernel-pool-knowledge", split="train")

mix.json and mix-manifest.parquet hold a plan (tier weights, repeat counts) that joins on path; the text stays in the shards. LICENSES.json records, per source, the declared licence, the evidence it was read from, and the exact commit.

kernel-llm pool — part 1: books, algorithms, mathematics, architecture

Contents

Licence flags — read before training

What is not here

Dedup and splits

How to load