Datasets:

ruthvikanne
/

kernel-mcu-c-pool

language:
  - en
pretty_name: Kernel and MCU-OS C pool (code + design docs)
task_categories:
  - text-generation
tags:
  - code
  - c
  - linux-kernel
  - rtos
  - embedded
  - domain-adaptation
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: heldout
        path: data/heldout-*.parquet

Kernel and MCU-OS C pool

Unmixed pool of C/assembly source and design documentation for domain-adaptive training of a kernel-level coding model. Apply the mix at sampling time using the bucket column.

bucket	content	est. tokens (train)
mcu	FreeRTOS, Zephyr, NuttX, RIOT, CMSIS, xv6, Contiki-NG, RT-Thread, U-Boot, OpenSBI, seL4, libopencm3	~116.5M
linux_gpu	drivers/gpu, video, accel	~25.3M
linux_cpu	arch/*, cpufreq, cpuidle, irqchip, clocksource, perf, firmware	~7.7M
linux_core	kernel, mm, fs, net, block, pci, usb, dma, include/linux ...	~45.2M
docs	Linux Documentation/, Zephyr doc/, NuttX Documentation/, U-Boot doc/, RISC-V ISA manual, seL4 docs	~14.4M

Token counts are byte-based estimates (3.6 bytes/token), not tokenizer counts.

Columns

text, path, bucket, source, spdx (per-file SPDX tag if present, else the repository licence), repo_license, commit (upstream commit the file came from), def_heavy (file is >60% #define lines, typically register headers; consider down-sampling), n_bytes.

Splits

heldout holds about 2% of directories (hash of the first 5 path components), so no directory appears in both splits. Caution: heldout is skewed toward mcu (about 19 percent of files in the pool but most of heldout tokens); evaluate per bucket.

Filters

Exact dedup (whitespace-normalised SHA-1 only; no near-duplicate removal), 300 B to 120 KB per code file (500 B to 150 KB for docs), generated-file and hex-table filters, test/sample/scripts/dts/board paths skipped for code, non-ASCII ratio limits.

Licences (read before use)

Content keeps its upstream licences, mostly GPL-2.0 variants (Linux, U-Boot, seL4), Apache-2.0 (Zephyr, NuttX, RT-Thread, CMSIS), MIT, BSD, LGPL, and CC-BY-4.0 (RISC-V ISA manual). Per-file SPDX tags are preserved in spdx; see LICENSES.json for the upstream commit and licence files per documentation source. Whether training on GPL code makes a model a derivative work is legally unsettled. This dataset does not resolve that; you are responsible for your use and for the licence terms when redistributing the data or derived weights.

Not included

Commercial textbooks. Open-licensed books (e.g. LDD3, Mastering the FreeRTOS Real Time Kernel) are not yet added; their licences must be verified at source first.