File size: 5,260 Bytes
de87138
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b6b3d24
de87138
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b6b3d24
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de87138
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
---
license: other
task_categories:
- text-generation
language:
- en
tags:
- code
- kernel
- osdev
- linux
- freebsd
- openbsd
- netbsd
- rust
- systems-programming
pretty_name: KernelLLM-2
size_categories:
- 100K<n<1M
---

# KernelLLM-2 Dataset

A high-quality raw dataset for training and fine-tuning LLMs on Operating System Development, significantly expanded for the second version.

## Overview

This dataset combines raw source code from a variety of mature and hobbyist operating system kernels with thousands of expert-level technical discussions and critiques from the Linux Kernel Mailing List, plus unique Git logic-diff traces.

> [!NOTE]
> The source code included in this dataset represents the latest stable versions of the respective repositories as of **February 27th 2026**.

## Dataset Structure

The dataset is provided in **Parquet** format for optimal compression and integration with the `datasets` library.

### 1. `kernelllm_dataset_v2.parquet` (~1.8 GB)
**Unified Knowledge Base:**
- **Files/Entries:** 367,769
- **Size:** ~1.8 GB (Parquet), ~6.1 GB (Raw JSONL)
- **Sources:** 20+ Kernel and System Projects.
- **Cleaning:**
    - Preserved `SPDX-License-Identifier` and `Copyright` lines while removing generic licensing boilerplate headers to prioritize logic.
    - Exact and "smart" deduplication across repositories.
    - Excluded files < 100 characters to remove boilerplate headers.

#### Repository Distribution:
| Repository | Entry Count |
|------------|------------|
| NetBSD | 83,959 |
| Linux | 69,197 |
| FreeBSD | 50,127 |
| OpenBSD | 27,522 |
| illumos | 24,703 |
| Zephyr  | 19,066 |
| ReactOS | 17,688 |
| Haiku | 15,472 |
| U-Boot | 12,674 |
| Coreboot | 12,082 |
| SerenityOS | 9,159 |
| QEMU | 7,029 |
| EDK2 | 5,869 |
| Xen Hypervisor | 2,311 |
| musl libc | 1,875 |
| GRUB | 1,706 |
| Asterinas | 1,417 |
| seL4 | 641 |
| Limine | 151 |
| Redox OS | 34 |

### 2. Specialized Data Types
- **`git_logdiff`:** (234 entries) Pairs commit messages with filtered logic-diffs to show how code evolves.
- **`lkml_critique`:** (~26,600 pairs) Contentious technical discussions pairing code snippets with maintainer feedback.

### 3. Dataset Schema

The unified dataset contains three primary data types, each with a specific schema:

#### Source Code (`data_type: "source"`)
```json
{
  "data_type": "source",
  "source": "repository_name",
  "filepath": "relative/path/to/file.c",
  "code": "cleaned_source_code_content"
}
```

#### Git Commit Diffs (`data_type: "git_logdiff"`)
```json
{
  "data_type": "git_logdiff",
  "commit": "commit_sha_hash",
  "source": "repository_name",
  "message": "commit_message",
  "code": "git_diff_output"
}
```

#### LKML Critiques (`data_type: "lkml_critique"`)
```json
{
  "data_type": "lkml_critique",
  "source": "mailing_list_name",
  "subject": "[PATCH] subject_line",
  "code": "patch_content",
  "critique": "maintainer_feedback",
  "metadata": {
    "author": "reviewer_email",
    "date": "email_date",
    "thread_id": "thread_identifier",
    "is_openbsd": boolean
  }
}
```

#### Data Processing Features
- **Deduplication:** Content-based hashing with normalized whitespace
- **License Header Cleaning:** Preserves SPDX identifiers and copyright lines while removing generic boilerplate
- **Size Filtering:** Excludes files under 100 characters
- **Git-specific:** Deduplicates by commit SHA when available
- **Critique-specific:** Deduplicates based on code + critique content pairs

## Dataset Release Strategy 

This is the **Raw** version of the dataset. It is intended for:
1.  **Intermediate Training:** Pre-training on high-quality systems code.
2.  **Fine-Tuning:** Instruction-tuning models to act as "Kernel Maintainers".
3.  **Research:** Analyzing developer communication patterns in low-level systems.

## Licensing and Usage

This dataset is a collection of publicly available open-source code and mailing list archives. It contains content governed by multiple licenses. For full details, see the [LICENSE](LICENSE) file.

*   **GNU General Public License (GPL) v2/v3** (Linux, seL4, Xen, QEMU, coreboot, GRUB, U-Boot, ReactOS)
*   **Common Development and Distribution License (CDDL)** (illumos)
*   **BSD 2-Clause / 3-Clause / 4-Clause & ISC** (FreeBSD, NetBSD, OpenBSD, SerenityOS, Limine)
*   **MIT License** (Redox, Haiku, musl)
*   **Apache License 2.0** (Zephyr)
*   **Mozilla Public License (MPL) 2.0** (Asterinas)
*   **BSD-2-Clause-Patent** (EDK2)

### Disclaimer

1.  **Original Licenses Apply:** The individual source code files and mailing list archives contained in this dataset remain governed by their original respective licenses. 
2.  **Attribution:** The `source` and `author` fields (where available) facilitate attribution.
3.  **Research & Ethics:** This dataset is provided primarily for research and educational purposes in the field of LLM training and systems programming analysis.
4.  **License Compliance:** Users of this dataset are responsible for ensuring that their use of the data complies with all applicable licenses.

### Opt-out

If you are a maintainer of one of the included repositories and would like your data removed from future versions of this dataset, please open an issue on the Hugging Face dataset page.