---
license: gpl-3.0
tags:
- agent
pretty_name: Shodan Queries
size_categories:
- 1M<n<10M
---
# Shodan Query Dataset

## What This Dataset Is

A supervised fine-tuning (SFT) dataset that teaches a language model to **write and explain Shodan search queries**. Every assistant response is either a verbatim Shodan query string composed from documented filter syntax (`filtername:value`), or a short factual explanation of what a given query matches. The dataset is grounded entirely in publicly documented Shodan filter names, banner fields, and example queries — the model learns to _compose_ queries from known primitives, not to hallucinate undocumented filters.

### Sources (provenance-tracked)

| Source | Type | Content |
|---|---|---|
| [Datapedia](https://datapedia.shodan.io/) | Embedded snapshot | Banner properties, example queries, tag values |
| [Shodan Search Filters](https://www.shodan.io/search/filters) | Embedded snapshot | Filter groups and their descriptions |
| [awesome-shodan-queries](https://github.com/jakejarvis/awesome-shodan-queries) | Embedded snapshot + raw fetch | ~75 curated queries organized by category |
| [memcrashed](https://github.com/diogobaeder/memcrashed.git) | Local clone parse | Memcached amplification facts (ports, service names) |
| [help.shodan.io — Search Fundamentals](https://help.shodan.io/the-basics/search-query-fundamentals) | Fetched + cached | Official query syntax documentation |
| [Awesome-Malicious-Infra-Hunting-Queries](https://github.com/frknaykc/Awesome-Malicious-Infra-Hunting-Queries) | Fetched + cached | Threat hunting Shodan queries |
| [Shodan Cheat Sheet](https://github.com/Daemon-AT4/Shodan_Cheet-Sheet) | Fetched + cached | Additional query patterns |

### Row format (Unsloth-compatible)

```json
{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user",   "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "system":   "...",
  "user":     "...",
  "response": "...",
  "metadata": { "row_kind": "...", "source_url": "...", "augmentation_kind": "...", ... }
}
```

### Row kinds

- **`nl_to_query`** — Natural language request → exact Shodan query string.
- **`query_to_explain`** — Verbatim query → short explanation of what it targets.

### Augmentation kinds

- **`none`** — Direct corpus-derived row.
- **`country_slot`** — Base query composed with `country:<ISO-3166-1-alpha-2>` and user prompt includes the country name in natural language.
- **`port_slot`** — Base query composed with `port:<N>` from corpus-observed and supplementary port lists.
- **`paraphrase_oversample`** — Same assistant query, user prompt rewritten with a different NL template.

### Scale

| Mode | Typical Row Count |
|---|---|
| Base (no `--augment`) | ~3,400 |
| `--augment --target-rows 1000000` | 1,000,000 |

---

## What This Dataset Is NOT

- **Not a Shodan API client dataset.** It does not teach API usage (`/api-info`, `/shodan/host`, REST endpoints). It teaches _search query syntax_ only.
- **Not a vulnerability scanner.** Queries identify _what is publicly indexed_; the dataset does not include exploitation steps, payloads, or post-access instructions.
- **Not a live data feed.** It contains query _patterns_, not query _results_. No IP addresses, banners, or host metadata from actual Shodan scans are included.
- **Not a general cybersecurity Q&A set.** It is narrowly scoped to Shodan's filter language and banner fields. It will not teach a model about Nmap, Metasploit, or attack methodologies.
- **Not a replacement for reading Shodan documentation.** Augmented rows are deterministic compositions of known-good filters and approved slot values (countries, ports). They do not cover every possible filter combination or edge case.
- **Not a safety-filtered dataset.** It intentionally does not include refusal patterns, disclaimers, or content warnings. If you need alignment-style guardrails, layer them in post-training or via system prompt at inference time.