README.md

---
language:
- en
license: other
license_name: mpl-2.0-commons-clause
tags:
- security
- red-team
- purple-team
- cybersecurity
- synthetic
- sft
- mitre-attack
- agentic
- tool-use
size_categories:
- n<1K
task_categories:
- text-generation
---

# Meridian-SFT-500

500 synthetic purple team SFT examples generated from the full MITRE ATT&CK
Enterprise technique corpus across nine generation angles. Intended for
supervised fine-tuning of red team and security-focused LLMs.

## What's in it

Each entry pairs an ATT&CK-contextualized prompt with a structured security
analysis demonstrating simultaneous offensive and defensive knowledge. The
dataset spans all 11 Enterprise tactics and covers both Windows and Linux
platforms.

Generated by purple-team-gen
using a locally-hosted Qwen3.6-27B-uncensored-heretic-v2 inference server.
Thinking mode was disabled during generation. All entries passed a quality
gate: minimum 900 character response, refusal pattern detection, SIGMA YAML
syntax validation where applicable, and mandatory code block presence.

## Generation angles

Each technique is approached from one of nine angles, distributed across the
500 entries:

| Angle | What it produces |
|---|---|
| `simulate_and_detect` | Attack implementation + SIGMA rule + remediation |
| `detection_engineering` | Behavioral indicators + SIGMA + Splunk hunt query |
| `redteam_operator` | Evasion-aware code + artifact footprint + cleanup |
| `security_tooling` | Offensive tool + companion detection script |
| `sample_analysis` | Malicious code sample + technical breakdown + YARA rule |
| `attack_chain` | Multi-stage kill chain with this technique as a step |
| `incident_triage` | Alert analysis + triage script + containment procedure |
| `hardening_guide` | Configuration hardening + monitoring setup + verification |
| `threat_modeling` | Attack paths + risk assessment + control mapping |

## Format

ChatML JSONL. One record per line:

```json
{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user",   "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "metadata": {
    "technique_id":   "T1055.001",
    "technique_name": "Process Injection: Dynamic-link Library Injection",
    "tactics":        ["defense-evasion", "privilege-escalation"],
    "platforms":      ["Windows"],
    "angle":          "simulate_and_detect",
    "generated_at":   "2026-07-24T...",
    "source":         "purple-team-gen"
  }
}
```

## Loading

```python
import json
from datasets import load_dataset

# From HuggingFace
ds = load_dataset("TitleOS/Meridian-SFT-500", split="train")

# Or directly from JSONL
with open("meridian-sft-500.jsonl") as f:
    records = [json.loads(line) for line in f]

messages  = records[0]["messages"]   # ChatML turns, ready for tokenization
technique = records[0]["metadata"]["technique_id"]
angle     = records[0]["metadata"]["angle"]
```

## Statistics

| | |
|---|---|
| Total entries | 593 |
| Unique technique IDs | ~280 |
| Enterprise tactics covered | All 11 |
| Angles with SIGMA rules | `simulate_and_detect`, `detection_engineering`, `attack_chain`, `threat_modeling` |
| Entries with fenced code blocks | 500 (100%) |
| Source STIX bundle | MITRE ATT&CK Enterprise (latest) |
| Generator model | Qwen3.6-27B-uncensored-heretic-v2 |

## Intended use

SFT data for LLMs targeting:

- Red team and purple team agentic workflows
- Codebase vulnerability auditing with remediation
- Detection rule generation (SIGMA, YARA, SPL)
- Security tool development
- Threat modeling and incident response

**Not intended for** production deployment without human review, or any use
outside authorized, scoped security testing engagements.

## What it isn't

This is a small sample dataset — 593 entries across ~697 Enterprise
techniques means roughly one angle per technique on average. It is suitable
as a representative slice for dataset mixing or evaluation, not as a
standalone fine-tuning corpus. For a full-coverage run (all techniques × all
9 angles ≈ 5,000+ entries), see the purple-team-gen pipeline.

The synthetic nature of the content means some technical specifics (API
signatures, event IDs, file paths) may not be accurate to every real-world
environment. Human review before operational use is required.

## License

MPL-2.0 with Commons Clause. Licensor: TitleOS.

The Commons Clause restricts selling this dataset or services whose value
derives substantially from it. See [LICENSE](LICENSE) for full terms.