---
license: apache-2.0
language:
- en
tags:
- cybersecurity
- mitre-attack
- ioc-extraction
- phishing-detection
- log-analysis
pretty_name: Astraea Unified Threat Dataset v4
size_categories: 100K<n<1M
task_categories:
- text-classification
- token-classification
---

# Astraea Unified Threat Dataset (v4)

Developed by **Auren Research**, **Astraea Unified Threat (v4)** is a production-balanced telemetry dataset for fine-tuning multi-task encoder architectures such as [`answerdotai/ModernBERT`](https://huggingface.co/answerdotai/ModernBERT-base) and [`chandar-lab/NeoBERT`](https://huggingface.co/chandar-lab/NeoBERT).

Each example is a structured security log or email payload with task-specific prefixes (for example, `[TASK=LOG_CLS]`), enabling a single model to perform log triage, IOC extraction, MITRE ATT&CK mapping, attack-chain classification, and phishing detection.

---

## Files

| File | Rows | Size | Purpose |
|---|---:|---:|---|
| `train.parquet` | 220,295 | ~84 MB | Primary training split |
| `val.parquet` | 31,843 | ~14 MB | Validation for log classification and IOC NER |
| `test.parquet` | 20,000 | ~22 MB | Held-out public phishing benchmark |
| `astraea-unified-threat.parquet` | 270,263 | ~127 MB | Legacy unified export (optional) |

**Recommended for training:** use `train.parquet`, `val.parquet`, and `test.parquet`. The unified file does not include all rows from the split release (2,971 additional training examples and a task-aware partition are only in the split files).

---

## Dataset Profile

| Metric | Value | Notes |
|---|---|---|
| **Total rows (splits)** | 272,138 | train + val + test |
| **Class balance (`is_malicious`)** | **44.2% malicious** | Globally balanced across splits |
| **Mean anomaly score** | 0.443 | Continuous target in [0.0, 1.0] |
| **Validated IOC spans** | 1,134,937 | Character offsets verified on IOC rows |
| **MITRE coverage** | 19 techniques, 10 tactics | On `mitre_classification` rows |
| **Log types** | 21 | Windows, Linux, cloud, EDR, network, IAM, K8s, WAF, email, and more |
| **Sources** | 23 | Synthetic telemetry + public phishing |
| **Language** | English | Standard telemetry and email syntax |
| **Format** | Parquet | Columnar, Hugging Face–compatible |

---

## Train / Validation / Test Splits

Splits are **task-aware** and **leakage-safe**: no overlapping `text` values across train, val, or test.

### `train.parquet` (220,295 rows)

| Task prefix | `label_type` | Rows |
|---|---|---:|
| `[TASK=LOG_CLS]` | `log_classification` | 126,805 |
| `[TASK=IOC_NER]` | `ioc_extraction` | 48,231 |
| `[TASK=MITRE]` | `mitre_classification` | 24,650 |
| `[TASK=ATTACK_CHAIN]` | `attack_chain` | 12,000 |
| `[TASK=PHISH]` | `phishing_detection` | 8,609 |

Use for fine-tuning all supported tasks. Includes synthetic phishing (`synthetic_email`) plus adversarial and false-positive hard negatives.

### `val.parquet` (31,843 rows)

| Task prefix | `label_type` | Rows |
|---|---|---:|
| `[TASK=LOG_CLS]` | `log_classification` | 24,405 |
| `[TASK=IOC_NER]` | `ioc_extraction` | 7,438 |

Focused on **network and supply-chain telemetry** (netflow, CI/CD, HTTP, firewall, DNS). Malicious rate for log classification is **~19.1%** (vs. ~4.3% in train), providing a harder validation signal than random sampling.

### `test.parquet` (20,000 rows)

| Task prefix | `label_type` | Rows |
|---|---|---:|
| `[TASK=PHISH]` | `phishing_detection` | 20,000 |

All rows are from **`public_phishing`** (Seven Phishing Email Datasets). Use this split for real-world phishing evaluation only.

> **Note:** MITRE and attack-chain tasks currently have no dedicated test split. Hold out a subset of `train.parquet` or report metrics on validation data for those tasks.

---

## Supported Tasks

| Task | Head type | Label columns | Eval split |
|---|---|---|---|
| Log triage | Sequence classification | `is_malicious` | `val` |
| IOC extraction | Token classification (NER) | `ioc_spans` | `val` |
| MITRE mapping | Multi-label classification | `mitre_tactic`, `mitre_technique` | train holdout |
| Attack chain | Sequence / multi-label classification | `attack_stage` | train holdout |
| Phishing detection | Sequence classification | `is_malicious` | `test` |

---

## Technical Features

- **2026 threat-vector telemetry:** Simulated enterprise attack surfaces including NHI/service-account abuse, CI/CD and OAuth compromise patterns, cloud audit anomalies, EDR alerts, and generative-AI-style spearphishing artifacts.
- **Benign background baseline:** Large volumes of benign Windows, Linux, database, cloud, and EDR telemetry to reduce false-positive saturation in production-like settings.
- **Malicious augmentations:** Additional malicious Windows, Linux, EDR, cloud, and netflow examples (not present in the legacy unified export) so models cannot rely on log-type shortcuts.
- **Token boundary fuzzing:** IPs, domains, hashes, and filenames are randomized across examples to encourage behavioral pattern learning over memorization.
- **Public holdout:** All 20,000 public phishing emails are isolated in `test.parquet` for unbiased evaluation.

---

## Data Schema

Each row contains 12 structured columns:

| Column | Type | Description |
|---|---|---|
| `row_id` | int64 | Row index **within each split file** (not a global ID across files) |
| `text` | string | Log stream or email payload; primary model input |
| `source` | string | Provenance label (e.g. `synthetic_cloud`, `public_phishing`) |
| `label_type` | string | Task router (`log_classification`, `ioc_extraction`, etc.) |
| `mitre_tactic` | string | MITRE ATT&CK tactic or `none` |
| `mitre_technique` | string | MITRE technique ID (e.g. `T1059.001`) or `none` |
| `ioc_spans` | JSON string | Span annotations: `[{"start": int, "end": int, "type": str, "value": str}]` |
| `anomaly_score` | float | Continuous score in [0.0, 1.0] |
| `attack_stage` | string | Kill-chain stage label(s) |
| `is_malicious` | bool | Binary threat label |
| `log_type` | string | Telemetry subtype (e.g. `windows_event`, `kubernetes`, `iam`) |
| `metadata` | JSON string | Ancillary provenance fields |

### IOC span types

`ip`, `domain`, `email`, `filename`, `hash_md5`, `hash_sha256`, `hash_sha1`, `punycode`, `registry_key`

---

## Quick Start

```python
from datasets import load_dataset

# Load from a local directory or Hugging Face Hub
train = load_dataset("parquet", data_files="train.parquet", split="train")
val   = load_dataset("parquet", data_files="val.parquet", split="train")
test  = load_dataset("parquet", data_files="test.parquet", split="train")

# Example: log classification
logs = train.filter(lambda x: x["label_type"] == "log_classification")
print(logs[0]["text"])
print(logs[0]["is_malicious"])
```

For multi-task fine-tuning, pass the full `text` field (including the `[TASK=...]` prefix) to models such as ModernBERT, and use task-specific classification or token heads keyed on `label_type`.

---

## Citation

```bibtex
@misc{astraea-unified-threat-v4,
  title        = {Astraea Unified Threat Dataset v4: Balanced Long-Context Telemetry for Multi-Task SecOps Encoders},
  author       = {Auren Research},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/auren-research/astraea-unified-threat}}
}
```