README.md

---
language:
- en
license: apache-2.0
tags:
- named-entity-recognition
- cybersecurity
- token-classification
- threat-intelligence
- ioc-extraction
- ner
- span-detection
task_categories:
- token-classification
task_ids:
- named-entity-recognition
pretty_name: Arcspan Cybersecurity NER Dataset
size_categories:
- 10K<n<100K
---

# Arcspan Cybersecurity NER Dataset

A multi-source cybersecurity named entity recognition dataset in OPF (OpenAI Privacy Filter) JSONL format, covering 5 entity classes across threat intelligence reports, CVE descriptions, MITRE ATT&CK entries, APT reports, and more.

Built as the training and evaluation corpus for the [Arcspan](https://huggingface.co/chairulridjal/arcspan) project — fine-tuning OpenAI's sparse MoE Privacy Filter for cybersecurity IOC extraction.

---

## Dataset Summary

| Split | File | Records | Spans | Purpose |
|---|---|---|---|---|
| **R9 Train** | `r9_5class_train.jsonl` | 24,518 | 63,457 | Main training set (latest, leakage-clean) |
| **R9 Valid** | `r9_5class_valid.jsonl` | 2,821 | 5,681 | Validation set |
| **R8 Train** | `r8_5class_train.jsonl` | 26,079 | 76,824 | Previous training set |
| **R8 Valid** | `r8_5class_valid.jsonl` | 2,821 | 5,681 | R8 validation set |
| **APTNER Test** | `aptner_5class_test_clean.jsonl` | 172 | 340 | Independent benchmark (APT reports) |
| **CyNER Test** | `cyner_test.jsonl` | 748 | 892 | CyNER benchmark test set |
| **SecureBERT2 Test** | `securebert2_5class_test.jsonl` | 200 | 283 | SecureBERT2 benchmark test set |
| **Enriched Test** | `enriched_5class_test.jsonl` | 3,853 | 5,512 | Held-out enriched evaluation set |

---

## Label Space

5-class cybersecurity NER schema:

| Label | Description | R9 Train Count |
|---|---|---|
| `Indicator` | IOCs — IPs, domains, URLs, file hashes, file paths, registry keys, email addresses | 16,265 |
| `Malware` | Malware families, ransomware, trojans, backdoors, botnets, campaigns | 15,585 |
| `Organization` | Threat actors, APT groups, vendors, affected organizations | 13,546 |
| `System` | Operating systems, software, platforms, infrastructure components | 11,947 |
| `Vulnerability` | CVEs, exploit names, vulnerability descriptions | 6,114 |

---

## Data Format

All files are JSONL in OPF (OpenAI Privacy Filter) format. Each line is a JSON object:

```json
{
  "text": "APT29 deployed Cobalt Strike via CVE-2021-44228 against Exchange servers.",
  "spans": {
    "Organization: APT29": [[0, 5]],
    "Malware: Cobalt Strike": [[16, 28]],
    "Vulnerability: CVE-2021-44228": [[33, 47]],
    "System: Exchange": [[56, 64]]
  },
  "info": {
    "id": "apt_reports_00042",
    "source": "apt_reports"
  }
}
```

**Span key format:** `"Label: surface_text"` → `[[start_char, end_char], ...]`

Offsets are character-level, zero-indexed, half-open `[start, end)`.

---

## Training Data Sources (R9)

The R9 training set aggregates 22 sources, deduplicated and leakage-cleaned:

| Source | Records | Description |
|---|---|---|
| `cyner2_train` | 4,563 | CyNER v2 training split |
| `cyberner_stix_train` | 3,723 | CyberNER harmonized (STIX-mapped) |
| `dnrti_train` | 2,834 | DNRTI dataset training split |
| `aptner_train` | 2,584 | APTNER training split |
| `apt_reports` | 2,263 | APT reports (LLM-annotated) |
| `nvd_v2` | 1,995 | NVD CVE descriptions v2 (LLM-annotated) |
| `mitre_attack_v2` | 1,485 | MITRE ATT&CK v2 (LLM-annotated) |
| `synthetic_v2` | 1,292 | Synthetically generated IOC examples v2 |
| `cyberner` | 1,204 | CyberNER base |
| `cyner_train` | 717 | Original CyNER training split |
| `defanged_augment` | 652 | Defanged IOC augmentation (e.g. `192[.]168[.]1[.]1`) |
| `exploitdb` | 500 | ExploitDB entries (LLM-annotated) |
| `nvd_cve` | 338 | NVD CVE descriptions (original) |
| `synthetic_ioc` | 92 | Synthetically generated IOC examples v1 |
| `vendor_blogs` | 61 | Security vendor blog posts (LLM-annotated) |
| `security_news` | 45 | Security news articles (LLM-annotated) |
| `cisa_advisories` | 39 | CISA advisories (LLM-annotated) |
| `mitre_attack` | 39 | MITRE ATT&CK (original) |
| `alienvault_otx` | 37 | AlienVault OTX pulses (LLM-annotated) |
| `securebert2_train` | 22 | SecureBERT2 training split |
| `malware_reports` | 21 | Malware analysis reports (LLM-annotated) |
| `dnrti_valid` | 12 | DNRTI validation (included in train) |

---

## Leakage Audit (R9)

Zero overlap between training data and all held-out evaluation sets:

| Held-out Set | Records | Exact Overlap | Prefix-80 Overlap |
|---|---|---|---|
| R9 Validation | 2,821 | **0** | **0** |
| Enriched Test | 3,853 | **0** | **0** |
| CyNER Test | 748 | **0** | **0** |
| SecureBERT2 Test | 200 | **0** | **0** |
| APTNER Test | 172 | **0** | **0** |

Internal duplicates: **0** exact, **0** prefix-80.

---

## Benchmark Evaluation Results

Evaluated using the [Arcspan R8 checkpoint](https://huggingface.co/chairulridjal/arcspan) with strict exact-match scoring (seqeval-style):

### APTNER (Independent benchmark — APT report style)

| Class | F1 | Precision | Recall | Support |
|---|---|---|---|---|
| Malware | **0.707** | 0.793 | 0.637 | 102 |
| Indicator | **0.667** | 0.661 | 0.673 | 55 |
| Vulnerability | 0.500 | 0.429 | 0.600 | 5 |
| Organization | 0.326 | 0.500 | 0.242 | 91 |
| System | 0.160 | 0.615 | 0.092 | 87 |
| **Micro avg** | **0.498** | **0.668** | **0.397** | 340 |

### CyNER Test

| Class | F1 | Precision | Recall | Support |
|---|---|---|---|---|
| Malware | **0.577** | 0.585 | 0.570 | 242 |
| System | 0.399 | 0.412 | 0.387 | 248 |
| Vulnerability | 0.375 | 0.500 | 0.300 | 10 |
| Organization | 0.316 | 0.288 | 0.351 | 131 |
| Indicator | 0.250 | 0.518 | 0.165 | 261 |
| **Micro avg** | **0.405** | **0.454** | **0.365** | 892 |

---

## Usage

### Loading with Hugging Face `datasets`

```python
from datasets import load_dataset

ds = load_dataset("chairulridjal/arcspan-cyber-ner")
# Splits: r9_train, r9_valid, aptner_test, cyner_test, securebert2_test, enriched_test
```

### Loading manually

```python
import json

with open("r9_5class_train.jsonl") as f:
    examples = [json.loads(line) for line in f]

# Access spans
for ex in examples[:3]:
    print(ex["text"][:80])
    for key, offsets in ex["spans"].items():
        label, surface = key.split(": ", 1)
        for start, end in offsets:
            print(f"  [{label}] {ex['text'][start:end]!r} @ {start}:{end}")
```

### Using with OpenAI Privacy Filter / Arcspan

```bash
# Evaluate directly with opf
opf eval r9_5class_train.jsonl \
  --checkpoint chairulridjal/arcspan \
  --device cpu
```

---

## Related Resources

- **Model:** [chairulridjal/arcspan](https://huggingface.co/chairulridjal/arcspan) — Fine-tuned cybersecurity NER model
- **Base model:** [openai/privacy-filter](https://huggingface.co/openai/privacy-filter) — OpenAI's sparse MoE Privacy Filter
- **Source datasets:** CyNER, APTNER, DNRTI, CyberNER, MITRE ATT&CK, NVD, ExploitDB

---

## License

Apache 2.0. Note that individual source datasets may carry their own licenses — see the original dataset repositories for details. LLM-annotated portions were generated from publicly available text.