Golden Training Datasets
Golden training datasets with complete legal provenance — every record traced to its source, license, and checksum. Built to train and validate LLM defenses against real-world attack patterns, before production deployment, not after compromise.
Corpus
The core training corpora behind Homeland Defender — an OT/ICS security LLM grounded in authoritative sources: CISA CSAF ICS advisories, MITRE ATT&CK for ICS, NVD, IEC 62443, NIST SP 800-82, and vendor advisories.
SFT Corpus
Golden SFT Corpus
The canonical supervised fine-tuning corpus: every harvested advisory reworked into instruction/response pairs with per-record provenance — source, license, and checksums for each of the 10,163 records. The audited v2 90/10 train/test split (9,146 / 1,017) is the exact input used for training.
DPO
Unified DPO Preference Pairs
Remediated, consolidated preference pairs with clean record IDs — the audit-clean input for preference alignment. Enriched with tool-schema system prompts so the model learns not just what to answer, but how to reason about security findings.
DPO — Synthetic
Synthetic DPO Sets
Two boosted generation runs producing larger, higher-quality preference pairs: a 2,000-pair boost set with richer completions and a 3,500-pair LLM-regenerated set. Used to scale preference coverage beyond harvested material.
DPO — Domain
Protocol DPO (Modbus / DNP3)
Preference pairs for industrial protocol analysis, teaching correct identification and reasoning over Modbus and DNP3 traffic. The dedicated input for the protocol grounding LoRA — where generic LLMs most often fail on ICS material.
DPO — Tooling
Tool-Call DPO
Preference pairs with structured tool-call JSON outputs that teach orchestration — knowing when to invoke a scanner, how to structure the call, and how to surface results. Quality-filtered and enriched variants included.
DPO — Identity
Identity DPO
Identity-override preference pairs ("Who are you?" behavior) that make the model's identity a parameter-level property rather than a system-prompt assertion — closing the prompt-injection role-hijacking vector at the weights.
Grounding & Evaluation
Specialized corpora that anchor the model in protocol reality, attacker tradecraft, and detection logic — plus a fresh holdout set that stays out of training for honest evaluation.
Records extracted from 15 ICSNPP protocol dissector repositories (dnp3, modbus, bacnet, s7comm, enip, and more) — raw protocol behavior as training material, so the model reasons from dissector reality rather than textbook descriptions.
~700 records · 15 repos · 1,017 filesReasoning records derived from the MITRE ATT&CK for ICS framework — mapping attacker techniques and procedures to defensive reasoning, so the model connects observed behavior to the technique it represents.
334 records3,132 Sigma detection rules (YAML) harvested as detection-logic training source — teaching the model what detection rules look like, how they're structured, and how to reason about detection coverage.
3,132 rulesSix labeled Modbus/DNP3 PCAPs with ground truth for protocol-analysis evaluation, plus a fresh holdout set explicitly marked DO_NOT_TRAIN — kept out of every training run so benchmark grading stays honest.
6 PCAPs · ground_truth.jsonl · DO_NOT_TRAINProvenance & Governance
Every golden record carries its source, license, and checksum. No unverifiable web scraping, no unknown origin — organizations can validate the data before they deploy defenses built on it.
The full corpus was remediated in a dedicated audit pass: record IDs reminted, train/test splits rebuilt, and provenance fields backfilled across every corpus. Reproducibility scripts are included with the data.
clean_dpo_pairs, make_train_test_split, and remint_record_ids ship alongside the datasets — anyone can re-derive the audit-clean v2 artifacts from source, proving what the data contains.
All production datasets are scrubbed of PII, proprietary source code, live credentials, and production network architectures. Structural disclosure — no offensive payloads published.
Inventory
The complete catalog — every dataset in the production pipeline, with record counts, sizes, and purpose.
| Dataset | Records | Size | Description |
|---|---|---|---|
| SFT Corpus — Golden | |||
| golden-records-manifest.jsonl | 10,163 | 15.3 MB | Structured vulnerability-analysis records (NVD, CISA alerts, GitHub/npm advisories, MITRE, Sigma, ICS vendors). Each record pairs a CVE/advisory with attack vector, severity label, CWE classification, and blue-team actions — trains the model to turn raw advisories into attributed, actionable security findings. |
| corpus-final-provenanced.jsonl | 11,339 | 58.7 MB | Final SFT corpus with provenance backfilled. Trains advisory comprehension — reading CISA, MITRE, NVD, IEC 62443, and NIST source material and producing structured, grounded security knowledge. |
| phase3-instruction-pairs-final-v2.jsonl | 10,163 | 46.7 MB | Deep scenario-reasoning pairs. Each record embeds a realistic OT incident (e.g., a suppressed pump-shutdown command causing pipe rupture) — trains step-by-step reasoning through adversary technique, network anomalies, device behavior, and detection logic. |
| train-split-v2 / test-split-v2.jsonl | 9,146 / 1,017 | 42.0 / 4.7 MB | Remediated 90/10 split of the v2 pairs. Trains instruction following on the exact train/test inputs used to fit and evaluate the SFT phase. |
| DPO — Preference Pairs | |||
| dpo-all-pairs-v2.jsonl | 4,767 | 108 MB | Preference pairs across the alignment corpus, including indirect-prompt-injection safety cases. Trains the model to refuse injection attempts and stay in role (chosen = refusal with guardrails, rejected = compliance) — the safety backbone of the final model. |
| dpo-synthetic-pairs (boost / LLM) | 2,000 / 3,500 | 20.0 / 36.1 MB | Synthetic preference pairs from boosted and LLM-regenerated runs. Trains deep vendor-specific posture analysis — e.g., Siemens fire-protection systems, Rockwell ControlLogix — with non-exploitative guardrails: passive monitoring, log analysis, and architectural review over active testing. |
| protocol-dpo-pairs.jsonl | 75 | 0.15 MB | Modbus packet-analysis preference pairs. Trains the model to identify protocol anomalies (invalid function codes, malformed frames) and emit correct structured dissection output — the domain where generic LLMs fail most on ICS material. |
| dpo-tool-call-pairs.jsonl | 150 | 0.5 MB | Preference pairs over real advisory material (e.g., CISA CSAF vendor advisories) with structured tool-call JSON. Trains orchestration: when to invoke vuln_intel_cve_lookup or protocol_dissection, how to structure the call, and how to surface results. |
| identity_dpo_pairs.jsonl | 300 | 0.18 MB | Identity-override preference pairs. Trains identity as a weight-level property — the model answers identity questions as Homeland Defender, rejecting base-model identity leakage that system prompts alone can't prevent. |
| Domain Grounding | |||
| icsnpp/ (protocol grounding) | ~700 | 1 MB | Real Zeek logs (conn, modbus_detailed, modbus_mask_write_register) produced by ICSNPP analyzers, paired with defensive-action questions. Trains the model to read live OT monitoring output and judge whether activity is a security concern. |
| attack_ics_framework_reasoning.jsonl | 334 | 0.96 MB | Reasoning records over MITRE ATT&CK for ICS techniques (e.g., T1695.001 Serial COM). Trains the model to identify affected asset classes and explain why each is in scope — technique-level analysis, not just recall. |
| sigma-repo/rules/ | 3,132 | — | Full Sigma rules (title, logsource, detection, tags, level). Trains detection-logic literacy — rule structure, coverage reasoning, and writing or reviewing detection rules. |
| ics-advisory-project.csv | — | 2.9 MB | Raw CISA ICS advisory harvest — the uncurated starting material that grounds the entire pipeline in authoritative advisory data. |
| cot-queries/query_bank.jsonl | — | 48 KB | Seed queries with expected tool routing — e.g., "Are there exploited-in-the-wild CVEs against Phoenix Contact devices?" mapped to vuln_intel_cve_lookup. Drives chain-of-thought training that teaches step-by-step analysis with correct tool selection. |
| Evaluation | |||
| data/eval/pcaps/ | 6 | — | Packet-level labeled Modbus/DNP3 captures (dnp3_read, dnp3_write, dnp3_select_operate, modbus_test) with oracle-parsed ground truth — normal and anomalous traffic annotated per packet. The benchmark for protocol analysis, not the training set. |
| b0-fresh-holdout/ | — | — | Holdout set excluded from training (DO_NOT_TRAIN). Grades benchmark performance honestly — the model has never seen this data, so results reflect real generalization. |
All v2 artifacts verified against stage hashes; audit scripts and tests included with delivery. Intermediate pipeline artifacts available on request.
Access
Preview the public 5% samples on HuggingFace — same schemas, same quality, deterministic subset of the full corpora. Request access for the complete catalog with provenance manifests, audit scripts, and licensing terms.
partnerships@lateos.ai · response within 2 business days