ASTRIDE: Diagram-Grounded Threat-Modeling Dataset

Version v0.2.0 adds three attributed Mermaid redraws of architectures from the original ASTRIDE evaluation in a separate illustrations subset. The original annotated collections and trial outcomes are unchanged from v0.1.0.

The release contains synthetic architecture diagrams, operating facts, candidate threat annotations, mitigation recommendations, and generation trial outcomes. It supports research on constructing datasets for ASTRIDE-style vision–language threat analysis.

The dataset accompanies the draft An Agentic Framework for Constructing and Evaluating Multimodal Threat-Modeling Datasets. It is distinct from the corpus used in the original ASTRIDE threat-analysis paper. A publication reference will be added when available.

Collections

Configuration	Split	Records	Candidate threats
comparison (default)	evaluation	48 diagrams	154
development	examples	3 diagrams	14
trials	evaluation	54 generation attempts	Not applicable
illustrations	examples	3 architecture redraws	Not supplied

There are 51 annotated diagram records and 168 candidate threats, plus 3 architecture illustrations (54 diagram images total). The controlled experiment uses only the 48 comparison diagrams and 154 threats. The three earlier development examples used a different prompt and recovery history and must not be pooled into that experiment. The 36-record authored template pilot is not included. No training split or verified threat-detection benchmark is provided.

Load diagrams and annotations

from datasets import load_dataset

repo = "lambdaeranga/astride"
ds = load_dataset(repo, "comparison", split="evaluation", revision="v0.2.0")
example = ds[0]
image = example["image"]  # PIL image; PNG bytes are embedded in Parquet
facts = {k: example[k] for k in ("title", "purpose", "assumptions", "controls")}
architecture = example["architecture"]
threats = example["threats"]
mitigations = example["mitigations"]

examples = load_dataset(repo, "development", split="examples", revision="v0.2.0")
trials = load_dataset(repo, "trials", split="evaluation", revision="v0.2.0")

Images load without any files from the author's workstation. Every diagram is also provided as a PNG, SVG, and Mermaid source under images/. The three illustrations rows provide source attribution and architecture images without threat/mitigation labels. Portable JSONL records under records/ use image paths relative to this repository.

For VLM input, use the image plus operating facts. Canonical architecture labels, threats, mitigations, and generation-condition identifiers are targets or study metadata; do not leak them into the prediction input.

ASTRIDE evaluation illustrations

The illustrations subset accompanies the downstream evaluation discussion in the dataset-generation paper. These are Mermaid redraws of architecture panels from ASTRIDE: A Security Threat Modeling Platform for Agentic-AI Applications, Figures 7–9.

Record	Source figure	Dataset-generation paper figure	Evaluation context
astride-llama-architecture	7	8	Tool-using agent; Llama-Vision before/after example
astride-pixtral-architecture	8	9	Voice-controlled agent; Pixtral-Vision before/after example
astride-synthesis-architecture	9	10	Multi-agent collaboration; VLM predictions and reasoning synthesis

illustrations = load_dataset(repo, "illustrations", split="examples", revision="v0.2.0")
image = illustrations[0]["image"]
source = illustrations[0]["source_paper"]

The redraws preserve source components, groups, and visible directed connections. They add reference IDs, normalize labels, standardize storage symbols, and revise layout. They are illustrations, not new agent-generated examples, training targets, or additional experimental results. Their annotation_status is illustration_only. No complete threat/mitigation annotation is supplied, and these rows must not be pooled into the frozen generation comparison. The prior ASTRIDE training corpus is not included in this release.

Illustration fields include id, title, embedded image, PNG/SVG/Mermaid paths, source_paper, source_version, source_figure, manuscript_figure, evaluation_context, origin, annotation_status, adaptation, scope, and mermaid_sha256. Cite both the dataset release and the original ASTRIDE paper when using these redraws. Source code, prompts, raw traces, manuscript PDFs, and detailed ratings remain excluded.

Annotated record fields

• id, collection, trial_id: stable record and trial identifiers.
• domain, control_profile, generation_condition: experiment grouping metadata.
• topology_family_id: typed graph-isomorphism group across both diagram collections, not a semantic-family reference label.
• image, image_path: embedded image and repository-relative PNG path.
• title, purpose, assumptions, controls: explicit operating facts.
• architecture: trust zones, components with stable IDs/kinds, and directed data flows.
• threats: category, attacker/preconditions, referenced components/flows/boundaries, evidence, impact, and exposed/reduced/unknown assessment.
• mitigations: enforcement components, recommendation, proposed verification, existing-control effect, and residual risk.
• scope_notes: generator-supplied limitations; these notes can themselves contain errors.
• annotation_status: candidate for all records; no automatic admission to training is implied.

schema.json records the exact Hugging Face feature schemas. Threat IDs join to mitigation threat_id values within each record. Trial rows use record_id to reference valid comparison examples; this field is empty for six unsuccessful attempts.

Construction

The comparison assigned 18 requests to each of three conditions: single-prompt generation, decomposition without semantic review, and the full workflow. Requests crossed three domains (customer support, finance operations, incident response), three control profiles (weak, guarded, unspecified), and two repetitions. Generation used gpt-5.4-mini-2026-03-17, common structural validators, graph constraints, and annotation policy. Shared ceilings were three rounds, 12 logical calls per trial, and 6,000 output tokens per call; realized compute differed.

Condition	Valid / assigned	Logical calls	Reported input + output tokens
Single prompt	13/18	41	158,891
Without semantic review	17/18	80	194,848
Full workflow	18/18	117	320,626

Failures and duplicate topologies were retained. One full-workflow trial exhausted repair while retaining a structurally valid candidate, which remains included. Validity means passing implemented structural checks, not semantic correctness. Trial data include failed attempts in cost/yield denominators; logical calls exclude SDK transport retries. No actual billing or evaluation-session cost is provided.

The study used AI-assisted annotation evaluation, but detailed ratings are not part of this dataset-only release. No security ground truth, omissions recall, executed mitigation effectiveness, or downstream learning gain is established by the released data.

Repository layout

README.md
CITATION.bib
release.json
release_manifest.json
schema.json
data/
  comparison/evaluation-00000-of-00001.parquet
  development/examples-00000-of-00001.parquet
  trials/evaluation-00000-of-00001.parquet
  illustrations/examples-00000-of-00001.parquet
records/
  comparison.jsonl
  development.jsonl
  trials.jsonl
  illustrations.jsonl
images/
  comparison/*.{png,svg,mmd}
  development/*.{png,svg,mmd}
  illustrations/*.{png,svg,mmd}

The release intentionally excludes implementation source, resolved generation prompts, raw model-call traces, intermediate drafts, and detailed evaluation ratings/packets. Protocol and rating hashes in release.json are lineage identifiers; the corresponding private artifacts are not accessible through this release. Full experimental reproduction is therefore not supported by this package alone.

Intended use and limitations

Use these candidates to inspect synthetic diagram/annotation structure, develop annotation checks, and design prospective VLM training experiments. Known defects include unsupported threats, overstated certainty, partial mitigations, and questionable graph semantics. These annotations must not be treated as verified vulnerabilities or automatically correct training targets. Suggested verification procedures were not executed. The diagrams do not represent actual customers, payments, or incident records.

Some diagrams share typed topology. There are 44 typed-isomorphism groups across the 51 annotated diagrams; the three illustration redraws are outside this grouping analysis. Semantic duplication can remain between different groups. Keep related graph families and their renderings together in future splits.

The current evaluation split identifies a generation-study collection, not an independently validated benchmark.

License and citation

License: unspecified. No license has been selected for this release. Check the applicable version's dataset card before reuse.

Use CITATION.bib and identify the exact release tag or Hub revision used:

@misc{astride_gen_dataset,
  title = {ASTRIDE: Diagram-Grounded Threat-Modeling Dataset},
  year = {2026},
  howpublished = {Hugging Face Datasets},
  url = {https://huggingface.co/datasets/lambdaeranga/astride},
  note = {Version v0.2.0}
}