Mermaid Expert Corpus 500k is a validated, deduplicated, metadata-rich text-to-Mermaid dataset for companies building, training, evaluating, or benchmarking diagram-generation systems.

This package contains the full 500,000-record accepted corpus and rendered SVG artifacts for every accepted record.

Licensing inquiries: corefidelity@proton.me

Public 1,000-record sample: CoreFidelity/Mermaid_500k_1KSample

Repository access is gated and granted only to licensed users.

Visual Examples

Representative examples from the corpus:

Diagram type	Domain	Mermaid source
Flowchart	auth/identity	`flowchart_145698.mmd`
ER	finance operations	`er_008810.mmd`
Sequence	auth/identity	`sequence_056143.mmd`
Class	ecommerce	`class_654245.mmd`
State	auth/identity	`state_092222.mmd`
Architecture	observability	`architecture_039411.mmd`

Quick Statistics

accepted records       500,000
diagram modalities           6
software/system domains     15
template families           82
rendered SVGs          500,000

Sample Row

A representative flat viewer row:

{
  "id": 56143,
  "diagram_type": "sequence",
  "domain": "auth_identity",
  "subdomain": "auth_login_seq",
  "complexity_tier": "intermediate",
  "template_name": "auth_login_seq",
  "description": "We need to map out the user login flow. The Mobile App sends login credentials to the Authn service. The Authn service then queries the Identity DB to look up the user. If the password is valid, the Authn service issues an access token back to the Mobile App; otherwise, it returns a 401 Unauthorized error.",
  "mermaid": "sequenceDiagram\n    autonumber\n    actor C as Mobile App\n    participant A as Authn service\n    participant D as Identity DB\n    C->>A: Login {credentials}\n    A->>D: Lookup user\n    D-->>A: User record\n    alt Valid password\n        A-->>C: Issue access token\n        else Invalid password\n        A->>C: 401 Unauthorized\n    end",
  "svg_path": "examples/sequence_056143.svg"
}

License Required

No use of this full dataset is permitted without a paid license.

Once licensed, licensees may use the dataset broadly for commercial and internal purposes, including model training, model evaluation, benchmarking, product integration, documentation automation, developer tooling, and derivative dataset workflows, subject to the written license agreement.

See LICENSE for the package terms and contact corefidelity@proton.me for licensing.

Package Contents

500k_full/
  README.md
  LICENSE
  manifest.json
  records.jsonl
  viewer.jsonl
  index.csv
  examples/
  svgs/

Notes:

records.jsonl contains the 500,000 validated accepted records.
viewer.jsonl is a flat Hugging Face Dataset Viewer split with core fields and SVG paths.
index.csv is a lightweight browsing index with ids, diagram metadata, validation flags, Mermaid/description sizes, coverage tags, and SVG paths.
examples/ contains six representative SVG, Mermaid, and compact JSON examples.
svgs/ contains rendered SVGs for the accepted records.
In this working copy, records.jsonl is hard-linked from the canonical accepted JSONL file to avoid duplicating the 2 GB record file.

Accepted Corpus Size

accepted records      500,000
rendered SVGs         500,000
diagram types               6
commercial domains         15
template families          82

Diagram Type Distribution

flowchart       159,727
er              105,371
sequence         85,207
class            53,220
state            50,824
architecture     45,651

Difficulty Distribution

expert          193,533
advanced        189,917
intermediate     68,809
principal        44,012
basic             3,729

Domain Coverage

The corpus spans practical software and system-design domains, including:

ecommerce
auth and identity
data platforms
messaging and streaming
distributed systems
DevOps and CI/CD
ML infrastructure
observability
finance operations
payments
security
cloud architecture
enterprise SaaS
healthcare operations
support operations

Top domain counts:

ecommerce              90,480
auth_identity          83,767
data_platforms         50,423
messaging_streaming    50,111
distributed_systems    49,754
devops_ci_cd           47,672
ml_infrastructure      40,460
observability          27,632
finance_ops            19,754
payments               18,191
security                9,587
cloud_architecture      4,347
enterprise_saas         4,076
healthcare_ops          1,936
support_ops             1,810

What The Dataset Is For

The full corpus is designed for companies that need high-quality Mermaid generation data across real software documentation and system-design use cases:

fine-tuning text-to-diagram models
training Mermaid-generation models
evaluating Mermaid syntax and semantic quality
building diagram-type routers and prompt classifiers
constructing benchmark and eval suites
curriculum training by diagram type, domain, and difficulty
developer-tool and documentation-automation workflows
synthetic data augmentation for software and systems documentation

Quality And Validation

The accepted corpus has been filtered and audited for:

parse/render-valid Mermaid
exact Mermaid duplicate removal
exact prompt/description duplicate removal
structural near-duplicate filtering
artifact and broken-text detection
metadata-rich slicing by diagram type, domain, difficulty tier, and template family
filtering provenance through the build pipeline

Final accepted-corpus audit:

accepted records                         500,000
exact Mermaid duplicate clusters               0
exact prompt duplicate clusters                0
structural near-duplicate clusters             0
artifact records                               0

SVG audit:

rendered SVG files       500,000
failed SVG renders             0

Release Integrity

The licensed corpus is distributed as clean training/evaluation data.

It does not include watermark strings, canary records, trap examples, poisoned samples, or intentionally degraded records. The accepted release is intended to be directly usable for model training, evaluation, and benchmark construction without hidden compromise mechanisms.

Dataset access and commercial rights are controlled through licensing rather than embedded watermark/canary contamination.

Record Format

Records are JSONL. Field sets vary by diagram type, but records generally include:

stable record id
diagram type
domain
complexity tier
template family
natural-language prompt/description
Mermaid source
diagram-specific metadata
quality/filtering provenance

The schema is intentionally metadata-rich so buyers can slice the corpus by modality, commercial domain, difficulty tier, template family, and workflow type.

Licensing

This full dataset is a commercial data product. No inspection, copying, redistribution, model training, evaluation, product integration, commercial implementation, or other use is permitted unless a paid license has been granted.

Paid licensees may use the dataset broadly according to their written license agreement.

Licensing contact: corefidelity@proton.me

Downloads last month: 16

Total file size:

45.6 GB