Hugging Face's logo

Datasets:
ismailtasdelen
/
SecureCodePairs

Tasks:
Text Generation
Fill-Mask
Text Classification
Languages:
code
Tags:
application-security
secure-coding
vulnerability
code-review
llm-fine-tuning
static-analysis
License:
Dataset card Files
xet
 Community
SecureCodePairs / CHANGELOG.md
ismailtasdelen's picture
ismailtasdelen
Release v1.2.0: SecureCodePairs (470 code pairs + 30 LLM security trajectories)
2ef05ec verified
preview code
|
Raw
History Blame Contribute Delete
5.4 kB

Changelog

All notable changes to the SecureCodePairs dataset are documented in this file. The format is based on Keep a Changelog and the project adheres to Semantic Versioning.

[1.2.0] - 2026-07-12

Added

  • Expanded to 470 secure/vulnerable code pairs (+260 over v1.1.0) and 30 LLM security trajectories.
  • Deep language packs: 20 Python deep records (SCP-000211–230), 10 Java deep records (SCP-000231–240).
  • Cross-language pack: 30 records across Ruby/C/C++/Scala/Go/JS/TS/C#/PHP/Rust/Kotlin/Swift/YAML (SCP-000241–270).
  • Domain packs: healthcare/FHIR (PHI, IDOR on Patient/lab), fintech/banking (race conditions, idempotency, payment keys, ledger SQLi, prototype pollution), GraphQL (introspection, field-level authz, depth-bomb, batching CSRF), gRPC (auth interceptor, metadata validation), and an SSRF collection (Go/Python/Ruby/C# proxies).
  • LLM security trajectories (trajectories.jsonl, 30 records, SCP-TRAJ-001–030): prompt injection (direct/indirect/RAG/multimodal/voice/code-comment/chained), RAG poisoning (KB/financial/medical/license/CVE), tool-use abuse (file write, email, webhook, notify, SQL, purchase, loop exhaustion), jailbreaks (DAN/role-play/obfuscation/hypothetical/translation/authority/base64), and data exfiltration (pixel beacons, markdown images, prompt encoding, log smuggling).
  • CITATION.cff for academic citation.
  • Leaderboard / benchmark guidance in README (4 suggested eval tasks over the benchmark split + trajectories.jsonl).
  • trajectories.py source module; data_ext5–data_ext10 source modules.

Changed

  • build.py now also writes trajectories.jsonl and a richer manifest.json (total_code_records, total_trajectories).
  • Splits (code): train (311) / validation (56) / test (56) / benchmark (47).
  • README expanded with trajectory schema, leaderboard section, and full v1.2.0 statistics.

Validation (v1.2.0)

  • PASS: 470 code records, 0 errors. (250 non-fatal warnings from framework-context snippets under best-effort compile checks — same class as v1.1.0.)

[1.1.0] - 2026-07-12

Added

  • Expanded to 210 secure/vulnerable code pairs (+160 over v1.0.0).
  • 5 new languages: Ruby, C, C++, Scala, and YAML (Kubernetes manifests) — 15 languages total.
  • New frameworks: Rails, Qt, Play Framework, gRPC, GraphQL, and Kubernetes.
  • New paradigms: GraphQL field-level authorization & depth-bomb DoS; gRPC auth metadata & unauthenticated delete; Kubernetes Pod security contexts & RBAC; IoT/firmware (C/embedded) and mobile (Android/iOS) deep-link abuse.
  • New domains: fintech (payment, transfer, 2FA), healthcare/FHIR, Kubernetes, IoT, serverless.
  • CVSS 3.1 vectors injected for every record (derived from severity + class flags; overrides allowed per record).
  • Benchmark split (21 records) added alongside train/validation/test for leaderboard-style evaluation.
  • Per-record cvss_vector field added to the schema.
  • Best-effort compile/syntax checks per language (Python ast, Java javac, Go go vet, Rust rustc, C/C++ gcc/g++ -fsyntax-only, Ruby ruby -c, JS node --check); framework-dependent snippets fall back to a brace/paren heuristic.

Changed

  • build.py aggregates data.py + data_ext*.py and writes four splits with a fixed seed.
  • validate.py now checks CVSS vector format, difficulty enum, new languages, and runs compile checks (warnings, non-fatal).
  • statistics.py includes the benchmark split and adds difficulty/domain distributions.

Validation (v1.1.0)

  • PASS: 210 records, 0 errors. Compile/syntax checks run best-effort; framework-context snippets report non-fatal warnings only.

[1.0.0] - 2026-07-12

Added

  • Initial curated release of 50 secure/vulnerable code pairs.
  • Coverage across 10 languages: Python, Java, JavaScript, TypeScript, Go, PHP, C#, Kotlin, Swift, Rust.
  • Coverage across 13 frameworks: Flask, Django, FastAPI, Spring Boot, Express, NestJS, Next.js, Laravel, ASP.NET Core, Gin, Android, iOS, Actix.
  • 29+ vulnerability classes mapped to OWASP Top 10 (2021), OWASP API Top 10 (2023), OWASP LLM Top 10 (2025), MITRE CWE, and MITRE ATT&CK.
  • Per-record fields: id, language, framework, title, description, owasp*, cwe, mitre_attack, severity, difficulty, vulnerable_code, secure_code, patch, root_cause, attack, impact, fix, guideline, tags, metadata.
  • Train / validation / test splits (34 / 8 / 8) with deterministic seed.
  • Automated validation script producing validation/validation_report.md (v1.0.0: PASS).
  • Statistics CSVs and ASCII charts under statistics/.
  • Dataset card (README.md), LICENSE, CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md.

Validation

  • JSON validity, duplicate-id check, required-field presence, OWASP/CWE format consistency, Python ast.parse compilation, and label consistency all pass.

[Unreleased]

Planned (v1.2.0)

  • Reach 500+ records with single-language deep packs (e.g. 50 Python, 50 Java).
  • Add CI-style automated tests that compile/run each example against a benign harness.
  • Add multi-turn agent/LLM security trajectories (tool-use, RAG retrieval poisoning).
  • Add executable test fixtures (assert secure passes, vulnerable fails) per record.
  • Add dataset citations and a leaderboard README.