README.md

Durinn Hacktoberfest Retrospective Dataset

Author: Ryan Marinelli & Victor Strandmoe Project: Durinn — Scaling Vibe Coding Auditing
Dataset Type: Security SFT (Supervised Fine-Tuning)
Sources: Scanning GitHub Hacktoberfest 2025 Format: HuggingFace DatasetDict with train and validation splits

---

📌 Overview

This dataset provides security-focused training data derived from analyzing Hacktoberfest 2025 GitHub repositories before and after the event using Semgrep’s OWASP Top 10 ruleset.

It contains two types of supervised fine-tuning (SFT) examples:

1. Repository-level summaries
   High-level assessments of how a repo’s security posture changed between
   before (Oct 1) and after (Nov 1) snapshots.

2. Snippet-level findings
   Per-finding examples showing:

   • Code before and after
   • Git diff between snapshots
   • Semgrep rule, severity, OWASP categories
   • Whether the finding persisted or was fixed

These examples are formatted in instruction–response form for use in training
LLMs to:

• Detect insecure code
• Explain vulnerabilities
• Understand OWASP semantics
• Recognize fixes in Git diffs
• Generate more secure code patterns

This dataset powers Durinn, a pipeline for auditing "vibe code" with Semgrep + LLM security critics.

---

📂 Files Included

train/ and validation/

Both contain rows with:

Field	Type	Description
id	string	Unique example ID
source	string	"repo" or "snippet"
prompt	string	The full instruction + context given to the model
response	string	The ideal model answer
meta	JSON	Structured metadata for downstream analysis

---

🧠 Dataset Schema

Each row follows this schema:

{
  "id": "unique_example_id",
  "source": "repo | snippet",
  "prompt": "<instruction + context + code + diff>",
  "response": "<model answer>",
  "meta": {
    "type": "repo|snippet",
    "repo": "owner/name",
    "path": "path/to/file" ,
    "rule_id": "semgrep.rule.name",
    "severity": "WARNING|ERROR",
    "owasp": ["A03:2025 - ..."],
    "status": "present|fixed"
  }
}

---

🏗️ How the Data Was Generated

✔️ 1. Repo Selection

We selected GitHub repos from hacktoberfest_repos_2025.csv, filtering the top N repositories by stars, creation date, and topical relevance.

✔️ 2. Two Snapshots per Repo

• Before: last commit ≤ 2025-10-01
• After: first commit ≥ 2025-11-01

✔️ 3. Semgrep Scan (OWASP Top 10 Multilang)

Both snapshots were scanned using:

owasp-topten-multilang.yml

Findings include:

• check_id
• message & metadata
• OWASP mappings
• severity
• file, path, and code ranges

✔️ 4. Per-Finding Code Extraction

Each finding includes:

• The affected code range (with ±1 line context)
• Git diff between snapshots
• The finding's persistence/fix status

✔️ 5. SFT Example Construction

We generated two aligned datasets:

🟦 Repo-level

Natural-language summary of security posture changes:

• Δ total findings
• Δ OWASP categories
• Δ rule counts
• High-risk flags
• Explanation sentences

🟩 Snippet-level

Instructional examples showing:

• The insecure code
• The fixed code (if applicable)
• The diff
• Explanation of why the original pattern was insecure

---

🎯 Intended Use Cases

• Fine-tuning models to detect security vulnerabilities
• Teaching LLMs how OWASP categories map to real code patterns
• Training critics to judge code security (“fortification”)
• Building automated code remediation agents
• Benchmarking LLM-based static analysis

This dataset is optimized for:

• TRL SFTTrainer
• LoRA fine-tuning
• Reward modeling / preference learning
• Agent-based security critics (e.g., Durinn MCP)

---

⚠️ Limitations

• Not all repos contain fixes; some only regress.
• Semgrep rule coverage varies by language.
• OWASP category extraction depends on rule metadata correctness.

---

📜 License

This dataset includes:

• Small code excerpts from public GitHub repositories (fair use for research)
• Derived annotations and summaries

Contact for DMCA/removal requests.

---

🙌 Acknowledgments

Special thanks to:

• Hacktoberfest maintainers
• Semgrep.io team
• Durinn Research Group

---

📎 Citation

@dataset{durinn_hacktoberfest_retrospective,
  title={Durinn Hacktoberfest Retrospective},
  author={Marinelli, Ryan},
  year={2025},
  publisher={HuggingFace Datasets},
  howpublished={\url{https://huggingface.co/datasets/zrmarine/Durinn_Hacktoberfest_Retrospective}}
}