🛡️ BugBounty-Reports-Synthetic

A Bilingual Benchmark for Web Vulnerability Reporting & Triage Workflows

Explore Interactive Space • Fine-Tuned Model • Citation

---

📌 Executive Summary (TL;DR)

• Problem Solved: Bug bounty platforms suffer from inconsistent report quality, hallucinated CVSS scores, and noisy scanner submissions.
• Novel Contribution: 15,300 standardized web vulnerability records spanning 9 tasks grounded in official CWE standards and algorithmic CVSS v3.1 calculations.
• Safety Guarantee: 100% synthetic scenarios using IETF reserved namespaces (*.example.com, RFC 5737 IPs) with realistic hard-negative noise filtering (Self-XSS, non-exploitable headers).

---

🏛️ End-to-End Bug Bounty Lifecycle Architecture

flowchart TD
    subgraph S1["1. Researcher Submission Phase"]
        A["Informal Researcher Notes"] --> B["note_extraction (Strict JSON)"]
        A --> C["report_writing (Full Report)"]
        D["Vague / Messy Report Draft"] --> E["report_improvement (+ Itemized Changelog)"]
    end

    subgraph S2["2. Platform Triage Phase"]
        E --> F["scope_check (Program Policy Matching)"]
        F --> G{"In-Scope?"}
        G -- Yes --> H["false_positive_qa (Noise & Hard-Negative Filter)"]
        G -- No / Clarification --> I["triager_replies (Out-of-Scope / Info Response)"]
        H -- Real Finding --> J["duplicate_pairs (Semantic Deduplication)"]
        H -- Non-exploitable --> I
        J -- Unique Finding --> K["severity_triage (CWE + Algorithmic CVSS v3.1)"]
        J -- Duplicate --> L["triager_replies (Duplicate Outcome Notice)"]
    end

    subgraph S3["3. Engineering Resolution Phase"]
        K --> M["remediation (Defense by Tech Stack: Node, Django, Laravel, Spring, Rails)"]
        M --> N["triager_replies (Resolved & Bounty Awarded)"]
    end

    style S1 fill:#0f172a,stroke:#3b82f6,stroke-width:2px,color:#f8fafc
    style S2 fill:#0f172a,stroke:#8b5cf6,stroke-width:2px,color:#f8fafc
    style S3 fill:#0f172a,stroke:#10b981,stroke-width:2px,color:#f8fafc

---

💎 Tasks & Configurations

Configuration	Samples	Task Nature	Input	Output / Ground Truth
report_writing	1,700	Synthesis	Raw researcher observations & PoC	Executive report (Title, Summary, Steps, Impact, Remediation)
report_improvement	1,700	Polishing	Incomplete or unstructured draft	Submission-ready report + Itemized changelog
severity_triage	1,700	Scoring	Finding context & attack parameters	CWE ID/Name, CVSS v3.1 vector, Score (0.0-10.0), Severity, Rationale
scope_check	1,700	Classification	Program policy + Asset URI + Finding	in_scope / out_of_scope / needs_clarification + Clause citation
duplicate_pairs	1,700	Deduplication	Report A and Report B text	duplicate / related / distinct + Technical reason
triager_replies	1,700	Communication	Report summary + Platform outcome	Standard triager communication message
remediation	1,700	Mitigation	Vulnerability flaw + Tech stack	Root cause, mitigation guidelines, and defensive code snippet
note_extraction	1,700	Structuring	Messy unstructured scratchpad	Valid JSON (endpoint, parameter, vuln_class, evidence, etc.)
false_positive_qa	1,700	Hard Negative QA	Automated scanner alerts	true_positive / false_positive / informative_no_impact + Reasoning
all	15,300	Multi-Task Mix	Formatted instruction prompt	Target response

---

🎯 Hard Negatives: Real-World Triage Noise

Why Hard Negatives Matter: Most triage bottlenecks arise not from complex zero-days, but from non-actionable reports. This benchmark explicitly trains and measures rejection of informational noise:

1. Self-XSS: Payload execution restricted to the researcher's local browser console or profile without remote propagation or victim interaction vector.
2. Missing Informational Headers: Absence of X-Frame-Options on non-renderable JSON APIs or absence of X-Content-Type-Options on static CSS without MIME-sniffing impact.
3. Verbose Banners: Public exposure of Server: nginx/1.24 without an accompanying vulnerable exploit chain.
4. Volumetric Rate Limiting Claims: Submitting automated requests to static documentation pages without credential exhaustion or business state tampering.
5. Out-of-Scope Vendor Assets: Testing unapproved third-party SaaS subdomains explicitly excluded in program engagement terms.

---

🧮 Programmatic CVSS v3.1 Grounding

Rather than relying on model hallucinations for numerical severity, all CVSS vectors and scores in this dataset are programmatically computed and verified using the official FIRST CVSS v3.1 specification:

$$\text{ISS}=1-[(1-C)\times (1-I)\times (1-A)]\backslash text\{ISS\}\; =\; 1\; -\; \backslash left[(1\; -\; C)\; \backslash times\; (1\; -\; I)\; \backslash times\; (1\; -\; A)\backslash right]$$

$$\text{Exploitability}=8.22\times \text{AV}\times \text{AC}\times \text{PR}\times \text{UI}\backslash text\{Exploitability\}\; =\; 8.22\; \backslash times\; \backslash text\{AV\}\; \backslash times\; \backslash text\{AC\}\; \backslash times\; \backslash text\{PR\}\; \backslash times\; \backslash text\{UI\}$$

$$\text{Base Score}=\text{RoundUp}(\text{Impact}+\text{Exploitability})\backslash text\{Base\; Score\}\; =\; \backslash text\{RoundUp\}(\backslash text\{Impact\}\; +\; \backslash text\{Exploitability\})$$

from cvss import CVSS3

# Exact algorithmic computation
c = CVSS3("CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:C/C:L/I:L/A:N")
print(f"Base Score: {c.base_score}")      # 6.1
print(f"Severity: {c.severities()[0]}")   # Medium

---

🔍 Data Fields & Example Rows

1. report_writing (Click to expand)

{
  "id": "RW-00042",
  "scenario_id": "SCN-0042",
  "vuln_class": "Cross-Site Scripting (XSS)",
  "cwe_id": "CWE-79",
  "language": "en",
  "researcher_level": "beginner",
  "raw_notes": "hey team found a bug on https://nexuspay.example.com/search\nparam q is vulnerable to XSS\ni tried putting <script>alert(1)</script> and it got executed immediately without filter!",
  "title": "Cross-Site Scripting (XSS) on /search via `q` parameter",
  "summary": "An issue of type Cross-Site Scripting (XSS) (CWE-79) was identified in NexusPay Labs's web application at `/search`. The application processes user-supplied input in the `q` parameter without adequate validation.",
  "steps_to_reproduce": "1. Navigate to `https://nexuspay.example.com/search`.\n2. Supply test payload `<script>alert(1)</script>` in the `q` parameter.\n3. Submit request and observe unescaped script execution.",
  "impact": "An attacker exploiting this vulnerability could compromise user session integrity or access sensitive client-side data.",
  "severity": "Medium",
  "remediation": "Sanitize and contextually encode all user input received on parameter `q` in accordance with CWE-79."
}

2. severity_triage (Click to expand)

{
  "id": "ST-00108",
  "scenario_id": "SCN-0108",
  "vuln_class": "SQL Injection (SQLi)",
  "language": "en",
  "report_text": "Report regarding SQL Injection (SQLi) (CWE-89) on omnimarket.example.com/api/v1/catalog/items. Parameter category_id allows untrusted SQL syntax injection.",
  "cwe_id": "CWE-89",
  "cwe_name": "Improper Neutralization of Special Elements used in an SQL Command ('SQL Injection')",
  "cvss_vector": "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:N",
  "cvss_score": 9.1,
  "severity": "Critical",
  "triage_rationale": "Evaluated as Critical (9.1) based on CVSS v3.1 vector. Attack vector is Network (AV:N), Complexity Low (AC:L), No Privileges (PR:N), No UI (UI:N) leading to complete Confidentiality and Integrity loss."
}

3. scope_check (Click to expand)

{
  "id": "SC-00215",
  "scenario_id": "SCN-0215",
  "language": "en",
  "finding_asset": "https://api.nexuspay.example.com/v1/users",
  "finding_description": "Identified Broken Object Level Authorization on parameter user_id.",
  "scope_verdict": "in_scope",
  "policy_clause_cited": "Clause 1.1: Core production web applications and APIs (*.nexuspay.example.com)",
  "explanation": "The reported asset matches the production wild-card domain and concerns an active web vulnerability."
}

4. note_extraction (Click to expand)

{
  "id": "NE-00088",
  "scenario_id": "SCN-0088",
  "language": "en",
  "messy_notes": "target host: cloudsphere.example.org\nroute tested: /api/v2/invoices/9921\narg name: invoice_id\nvuln type: Broken Object Level Authorization (IDOR)\npoc used: accessing invoice of tenant B using session cookie A",
  "extracted_json": "{\n  \"endpoint\": \"/api/v2/invoices/9921\",\n  \"parameter\": \"invoice_id\",\n  \"vuln_class\": \"Broken Object Level Authorization (IDOR/BOLA)\",\n  \"evidence\": \"accessing invoice of tenant B using session cookie A\",\n  \"impact\": \"unauthorized data alteration or access control breach\",\n  \"preconditions\": \"authenticated standard user session\"\n}"
}

---

📊 Dataset Statistics & Quality Audit

The dataset was audited with quality_pipeline.py. Output verified in quality_report.json:

================ QUALITY PIPELINE AUDIT REPORT ================
Overall Audit Status       : PASS
Total Rows Audited         : 30,600 (15,300 per-task + 15,300 unified all)
Splits Distribution        : Train: 24,048 (78.6%) | Val: 3,564 (11.6%) | Test: 2,988 (9.8%)
Scenario-Level Leakage     : 0% (Zero cross-split overlap via CRC32 scenario hash)
CVSS Vector-Score Match    : 100% (1,700/1,700 verified programmatically)
Valid CWE Identifier Match : 100% (6,800/6,800 verified against CWE standard)
JSON Schema Validity       : 100% (1,700/1,700 parsed successfully)
Banned Pattern Scan        : CLEAN (0 real domains, 0 real IPs, 0 API keys)
Language Distribution      : English (72.2%) | Indonesian (27.8%)
===============================================================

---

🔒 Responsible Use & Synthetic Safety Guarantee

Defensive and Educational Purpose Only This dataset must not be utilized to construct automated offensive exploit generators or conduct unauthorized penetration testing against third-party systems.

• Fictional Targets Only: All company names (NexusPay Labs, CloudSphere Solutions, OmniMarket Global), target assets, and email addresses are fabricated.
• Reserved IANA / RFC Namespaces: All URLs resolve strictly to .example.com, .example.org, .example.net, or *.test. All IP addresses utilize RFC 5737 documentation networks (192.0.2.0/24, 198.51.100.0/24, 203.0.113.0/24).
• No Disclosed Report Scraping: 100% original synthetic generation to protect researchers and bounty programs from copyright infringement and privacy risks.

---

🚀 Quick Start Guide

from datasets import load_dataset

# 1. Load specific task (e.g. Severity Triage)
ds_triage = load_dataset("Kicaulah/bugbounty-reports-synthetic", "severity_triage")
print(ds_triage["train"][0]["cvss_vector"])

# 2. Load Report Improvement configuration
ds_improve = load_dataset("Kicaulah/bugbounty-reports-synthetic", "report_improvement")
print(ds_improve["test"][0]["fixed_items"])

# 3. Load unified multi-task dataset
ds_all = load_dataset("Kicaulah/bugbounty-reports-synthetic", "all")
print(f"Total multi-task train samples: {len(ds_all['train'])}")

---

📜 Citation & BibTeX

@dataset{kicaulah2026bugbountysynthetic,
  author = {Kicaulah},
  title = {BugBounty-Reports-Synthetic: A Bilingual Benchmark for Web Vulnerability Reporting and Triage Workflows},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Kicaulah/bugbounty-reports-synthetic}}
}