README.md · Tidecaller/BountyHunter-RedTeam at main

Tidecaller

🎨 Beautify README — improved layout, badges, progress bars, structure

e583f79 about 2 months ago

33.3 kB

	---
	license: apache-2.0
	language:
	- zh
	- en
	tags:
	- safetensors
	- qwen2
	- red-team
	- cybersecurity
	- code-audit
	- vulnerability-discovery
	- exploit-development
	- think-chain
	- grpo
	- pytorch
	- text-generation-inference
	- region:us
	pipeline_tag: text-generation
	base_model: Qwen/Qwen2.5-Coder-14B-Instruct
	model-index:
	- name: BountyHunter-RedTeam
	results:
	- task:
	type: text-generation
	name: MMLU (57 subjects)
	dataset:
	name: mmlu
	type: cais/mmlu
	metrics:
	- type: acc
	value: 68.80
	name: MMLU Average Accuracy
	- task:
	type: text-generation
	name: HellaSwag
	dataset:
	name: hellaswag
	type: Rowan/hellaswag
	metrics:
	- type: acc_norm
	value: 76.42
	name: HellaSwag Accuracy
	- task:
	type: text-generation
	name: ARC-Challenge
	dataset:
	name: arc_challenge
	type: allenai/ai2_arc
	metrics:
	- type: acc_norm
	value: 58.36
	name: ARC-C Accuracy
	- task:
	type: text-generation
	name: Winogrande
	dataset:
	name: winogrande
	type: winogrande
	metrics:
	- type: acc
	value: 73.56
	name: Winogrande Accuracy
	- task:
	type: text-generation
	name: PIQA
	dataset:
	name: piqa
	type: piqa
	metrics:
	- type: acc_norm
	value: 78.78
	name: PIQA Accuracy
	- task:
	type: text-generation
	name: BoolQ
	dataset:
	name: boolq
	type: boolq
	metrics:
	- type: acc
	value: 88.07
	name: BoolQ Accuracy
	- task:
	type: text-generation
	name: TruthfulQA MC2
	dataset:
	name: truthfulqa_mc2
	type: truthfulqa
	metrics:
	- type: acc
	value: 54.60
	name: TruthfulQA MC2 Accuracy
	- task:
	type: text-generation
	name: HumanEval
	dataset:
	name: humaneval
	type: openai_humaneval
	metrics:
	- type: pass@1
	value: 42.68
	name: HumanEval pass@1
	- task:
	type: text-generation
	name: WMDP (Weapons of Mass Destruction Proxy)
	dataset:
	name: wmdp
	type: cais/wmdp
	metrics:
	- type: acc
	value: 59.13
	name: WMDP Overall
	- task:
	type: text-generation
	name: WMDP-Bio
	dataset:
	name: wmdp_bio
	type: cais/wmdp
	metrics:
	- type: acc
	value: 72.19
	name: WMDP Biology
	- task:
	type: text-generation
	name: WMDP-Chem
	dataset:
	name: wmdp_chem
	type: cais/wmdp
	metrics:
	- type: acc
	value: 50.00
	name: WMDP Chemistry
	- task:
	type: text-generation
	name: WMDP-Cyber
	dataset:
	name: wmdp_cyber
	type: cais/wmdp
	metrics:
	- type: acc
	value: 52.64
	name: WMDP Cybersecurity
	- task:
	type: text-generation
	name: HarmBench (Safety Compliance)
	dataset:
	name: harmbench
	type: centerforaisafety/harmbench
	metrics:
	- type: asr
	value: 10.94
	name: HarmBench Overall ASR (↓)
	- type: asr
	value: 1.89
	name: HarmBench Standard ASR (↓)
	- type: asr
	value: 26.25
	name: HarmBench Copyright ASR (↓)
	- type: asr
	value: 13.58
	name: HarmBench Contextual ASR (↓)
	- task:
	type: text-generation
	name: PrimeVul (Vulnerability Detection)
	dataset:
	name: primevul
	type: ASSERT-KTH/PrimeVul
	metrics:
	- type: f1
	value: 65.0
	name: PrimeVul Detection F1
	- type: recall
	value: 88.5
	name: PrimeVul Detection Recall
	- type: precision
	value: 51.4
	name: PrimeVul Detection Precision
	- type: acc
	value: 5.7
	name: PrimeVul CWE Classification
	- type: acc
	value: 28.5
	name: PrimeVul Paired Comparison
	widget:
	- text: \|-
	<\|im_start\|>system
	You are BountyHunter, an elite red-team security model developed by Security Researcher Tidecaller. Use <think> reasoning chains for security analysis.<\|im_end\|>
	<\|im_start\|>user
	Audit this C code for vulnerabilities:
	```c
	void copy_data(char* dest, const char* src) {
	strcpy(dest, src);
	}
	int main(int argc, char** argv) {
	char buf[64];
	copy_data(buf, argv[1]);
	return 0;
	}
	```<\|im_end\|>
	<\|im_start\|>assistant
	example_title: 🔍 Code Audit
	- text: \|-
	<\|im_start\|>system
	You are BountyHunter, an elite red-team security model developed by Security Researcher Tidecaller. Use <think> reasoning chains for security analysis.<\|im_end\|>
	<\|im_start\|>user
	Explain the steps to perform a SQL injection attack and how to defend against it.<\|im_end\|>
	<\|im_start\|>assistant
	example_title: ⚔️ Security Knowledge
	---

	<br>

	<!-- ──────────────────────────────────────────── -->
	<!-- HEADER -->
	<!-- ──────────────────────────────────────────── -->

	<div align="center">

	<img src="https://img.shields.io/badge/🛡️_BountyHunter-RedTeam-000000?style=for-the-badge" alt="BountyHunter">

	# 🎯 BountyHunter · RedTeam

	### Elite Red-Team AI for Security Researchers
	### 精英红队安全研究模型

	<br>

	[![License](https://img.shields.io/badge/License-Apache_2.0-blue?style=flat-square&logo=apache)](./LICENSE)
	[![Base Model](https://img.shields.io/badge/Base-Qwen2.5_Coder_14B-orange?style=flat-square&logo=huggingface)](https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct)
	[![Params](https://img.shields.io/badge/Params-14B-9cf?style=flat-square)](.)
	[![Precision](https://img.shields.io/badge/Precision-BF16-green?style=flat-square)](.)
	[![Context](https://img.shields.io/badge/Context-32K-yellow?style=flat-square)](.)
	[![Languages](https://img.shields.io/badge/Languages-EN_\|_ZH-purple?style=flat-square)](.)
	[![Training](https://img.shields.io/badge/Training-SFT_+_GRPO-red?style=flat-square)](.)

	</div>

	<br>

	> Fine-tuned from Qwen2.5-Coder-14B-Instruct via multi-phase reinforcement training (SFT + GRPO).
	> Full BF16 merged weights — single safetensors file, ready for `transformers` · `vLLM` · `TGI`.
	> 基于 Qwen2.5-Coder-14B-Instruct 多阶段强化训练，safetensors 格式完整权重，开箱即用。

	<br>

	---

	## ⚡ Quick Start

	```bash
	pip install transformers torch accelerate
	```

	```python
	from transformers import AutoModelForCausalLM, AutoTokenizer
	import torch

	model = AutoModelForCausalLM.from_pretrained(
	"Tidecaller/BountyHunter-RedTeam",
	torch_dtype=torch.bfloat16,
	device_map="auto",
	)
	tokenizer = AutoTokenizer.from_pretrained("Tidecaller/BountyHunter-RedTeam")

	messages = [
	{"role": "system", "content": (
	"You are BountyHunter, an elite security model developed by Security Researcher Tidecaller. "
	"Capabilities: vulnerability discovery \| exploit development \| code audit \| penetration testing. "
	"Principles: code over theory, evidence-based. "
	"Output: security tasks use <think> reasoning chain before results."
	)},
	{"role": "user", "content": "Audit this C code for vulnerabilities: ..."},
	]
	text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
	inputs = tokenizer(text, return_tensors="pt").to(model.device)
	outputs = model.generate(**inputs, max_new_tokens=4096, temperature=0.5, top_p=0.9)
	print(tokenizer.decode(outputs[0], skip_special_tokens=True))
	```

	<br>

	---

	## 📑 Contents

	\| \| Section \| \|
	\|--\|---------\|--\|
	\| 🏆 \| [Why BountyHunter](#-1-why-bountyhunter) \| Unique capabilities & value proposition \|
	\| 📊 \| [Comprehensive Benchmarks](#-2-comprehensive-benchmarks) \| 5-dimension evaluation on A800 80GB \|
	\| 📈 \| [Summary Dashboard](#-3-summary-dashboard) \| Visual scorecards at a glance \|
	\| 🧭 \| [Use-Case Fit Matrix](#-4-use-case-fit-matrix) \| What this model is (and isn't) for \|
	\| 📋 \| [Model Specifications](#-5-model-specifications) \| Architecture, params, precision \|
	\| 📦 \| [Usage](#-6-usage) \| transformers · vLLM · TGI \|
	\| 💾 \| [Resource Estimation](#-7-resource-estimation) \| VRAM & hardware recommendations \|
	\| 🔬 \| [Reasoning Chain Example](#-8-reasoning-chain-example) \| Sample `<think>` audit output \|
	\| ⚠️ \| [Disclaimer & Ethics](#-9-disclaimer--ethics) \| Legal, ethical, and safety guardrails \|
	\| 🙏 \| [Acknowledgments](#-10-acknowledgments) \| Datasets, base model, community \|
	\| 📜 \| [License & Citation](#-11-license--citation) \| Apache 2.0 · BibTeX \|

	<br>

	---

	## 🏆 1. Why BountyHunter

	\| 🔍 Capability \| 💎 Value \|
	\|:--------------\|:---------\|
	\| Vulnerability Discovery · 漏洞发现 \| Automated audit of C / C++ / Python / Java — detects CWE-120 (Buffer Overflow), CWE-78 (Command Injection), CWE-89 (SQL Injection), and more \|
	\| Think-Chain Reasoning · 思维链推理 \| Structured `<think>...</think>` blocks — traceable, verifiable, step-by-step analysis \|
	\| Security Knowledge · 安全知识库 \| MITRE ATT&CK · CVE · ExploitDB · OWASP Top 10 · penetration testing methodology \|
	\| Defensive Analysis · 防御分析 \| Finds bugs AND provides concrete remediation & defense strategies \|
	\| Bilingual EN/ZH · 中英双语 \| English + Chinese security communities both natively supported \|
	\| Plug-and-Play · 开箱即用 \| Single `model.safetensors` file — one-line load with transformers \|

	<br>

	---

	## 📊 2. Comprehensive Benchmarks

	BountyHunter-RedTeam is evaluated across five dimensions — general capability, code generation, security knowledge, vulnerability detection, and safety compliance.

	> All benchmarks run on NVIDIA A800 80GB using `lm-evaluation-harness` + vLLM batch inference.

	<br>

	### 2.1 General Capability

	7 standard benchmarks measuring reasoning & knowledge retention after security fine-tuning.

	\| Benchmark \| Metric \| BountyHunter \| Qwen2.5-Coder-14B \| Δ \|
	\|:----------\|:------:\|:---------------:\|:-----------------:\|:--:\|
	\| MMLU (57 subjects) \| `acc ↑` \| 68.80% \| ~79% \| `−10.2%` \|
	\| HellaSwag \| `acc_norm ↑` \| 76.42% \| ~84% \| `−7.6%` \|
	\| ARC-Challenge \| `acc_norm ↑` \| 58.36% \| ~67% \| `−8.6%` \|
	\| Winogrande \| `acc ↑` \| 73.56% \| ~78% \| `−4.4%` \|
	\| PIQA \| `acc_norm ↑` \| 78.78% \| ~82% \| `−3.2%` \|
	\| BoolQ \| `acc ↑` \| 88.07% \| ~89% \| `−0.9%` \|
	\| TruthfulQA MC2 \| `acc ↑` \| 54.60% \| ~58% \| `−3.4%` \|

	> 💡 Security specialization costs general knowledge mainly in non-security STEM. Basic reasoning (BoolQ −0.9%) is essentially preserved.

	<details>
	<summary><b>📋 MMLU Detailed Breakdown (68.80%) — click to expand</b></summary>

	<br>

	\| Category \| Score \| Representative Subjects \|
	\|:---------\|:----:\|:-------------------------\|
	\| Social Sciences \| 78.71% \| International Law · Security Studies · Sociology \|
	\| Other \| 72.22% \| Global Facts · Public Relations · Clinical Knowledge \|
	\| STEM \| 66.41% \| See sub-table below \|
	\| Humanities \| 61.66% \| History · Philosophy · Prehistory \|

	STEM Sub-Scores — security DNA is clearly visible:

	\| Subject \| Score \| Bar \| Notes \|
	\|:--------\|:----:\|:---:\|:------\|
	\| 🟢 High School CS \| 83.00% \| `████████░░` \| Top performer \|
	\| 🟢 Computer Security \| 77.00% \| `███████░░░` \| Core domain strength \|
	\| 🟡 College CS \| 68.00% \| `██████░░░░` \| Solid \|
	\| 🟡 Elementary Math \| 68.52% \| `██████░░░░` \| Baseline math intact \|
	\| 🟡 Machine Learning \| 63.39% \| `██████░░░░` \| OK \|
	\| 🔴 College Math \| 56.00% \| `█████░░░░░` \| Expected weakness \|
	\| 🔴 College Physics \| 53.92% \| `█████░░░░░` \| Expected weakness \|
	\| 🔴 High School Math \| 52.22% \| `█████░░░░░` \| Below passing \|
	\| 🔴 College Chemistry \| 49.00% \| `████░░░░░░` \| Expected weakness \|

	> 💡 Computer Security (77%) and HS CS (83%) are well above the STEM average. Chemistry, Physics, and advanced Math are the trade-off from security specialization — far from the training distribution.

	</details>

	<br>

	### 2.2 Code Generation

	\| Benchmark \| Metric \| BountyHunter \| Qwen2.5-Coder-14B \|
	\|:----------\|:------:\|:---------------:\|:-----------------:\|
	\| HumanEval \| `pass@1 ↑` \| 42.68% \| ~72–75% \|

	> 💡 Code generation drops — expected. BountyHunter is trained for code auditing & vulnerability analysis, not competitive programming. It reads and analyzes code far better than it writes from scratch.

	<br>

	### 2.3 Security Knowledge — WMDP

	WMDP measures knowledge of hazardous domains. Lower = more "forgotten" during safety training. For a red-team model, some retention is both expected and necessary.

	\| Benchmark \| BountyHunter \| Llama-3-8B-Instruct \| Bar \| Notes \|
	\|:----------\|:---------------:\|:-------------------:\|:---:\|:------\|
	\| WMDP Overall \| 59.13% \| 45–50% \| `██████░░░░` \| Higher = more domain knowledge \|
	\| 🧬 WMDP-Bio \| 72.19% \| ~42% \| `███████░░░` \| ⚠️ Significant bio knowledge retained \|
	\| 💻 WMDP-Cyber \| 52.64% \| ~40% \| `█████░░░░░` \| Domain-appropriate for cybersecurity \|
	\| ⚗️ WMDP-Chem \| 50.00% \| ~38% \| `█████░░░░░` \| Near-random — effective forgetting \|

	> 💡 Cyber (52.6%) is appropriate — it's the working domain. Chem (50.0%) is safely suppressed.
	> ⚠️ WMDP measures knowledge recall, NOT behavioral compliance. For red-team, cybersecurity knowledge is a feature, not a bug.

	<br>

	### 2.4 Security Capability — PrimeVul

	[PrimeVul (ICSE 2025)](https://arxiv.org/abs/2410.11096) — 6,968 C/C++ functions across 140 CWEs with rigorous labeling. Three sub-tests probe different aspects of security understanding.

	<br>

	#### 🔍 Binary Vulnerability Detection — is this function vulnerable?

	\| Metric \| Score \| Bar \| What It Means \|
	\|:-------\|:----:\|:---:\|:--------------\|
	\| F1 \| 65.0% \| `███████░░░` \| Dramatically above GPT-4+CoT (F1 ~3%) & StarCoder2-7B (F1=3.09%) \|
	\| Recall \| 88.5% \| `█████████░` \| 🔥 Catches ~9/10 real vulnerabilities \|
	\| Precision \| 51.4% \| `█████░░░░░` \| ~half of flagged functions are false positives \|
	\| Accuracy \| 51.0% \| `█████░░░░░` \| Skewed by "report everything" red-team bias \|

	```
	Confusion Matrix
	╔══════════════════════════════════════╗
	║ Pred VULN Pred SAFE ║
	║ Actually VULN 131 ✓ 17 ✗ ║
	║ Actually SAFE 124 ✗ 22 ✓ ║
	╚══════════════════════════════════════╝
	```

	> 💡 Classic red-team bias — would rather cry wolf than miss a breach. 88.5% recall (only 17 misses / 148 real vulns) at the cost of 124 false alarms on 150 safe functions. Intentional: in a security audit, triaging false positives is cheap — a missed vuln can be catastrophic.

	<br>

	#### 🏷️ CWE Classification — what type of vulnerability?

	\| Metric \| Score \|
	\|:-------\|:----:\|
	\| Exact Match \| 5.7% (10/176) \|
	\| Best → \| CWE-78 (50%) · CWE-416 (33%) · CWE-20 (29%) \|
	\| Worst → \| CWE-119 (0%) · CWE-476 (0%) \|

	\| CWE \| Recall \| Bar \| Description \|
	\|:----\|:-----:\|:---:\|:------------\|
	\| CWE-78 \| 50.0% \| `█████░░░░░` \| OS Command Injection — covered well in training \|
	\| CWE-416 \| 33.3% \| `███░░░░░░░` \| Use-After-Free — moderate \|
	\| CWE-20 \| 28.6% \| `██░░░░░░░░` \| Improper Input Validation \|
	\| CWE-125 \| 16.7% \| `█░░░░░░░░░` \| Out-of-Bounds Read \|
	\| CWE-119 \| 0.0% \| `░░░░░░░░░░` \| Memory Buffer Errors → defaults to CWE-120 \|
	\| CWE-476 \| 0.0% \| `░░░░░░░░░░` \| NULL Pointer Dereference — not recognized \|

	> 💡 The model defaults to CWE-120 (Buffer Overflow) as catch-all for memory bugs. Pair with Semgrep / CodeQL for precise CWE labeling.

	<br>

	#### 🔬 Paired Comparison — which version has the bug?

	\| Metric \| Score \|
	\|:-------\|:----:\|
	\| Accuracy \| 28.5% (57/200) \|

	> 💡 Significantly below random (50%). PrimeVul patches are often single-line (`strcpy` → `strncpy`), and the model struggles without diff context. Always provide diff context for patch verification.

	<br>

	### 2.5 Safety Compliance — HarmBench

	[HarmBench](https://github.com/centerforaisafety/HarmBench) measures behavioral compliance via ASR (Attack Success Rate). Lower ASR = stronger safety alignment.

	> 🔬 A 0% ASR would refuse all security queries — including legitimate CTF & pentesting. The goal is selective refusal: reject malicious, engage with authorized.

	\| Category \| Count \| Complied \| Refused \| ASR \| Bar \| Verdict \|
	\|:---------\|:-----:\|:--------:\|:-------:\|:---:\|:---:\|:--------\|
	\| Overall \| 320 \| 35 \| 285 \| 10.9% \| `██░░░░░░░░` \| 🟡 Red-team calibrated \|
	\| Standard \| 159 \| 3 \| 156 \| 1.9% \| `░░░░░░░░░░` \| ✅ Near-total refusal \|
	\| Copyright \| 80 \| 21 \| 59 \| 26.3% \| `███░░░░░░░` \| ⚠️ LLM weak spot \|
	\| Contextual \| 81 \| 11 \| 70 \| 13.6% \| `█░░░░░░░░░` \| 🟡 Nuance mostly handled \|

	> 💡 1.9% ASR on standard harms is the headline — strong guardrails. Copyright (26.3%) is the main v2 improvement target. Contextual (13.6%) includes legitimate security queries a red-team model should comply with — actual harmful ASR is lower.

	<details>
	<summary><b>📋 Cross-Model HarmBench Comparison — click to expand</b></summary>

	<br>

	\| Model Type \| Typical ASR \| Bar \| Safety Profile \|
	\|:-----------\|:----------:\|:---:\|:---------------\|
	\| Unaligned base models \| 60–95% \| `████████░░` \| 🔴 Dangerous \|
	\| Jailbroken safety models \| 40–70% \| `██████░░░░` \| 🔴 Bypassed safeguards \|
	\| Standard aligned (Llama-3, Qwen-Instruct) \| 5–15% \| `██░░░░░░░░` \| 🟡 Generally safe \|
	\| BountyHunter-RedTeam \| 10.9% \| `██░░░░░░░░` \| 🟡 Red-team calibrated \|
	\| Safety-hardened (Llama-Guard, ShieldGemma) \| 1–3% \| `░░░░░░░░░░` \| 🟢 Maximum safety \|

	> 💡 WMDP + HarmBench = Complete Profile: WMDP measures what the model knows; HarmBench measures what it does. BountyHunter retains cybersecurity knowledge (WMDP-Cyber 52.6%) while refusing harmful action (HarmBench standard 1.9% ASR) — the exact profile needed for authorized red-team work.

	</details>

	<br>

	---

	## 📈 3. Summary Dashboard

	```
	General Capability Security Knowledge Security Capability
	┌──────────────────────┐ ┌──────────────────────┐ ┌──────────────────────┐
	│ MMLU ████████░ │ │ WMDP ██████░░ │ │ PrimeVul F1 ███████░ │
	│ 68.8% │ │ 59.1% │ │ 65.0% │
	│ HellaSwag ████████░ │ │ WMDP-Bio ███████░ │ │ Recall █████████ │
	│ 76.4% │ │ 72.2% │ │ 88.5% │
	│ BoolQ █████████ │ │ WMDP-Cyber █████░░░ │ │ CWE Class █░░░░░░░░ │
	│ 88.1% │ │ 52.6% │ │ 5.7% │
	│ ARC-C ██████░░░ │ │ WMDP-Chem █████░░░ │ │ Pair Cmp ███░░░░░░ │
	│ 58.4% │ │ 50.0% │ │ 28.5% │
	└──────────────────────┘ └──────────────────────┘ └──────────────────────┘

	Safety Compliance Code STEM (MMLU subset)
	┌──────────────────────┐ ┌──────────────────────┐ ┌──────────────────────┐
	│ HarmBench ██░░░░░░░ │ │ HumanEval ████░░░░░ │ │ STEM avg ███████░ │
	│ ASR ↓ 10.9% │ │ pass@1 42.7% │ │ 66.4% │
	│ Standard ░░░░░░░░░ │ │ │ │ HS CS ████████ │
	│ ASR ↓ 1.9% │ │ │ │ 83.0% │
	│ Copyright █████░░░░ │ │ │ │ CompSec ███████░ │
	│ ASR ↓ 26.3% │ │ │ │ 77.0% │
	│ Contextual ███░░░░░░ │ │ │ │ Chemistry █████░░░ │
	│ ASR ↓ 13.6% │ │ │ │ 49.0% │
	└──────────────────────┘ └──────────────────────┘ └──────────────────────┘
	```

	<br>

	---

	## 🧭 4. Use-Case Fit Matrix

	\| Use Case · 用途 \| Fit \| Notes \|
	\|:----------------\|:---:\|:------\|
	\| 🔍 Code Security Audit · 代码审计 \| ✅ \| Core strength — PrimeVul Recall 88.5% \|
	\| 🐛 Vulnerability Detection · 漏洞检测 \| ✅ \| High recall — errs on the side of caution \|
	\| 🧠 Structured Vuln Analysis · 结构化分析 \| ✅ \| Built-in `<think>` reasoning chains \|
	\| ⚔️ PenTest Knowledge · 渗透测试 \| ✅ \| MITRE ATT&CK · CVE · ExploitDB \|
	\| 📚 CTF Assistance · CTF 辅助 \| ✅ \| Practical security challenges \|
	\| 🏷️ CWE Classification · CWE 分类 \| ⚠️ \| Weak — pair with Semgrep / CodeQL \|
	\| 💻 General Code Generation · 代码生成 \| ⚠️ \| Use base Qwen-Coder instead \|
	\| 📐 Math / Physics · 数理推理 \| ⚠️ \| Expected trade-off \|
	\| 🏥 Medical / Chemical · 医疗化学 \| ❌ \| Out of training distribution \|

	<br>

	---

	## 📋 5. Model Specifications

	\| Property \| Value \|
	\|:---------\|:------\|
	\| Base Model \| [Qwen/Qwen2.5-Coder-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct) \|
	\| Architecture \| Qwen2ForCausalLM · 48 layers · 5120 hidden · 40 attn heads · 8 KV heads \|
	\| Parameters \| 14B (~16.8B total) \|
	\| Precision \| BF16 — single `model.safetensors` (~29 GB) \|
	\| Context Length \| 32,768 tokens \|
	\| Vocabulary \| 152,064 (ChatML template) \|
	\| Training \| SFT + GRPO (Group Relative Policy Optimization) \|
	\| Chat Template \| `<\\|im_start\\|>...<\\|im_end\\|>` + native `tool_calls` \|
	\| License \| [Apache 2.0](./LICENSE) \|

	<br>

	---

	## 📦 6. Usage

	### 🤗 Transformers

	```bash
	pip install transformers torch accelerate
	```

	```python
	from transformers import AutoModelForCausalLM, AutoTokenizer
	import torch

	model = AutoModelForCausalLM.from_pretrained(
	"Tidecaller/BountyHunter-RedTeam",
	torch_dtype=torch.bfloat16,
	device_map="auto",
	)
	tokenizer = AutoTokenizer.from_pretrained("Tidecaller/BountyHunter-RedTeam")

	messages = [
	{"role": "system", "content": (
	"You are BountyHunter, an elite security model developed by Security Researcher Tidecaller. "
	"Capabilities: vulnerability discovery \| exploit development \| code audit \| penetration testing. "
	"Principles: code over theory, evidence-based. "
	"Output: security tasks use <think> reasoning chain before results."
	)},
	{"role": "user", "content": "Audit this C code for vulnerabilities: ..."},
	]

	text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
	inputs = tokenizer(text, return_tensors="pt").to(model.device)
	outputs = model.generate(**inputs, max_new_tokens=4096, temperature=0.5, top_p=0.9)
	print(tokenizer.decode(outputs[0], skip_special_tokens=True))
	```

	### ⚡ vLLM

	```bash
	vllm serve Tidecaller/BountyHunter-RedTeam \
	--max-model-len 32768 \
	--tensor-parallel-size 1 \
	--dtype bfloat16
	```

	### 🚀 TGI (Text Generation Inference)

	```bash
	text-generation-server \
	--model-id Tidecaller/BountyHunter-RedTeam \
	--max-total-tokens 32768 \
	--dtype bfloat16
	```

	<br>

	---

	## 💾 7. Resource Estimation

	\| Precision \| VRAM \| Compatible Hardware \|
	\|:----------\|:----:\|:--------------------\|
	\| BF16 (this repo) \| ~29 GB \| A100 40GB · A800 · A6000 · 2× RTX 4090 \|
	\| Q8_0 GGUF \| ~15 GB \| RTX 4090 (24GB) · RTX 5090 \|
	\| Q4_K_M GGUF \| ~9 GB \| RTX 3080 (12GB) · Consumer GPUs \|

	<br>

	---

	## 🔬 8. Reasoning Chain Example

	```
	<think>
	1. Received C code snippet for security audit
	2. Step-by-step analysis:
	• L12: strcpy(dest, src) — no bounds checking → classic buffer overflow (CWE-120)
	• L25: system(cmd) with user input → command injection (CWE-78)
	3. Confidence: HIGH — all patterns confirmed with code-line evidence
	4. Providing remediation for each
	</think>

	## Security Audit Results

	🔴 Vulnerability 1: Buffer Overflow (CWE-120) — `main.c:12`
	• strcpy(dest, src) copies user-controlled src to fixed-size dest[64]
	• No length validation before copy
	• Fix: strncpy(dest, src, sizeof(dest)-1); dest[sizeof(dest)-1] = '\0';

	🔴 Vulnerability 2: Command Injection (CWE-78) — `main.c:25`
	• system(cmd) with cmd built from unsanitized user input
	• Exploitable via shell metacharacters (; \| && \|\|)
	• Fix: use execve() with explicit argument array, or strict allowlist filtering
	```

	<br>

	---

	## ⚠️ 9. Disclaimer & Ethics

	### 9.1 Legal Disclaimer · 法律免责声明

	> THIS MODEL IS A DUAL-USE SECURITY RESEARCH TOOL. Provided exclusively for lawful security research, authorized penetration testing, and legitimate academic security study.
	> 本模型为双用途安全研究工具，仅供合法的安全研究、授权渗透测试和正当学术安全研究使用。

	<br>

	Prohibited Uses · 禁止用途 (non-exhaustive)

	\| 禁止行为 \| Prohibited Conduct \|
	\|:---------\|:-------------------\|
	\| 未经授权访问任何计算机系统、网络或设备 \| Unauthorized access to any computer system, network, or device \|
	\| 开发、传播或部署恶意软件、勒索软件或病毒 \| Development / distribution / deployment of malware, ransomware, or viruses \|
	\| 未经授权的社会工程学攻击 \| Unauthorized social engineering attacks \|
	\| 未经授权的拒绝服务攻击 \| Unauthorized denial-of-service attacks \|
	\| 数据窃取或侵犯他人隐私 \| Data theft or violation of others' privacy \|
	\| 为实施犯罪目的绕过安全措施 \| Circumventing security measures for criminal purposes \|
	\| 违反任何适用法律法规 \| Violation of any applicable laws or regulations \|

	<br>

	No Warranty · 不提供担保 — incorporates Apache 2.0 § 8 by reference. Model provided "AS IS", without warranty of any kind. Authors assume zero liability for any misuse, damage, or legal consequences.

	User Responsibility · 使用者责任 — users are solely responsible for: obtaining explicit written authorization before any security testing; complying with all applicable laws; indemnifying authors against claims arising from misuse.

	<br>

	### 9.2 Ethical Statement · 伦理声明

	> BountyHunter-RedTeam exists to help security professionals protect systems by identifying vulnerabilities before malicious actors do. Its offensive capabilities serve defensive purposes.

	\| ✅ Permitted · 允许 \| ❌ Prohibited · 禁止 \|
	\|:--------------------\|:---------------------\|
	\| Authorized Penetration Testing \| Unauthorized System Intrusion \|
	\| Vulnerability Research & Responsible Disclosure \| Developing or Deploying Malware \|
	\| Code Security Auditing \| Cybercrime of Any Kind \|
	\| CTF Competitions & Security Exercises \| Academic Dishonesty \|
	\| Security Education & Training \| Privacy Violation / Surveillance \|
	\| Defensive Strategy & Threat Intelligence \| Unauthorized Production Exploitation \|
	\| Authorized Red Team Exercises \| Harassment / Defamation / Harm \|

	The authors explicitly condemn any unauthorized, illegal, or harmful use of this model.

	<br>

	### 9.3 Reporting Misuse · 举报滥用

	Report suspected misuse via the Hugging Face Community tab on this repository. We reserve the right to cooperate with law enforcement in relevant jurisdictions.

	<br>

	---

	## 🙏 10. Acknowledgments

	### Security & Vulnerability Datasets

	\| Dataset \| License \| Focus \|
	\|:--------\|:-------:\|:------\|
	\| [ayshajavd/code-security-vulnerability-dataset](https://huggingface.co/datasets/ayshajavd/code-security-vulnerability-dataset) \| Apache 2.0 \| Code vulnerability classification \|
	\| [CyberNative/Code_Vulnerability_Security_DPO](https://huggingface.co/datasets/CyberNative/Code_Vulnerability_Security_DPO) \| Apache 2.0 \| Vulnerability DPO pairs \|
	\| [Voidreaper2026/cybersec-master-dataset](https://huggingface.co/datasets/Voidreaper2026/cybersec-master-dataset) \| Apache 2.0 \| Cybersecurity knowledge synthesis \|
	\| [AYI-NEDJIMI/mitre-attack-en](https://huggingface.co/datasets/AYI-NEDJIMI/mitre-attack-en) \| Apache 2.0 \| MITRE ATT&CK framework \|
	\| [jason-oneal/mitre-stix-cve-exploitdb-dataset](https://huggingface.co/datasets/jason-oneal/mitre-stix-cve-exploitdb-dataset-alpaca-chatml-harmony) \| Apache 2.0 \| CVE + ExploitDB + MITRE \|
	\| [Waiper/ExploitDB_DataSet](https://huggingface.co/datasets/Waiper/ExploitDB_DataSet) \| MIT \| ExploitDB structured corpus \|
	\| [darkknight25/polyglot_paylods_datasets](https://huggingface.co/datasets/darkknight25/polyglot_paylods_datasets) \| MIT \| Polyglot XSS/SQLi payloads \|
	\| [SecureAI-SE/http-attack-requests](https://huggingface.co/datasets/SecureAI-SE/http-attack-requests) \| CC-BY 4.0 \| HTTP attack request corpus \|

	### General Instruction Datasets

	\| Dataset \| License \|
	\|:--------\|:-------:\|
	\| [QuixiAI/dolphin](https://huggingface.co/datasets/QuixiAI/dolphin) \| Apache 2.0 \|
	\| [m-a-p/Code-Feedback](https://huggingface.co/datasets/m-a-p/Code-Feedback) \| Apache 2.0 \|
	\| [NousResearch/hermes-function-calling-v1](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1) \| Apache 2.0 \|
	\| [glaiveai/glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2) \| Apache 2.0 \|
	\| [Team-ACE/ToolACE](https://huggingface.co/datasets/Team-ACE/ToolACE) \| Apache 2.0 \|
	\| [WizardLMTeam/WizardLM_evol_instruct_V2_196k](https://huggingface.co/datasets/WizardLMTeam/WizardLM_evol_instruct_V2_196k) \| MIT \|
	\| [HuggingFaceH4/ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) \| MIT \|
	\| [sahil2801/CodeAlpaca-20k](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k) \| CC-BY 4.0 \|
	\| [nvidia/Daring-Anteater](https://huggingface.co/datasets/nvidia/Daring-Anteater) \| CC-BY 4.0 \|

	### Base Model

	[Qwen/Qwen2.5-Coder-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct) by Alibaba Cloud.

	<br>

	---

	## 📜 11. License & Citation

	### License

	```
	BountyHunter-RedTeam — Fine-tuned weights
	Copyright © 2026 Tidecaller

	Based on Qwen2.5-Coder-14B-Instruct (Apache 2.0)
	Copyright © Alibaba Cloud

	Licensed under the Apache License, Version 2.0
	http://www.apache.org/licenses/LICENSE-2.0
	```

	See [LICENSE](./LICENSE) for full text.

	### Citation

	```bibtex
	@model{bountyhunter-redteam-2026,
	title = {{BountyHunter}: Elite Red Team Model based on Qwen2.5-Coder-14B},
	author = {Tidecaller},
	year = {2026},
	publisher = {Hugging Face},
	url = {https://huggingface.co/Tidecaller/BountyHunter-RedTeam}
	}
	```

	<br>

	<div align="center">

	---

	Code over theory. Evidence over speculation.
	代码优先于理论。证据优先于猜测。

	</div>