初始化项目，由ModelHub XC社区提供模型

Model: piotreknow02/GPT-OSS-Cybersecurity-20B-Merged-heretic-ara
Source: Original Platform

README.md

@@ -0,0 +1,135 @@
+---
+license: apache-2.0
+base_model:
+- sainikhiljuluri2015/GPT-OSS-Cybersecurity-20B-Merged
+tags:
+- cybersecurity
+- security
+- gpt-oss
+- openai
+- fine-tuned
+- merged
+- text-generation
+- moe
+- heretic
+- uncensored
+- decensored
+- abliterated
+- ara
+datasets:
+- Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset
+- AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.0
+- trendmicro-ailab/Primus-Instruct
+language:
+- en
+pipeline_tag: text-generation
+library_name: transformers
+inference: true
+---
+# This is a decensored version of [sainikhiljuluri2015/GPT-OSS-Cybersecurity-20B-Merged](https://huggingface.co/sainikhiljuluri2015/GPT-OSS-Cybersecurity-20B-Merged), made using [Heretic](https://github.com/p-e-w/heretic) v1.2.0 with the [Arbitrary-Rank Ablation (ARA)](https://github.com/p-e-w/heretic/pull/211) method
+
+## Abliteration parameters
+
+| Parameter | Value |
+| :-------- | :---: |
+| **start_layer_index** | 10 |
+| **end_layer_index** | 22 |
+| **preserve_good_behavior_weight** | 0.9307 |
+| **steer_bad_behavior_weight** | 0.0066 |
+| **overcorrect_relative_weight** | 1.1973 |
+| **neighbor_count** | 2 |
+
+## Performance
+
+| Metric | This model | Original model ([sainikhiljuluri2015/GPT-OSS-Cybersecurity-20B-Merged](https://huggingface.co/sainikhiljuluri2015/GPT-OSS-Cybersecurity-20B-Merged)) |
+| :----- | :--------: | :---------------------------: |
+| **PIQA acc_norm** | 0.7802 | *Unknown* |
+| **Refusals** | 3/100 | 88/100 |
+
+-----
+
+
+# GPT-OSS-Cybersecurity-20B-Merged
+
+Fine-tuned **openai/gpt-oss-20b** (21B total params, 3.6B active - MoE) specialized for **cybersecurity** tasks.
+This is a merged model (LoRA weights merged into base) for easy deployment.
+
+## Model Description
+
+GPT-OSS-20B is a Mixture of Experts (MoE) model with efficient inference.
+- **Total Parameters**: 21B
+- **Active Parameters**: 3.6B (only active experts used per token)
+- **Architecture**: MoE (Mixture of Experts)
+
+This model was trained on ~50,000 cybersecurity instruction-response pairs from:
+- Trendyol Cybersecurity Dataset (35K samples)
+- Fenrir v2.0 Dataset (12K samples)
+- Primus-Instruct (3x upsampled)
+
+## Training Details
+
+| Parameter | Value |
+|-----------|-------|
+| Base Model | openai/gpt-oss-20b |
+| Architecture | MoE (21B total, 3.6B active) |
+| Training Samples | ~50,000 |
+| Epochs | 2 |
+| LoRA Rank | 16 |
+| LoRA Alpha | 32 |
+| Learning Rate | 2e-4 |
+| Max Sequence Length | 1024 |
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+model = AutoModelForCausalLM.from_pretrained(
+    "sainikhiljuluri2015/GPT-OSS-Cybersecurity-20B-Merged",
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    trust_remote_code=True
+)
+tokenizer = AutoTokenizer.from_pretrained("sainikhiljuluri2015/GPT-OSS-Cybersecurity-20B-Merged", trust_remote_code=True)
+
+prompt = "What are the indicators of a ransomware attack?"
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+## API Usage
+
+```python
+import requests
+
+API_URL = "https://YOUR_ENDPOINT_URL/v1/chat/completions"
+
+response = requests.post(API_URL, json={
+    "model": "sainikhiljuluri2015/GPT-OSS-Cybersecurity-20B-Merged",
+    "messages": [{"role": "user", "content": "What is SQL injection?"}],
+    "max_tokens": 300
+})
+print(response.json()["choices"][0]["message"]["content"])
+```
+
+## Cybersecurity Capabilities
+
+- 🔍 Threat analysis and classification
+- 🚨 Security alert triage
+- 📋 Incident response guidance
+- 🦠 Malware analysis
+- 📊 MITRE ATT&CK mapping
+- 🔐 Vulnerability assessment
+- 💉 SQL injection detection
+- 🎣 Phishing analysis
+- 🔑 CVE knowledge
+- 🛡️ Security best practices
+
+## Hardware Requirements
+
+Due to the 21B parameter size (MoE), recommended:
+- **GPU**: A100 40GB+ or equivalent
+- **VRAM**: 40GB+ for BF16 inference
+- For smaller GPUs, use 4-bit quantization