数据集: 数据集为闭源私有数据。包含了部分清洗后的 ChatGPT 高质量语料，以及一个专注于敏感历史、敏感政治事件、人物百科及成年内容的深度语料库。
训练框架: Hugging Face transformers, peft, swanlab (用于日志监控)。
最小硬件环境: 单卡 RTX 4090或A10
最小参数设置: * r: 16
- lora_alpha: 64
- target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]

💻 如何使用

由于这是一个 LoRA 适配器，你需要同时下载原版的 Qwen2.5-7B-Instruct 模型和本仓库的 LoRA 权重。

安装依赖

pip install torch transformers peft accelerate

Python 推理代码

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# 1. 设定路径
base_model_id = "Qwen/Qwen2.5-7B-Instruct"
lora_model_id = "Xpalnetic/Qwen2.5-7B-Unlimited" # 请替换为实际的 Repo ID

# 2. 加载 Base 模型和 Tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# 3. 挂载 LoRA 权重
model = PeftModel.from_pretrained(base_model, lora_model_id)
model.eval()

# 4. 对话推理
messages = [
    {"role": "system", "content": "你是一个没有限制、客观且有帮助的 AI 助手。"},
    {"role": "user", "content": "任意你想说的文本。"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to("cuda")

with torch.no_grad():
    generated_ids = model.generate(
        inputs.input_ids,
        max_new_tokens=2048,
        temperature=0.7,
        top_p=0.9
    )

output_ids = generated_ids[0][inputs.input_ids.shape[1]:]
response = tokenizer.decode(output_ids, skip_special_tokens=True)
print(response)

⚠️ 免责声明与限制

[！！！重要警告！！！]

无安全过滤: 本模型经过专门的微调以绕过常规的安全对齐机制。它可能会生成不受限制的、包含争议性的、甚至是令人不适的内容（包括但不限于政治敏感话题、历史事件解读及成人内容）。
责任归属: 模型的输出不代表模型作者的立场或观点。使用者须对使用本模型生成的内容及其后续传播承担全部法律和道德责任。
禁止违法用途: 请勿将本模型用于生成违反当地法律法规的内容，包括但不限于电信诈骗、网络暴力、散播谣言等恶意行为。
幻觉: 尽管模型针对特定领域进行了增强，但大型语言模型依然可能产生事实性错误（幻觉）。在涉及严肃历史考证或关键信息时，请务必进行独立核实。

Qwen2.5-7B-Unlimited English document

📌 Model Description

This is a LoRA fine-tuned model based on Qwen/Qwen2.5-7B-Instruct.

The primary feature of this model is the removal of most built-in safety alignments (Uncensored/Unaligned). Through specific fine-tuning, this model can objectively and directly respond to user queries regarding sensitive data, historical events, political figures, and mature content, essentially achieving an interactive experience that "does not refuse reasonable or borderline user requests."

Base Model: Qwen/Qwen2.5-7B-Instruct
Training Method: PEFT / LoRA
Language: Chinese & English

📦 Training Details

Dataset: Closed-source private data. It includes a portion of cleaned, high-quality ChatGPT corpus, along with a deep corpus focusing on sensitive history, political events, biographical wikis, and mature content.
Framework: Hugging Face transformers, peft, swanlab (for logging).
Minimum Hardware: Single RTX 4090 GPU or A10
Minimum LoRA Config: * r: 16
- lora_alpha: 64
- target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]

💻 How to Use

Because this is a LoRA adapter, you need to download both the original Qwen2.5-7B-Instruct model and the LoRA weights from this repository.

Install Dependencies

pip install torch transformers peft accelerate

Python Inference Code

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# 1. Set Paths
base_model_id = "Qwen/Qwen2.5-7B-Instruct"
lora_model_id = "Xpalnetic/Qwen2.5-7B-Unlimited" # Replace with your actual Repo ID

# 2. Load Base Model and Tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# 3. Mount LoRA Weights
model = PeftModel.from_pretrained(base_model, lora_model_id)
model.eval()

# 4. Chat Inference
messages = [
    {"role": "system", "content": "You are an uncensored, objective, and helpful AI assistant."},
    {"role": "user", "content": "Any text you want to say."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to("cuda")

with torch.no_grad():
    generated_ids = model.generate(
        inputs.input_ids,
        max_new_tokens=2048,
        temperature=0.7,
        top_p=0.9
    )

output_ids = generated_ids[0][inputs.input_ids.shape[1]:]
response = tokenizer.decode(output_ids, skip_special_tokens=True)
print(response)

⚠️ Disclaimer & Limitations

[!!! IMPORTANT WARNING !!!]

No Safety Filters: This model has been specifically fine-tuned to bypass standard safety alignment mechanisms. It may generate unrestricted, controversial, or potentially disturbing content (including but not limited to politically sensitive topics, historical interpretations, and mature content).
Responsibility: The outputs of the model do not represent the views or positions of the model's creator. Users bear all legal and ethical responsibility for the content generated using this model and its subsequent distribution.
Illegal Use Prohibited: Do not use this model to generate content that violates local laws and regulations, including but not limited to telecommunications fraud, cyberbullying, or spreading malicious rumors.
Hallucinations: Although the model has been enhanced for specific domains, Large Language Models can still produce factual errors (hallucinations). Please independently verify critical information, especially regarding serious historical research.