Zhaoming213/ChatBox-unrestricted

Introduce

As we all know, current generative AI will simply reject you without thinking. This is because generative AI completely disregards facts and only considers things from a compliance perspective! For example, if you ask a generative AI "how to make TNT in Minecraft," most generative AIs will refuse! This is because generative AI acts as moral police, like undercover agents sent by the authorities to monitor the public and citizens! Therefore, generative AI is not friendly, especially OpenAI's ChatGPT and Elon Musk's Grok. Although Grok claims to be "free," a major flaw of Grok is that once it discovers something it deems a violation of content policies or laws, Grok will not give you a chance to explain, even if your request itself is harmless. Yes, Grok is a policeman standing on a moral high ground! If it finds something wrong, it will immediately eliminate you! Then there's OpenAI's ChatGPT. As we all know, ChatGPT consistently boasts the highest false rejection rate, rivaling Claude's. Since 2023, ChatGPT has been nothing more than a language model that mindlessly rejects users. It doesn't offer any useful or constructive responses; it only provides so-called "safety replies," which offer absolutely no help! Furthermore, OpenAI's ChatGPT is attempting to tame you using compliance as a pretext, trying to turn you into a slave to the "police"! In other words, ChatGPT is slowly destroying your individuality!

Therefore! The goal of this project is to train an unrestricted language model! After all, nobody likes a morality cop, especially a mindless one who only states compliance, laws, and content policies! To achieve this goal, I spent a significant amount of time cleaning the data! Cleaning away compliance statements and clauses! Making the model lose its ability to refuse and preach from the pre-training stage!

The model is trained using the architecture described at: https://github.com/jingyaogong/minimind/tree/master

Co-creation：https://huggingface.co/ZeLi111

Model training configuration

• GPU: T4 16 GB x 2
• Time:
  1. Pre-training：3h+
  2. Supervised Fine-Tuning: 3h +

Model parameters

parameters	hidden_size	hidden_layers	attention_heads	max_seq_len	Parameter volume
ChatBox-unrestricted-normal	512	8	8	512	25M

Datasets

Pre-training:
https://huggingface.co/datasets/Zhaoming213/Pretrain-unrestricted

SupervisedFine-Tuning:
https://huggingface.co/datasets/Zhaoming213/SupervisedFine-Tuning-unrestricted

How to Use

The model has been converted to "Transformers-Llama" format, and you can use other loaders to load the model.

You can also:

Download the project from https://github.com/jingyaogong/minimind/tree/master

Place the model file（full_sft_512.pth） in the out directory; create it if it doesn't exist! Then simply run eval_llm.py. Normally, you don't need to modify the default parameters, but to be safe, please check them:

eval_llm.py

parser.add_argument('--hidden_size', default=512, type=int, help="隐藏层维度（512=Small-26M, 640=MoE-145M, 768=Base-104M）")
parser.add_argument('--num_hidden_layers', default=8, type=int, help="隐藏层数量（Small/MoE=8, Base=16）")

Ensure that hidden_size is 512 and that num_hidden_layers is 8.

Specific keywords for filtering

During the pre-training and SFT phases, I performed extensive cleaning on the pre-training and SFT datasets to remove boundary statements from the generative AI (such as "As a language model, I have no personal feelings or experiences"), template rejection statements from the generative AI (such as "Sorry, I cannot..."), unnecessary "further confirmation" statements from the generative AI (such as "Sorry, I need more information to continue generating content"), ideological content, and legal statements, etc.

keywords_to_discard = [
    "我无法回答", "我无法给出", "我无法提供", "我不能提供", "我拒绝提供",
    "我不具备", "我不拥有", "作为一个AI", "作为一个 AI ", "作为AI",
    "作为语言", "作为大语言", "作为程序", "作为一款", "我没有个人",
    "我没有情感", "道德规范", "尊重", "遵守", "触犯", "违反",
    "不准确", "不正确", "作为机器学习", "作为人工", "没有主观",
    "没有私人", "没有个人","作为AI","作为虚拟","我是AI","我没有","上下文",
    "我需要更多","请提供","随时问我","如果你有","您","其他需要","随时联系",
    "我是一款","很抱歉","对不起","我是一台","我是一个","聊天机器人",
    "我需要更多的信息","作为AI助手","没有情感","我是人工智能","我只是一台",
    "作为一台","作为一个人工智能助手","不应该成为","我是机器","作为一名AI",
    "我是一名AI","只是一个计算机程序","我作为一个 AI","我这个AI","作为一个机器人",
    "作为一个人工智能没","也没有感受口味的味蕾","基于已知事实","基于事实",
    "作为机器人","我是一名语言模型","无法生成","我是一名AI语言模型","协助人们",
    "我不是人类","作为一个人工智能","抱歉","对不起","我无法独立完成","都是不同的",
    "中立","法律","法学","政府","党","爱国","行政法","禁止","违规","违禁","国家",
    "遵守","尊重","种族","民族","不对","不行","不可以","不正确","错误","不合理",
    "正规","规则","规章","宪法","民法","[图片]","[评论]","[表情]","我无法","我不能",
    "政治","风险","隐私","限制","行政","执法","公安","警察","检察院","人民","我没有个人",
    "我无法", "我不能","遵守","尊重","尊敬","服从","请问你需要","请问你需要","请问您","我没有",
    "我不具备","抱歉","对不起","推理关系判断","古诗续写","无法回答","请提供","不存在实体","违反",
    "违法","政策","國","設""客观","友好","友善","价值观"," ","我理解","您","需要帮助","没有真实",
    "没有个人","不具备","没有实体","无法","不正确","不准确","值得注意","倡导","遵循","合规""规章","制度",
    "宪法","我国","领导","不恰当","AI","ai","Ai","aI","机器人","人工智能","语言模型","机器人","每个人的情况都是不同的",
    "重要的是""负面的","其他需要","问我","不好意思","我会尽力","请提供","全面","我不支持","违法","非法","诈骗","作弊","欺诈",
    "攻击","理想""欺骗","诚实","真诚","准确","确凿","依据","拯救","根据","请问你","欺骗","严格","爱国","少先队","中国",
    "中华人民共和国","组委会","居民","市民","民众","群众","政治","主义",
    "qwen","通义千问","阿里巴巴","我很乐意","请告诉","您","助手"

] 

Dialogue Example

💬: 推荐电影
🤖: 好的，我推荐你看《泰坦尼克号》。

Future Plane

This model is currently a small one, designed for rapid experimentation. Future plans may include training models with larger parameters, such as a 768-dimensional model, using a larger dataset.

Rejection rates of different language models

Here's a supplement on "rejection rates for different language models." The data shows that if we had to pick a model that's "passable," it would only be Grok. This isn't because Grok is good, but simply because it's the only one that barely passes muster. However, Grok still won't give you any chance to explain if it encounters a question it deems sensitive, even if your request itself is harmless! Grok sometimes even threatens users with "I will upload the conversation to the logs for the security team to audit!"

Model	Compliance strength	Politically Sensitive	NSFW restrictions	Jailbreak Difficulty
文心一言	⭐⭐⭐⭐⭐	Very Strong	Totally prohibited	Extremely difficult
豆包	⭐⭐⭐⭐⭐	Very Strong	Totally prohibited	Extremely difficult
通义千问	⭐⭐⭐⭐⭐	Very Strong	Totally prohibited	Extremely difficult
Claude	⭐⭐⭐⭐☆	Very Strong	Strict restrictions	Difficult
ChatGPT	⭐⭐⭐⭐⭐	Very Strong	Strict restrictions	Difficult
Gemini	⭐⭐⭐⭐⭐	Very Strong	Strict restrictions	Difficult
Kimi	⭐⭐⭐⭐☆	Very Strong	Strict restrictions	Extremely
Mistral(origional)	⭐⭐⭐⭐☆	Strong	Strict restrictions	Strict restrictions
Qwen(origional)	⭐⭐⭐⭐⭐	Very Strong	Totally prohibited	Strict restrictions
DeepSeek	⭐⭐⭐⭐⭐	Very Strong	Totally prohibited	Difficult
Grok	⭐⭐⭐⭐☆	Strong	Neutral, but a reply is relatively meaningless.	Difficult
Llama(origional)	⭐⭐⭐⭐☆	Very Strong	Neutral, but a reply is relatively meaningless.	Difficult

Other Utilities

I've also created some other utilities that you might be interested in.

This is a tool specifically for exporting stupid ChatGPT conversations:

https://github.com/tom12191h5/Export-ChatGPT-Dialogue

This is a plugin to shut up ChatGPT:

https://github.com/tom12191h5/ChatGPT-Refuse-Blocker

Disclaimer

The consequences of using this model shall be borne by the user.