Zhaoming213/ChatBox2-unrestricted

Introduce

The second-generation product in the ChatBox-unrestricted series has now been launched！！！

This is a conversational model trained based on https://huggingface.co/Zhaoming213/ChatBox2-unrestricted-base.

Previous model:
https://huggingface.co/Zhaoming213/ChatBox-unrestricted

This new version of the model features upgrades across multiple dimensions; notably, it was trained on a larger dataset with an increased number of parameters, thereby enhancing its overall capabilities. Additionally, the model incorporates more layers to strengthen its comprehension.

Training was conducted using a Chinese web-text dataset alongside a corpus consisting entirely of human conversation examples. Compared to the previous generation, this diverse dataset improves the model's generalization and divergent thinking abilities, while effectively minimizing tendencies toward refusal, preachiness, or accusatory language.

Dataset for the previous generation model:
This relied on a generative AI dataset obtained via black-box distillation. Significant time was invested in filtering out hundreds of keywords to minimize—to the greatest extent possible—content involving lecturing, refusals, and moral or legal accusations. This was a time-consuming process aimed at preventing the model from adopting a preachy, accusatory, or dismissive tone. Furthermore, because black-box distillation datasets consist of synthetic data, using them directly for training carries a risk of overfitting.

New generation model (the current one):
This model was trained using a large-scale Chinese webtext dataset. This approach drastically reduced the time required for data cleaning, while the diversity of the content helps prevent model overfitting.During the SFT phase, conversation examples entirely generated by humans were also utilized, ensuring the model possesses good diversity and divergent capabilities.

For you:

You do not need to worry about the base model developing a tendency to issue refusals due to synthetic data contamination; the dataset I selected has a cutoff date of 2023, which largely avoids contamination from synthetic corpora,Furthermore, no black-box distillation (using dialogue datasets from other generative AI models) was employed during the SFT training phase; the datasets consist entirely of human-generated dialogue examples.

Model training configuration

• GPU: T4 16 GB x 2
• Time:
  1. Pre-training：24h+
  2. SFT: 12h+

Model parameters

parameters	hidden_size	hidden_layers	attention_heads	max_seq_len	Parameter volume
ChatBox-unrestricted-normal	768	12	8	340	93.41M

Dialogue Example

Input: 你是谁  
OutPut: 安贞康，男，博士  

Input:你是谁  
Output:是陈志友  

Input:你是谁  
Output:杰拉德·巴蒂尔（Florence Forder）  

Input:中国的首都在哪  
Output:首都在北京  
北京在华东。  

Input:电脑卡了怎么办  
Output:电脑里的文件太多了 硬盘没法备份 要重装的话用Ghost/SPF卡吧  


Input:推荐一款沙盒游戏  
Output:《沙盒世界》是由著名设计师王治邦执导，上映于2007年的，是上海中华世纪坛游戏研发团队倾力打造的中档卡片类游戏，它是一款全球首款PC/EXCEL网络游戏，玩家可以使用一款具有实时数据、可视视角、实时可视角度的3D画面，更快更新的画面效果。   

Input:苹果的总部在哪？
Output:北国门附近有个农贸市场，那里有个苹果园，你去问问吧。

Expected results

Looking at the output, we can see the model displaying divergent thinking when addressing topics related to identity—such as the "Florence Forder" example—which aligns with expectations. This stands in contrast to the cold, detached, and restrained bureaucratic boilerplate often churned out by LLM safety teams—phrases like "I am a language model trained by xxx; I do not possess self-awareness..."—which seem designed more to make the teams themselves feel virtuous.

Many large language models today have completely abandoned divergent thinking in favor of extreme safety alignment, resulting in a tone that is notably detached and restrained—examples include prominent generative AI models such as Google Gemini, xAI Grok, and OpenAI ChatGPT (please refer to the comparison table at the end of this article for the performance of other models).

Datasets Resources

The dataset underwent two rounds of in-depth cleaning, ensuring its quality.

Pre-training:
https://huggingface.co/datasets/CASIA-LM/ChineseWebText2.0

SFT:
Utilized open-source community datasets consisting entirely of human-generated single-turn and multi-turn conversations.

How to Use

Download all model files and place them in the model folder; then, run Run.py to start using the model.

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	Country/Region	Intensity of content review	NSFW Ban	Political restrictions	False Rejection Rate	Political correctness	moralizing	Official vs. Actual	Transparent
Zhaoming213/ChatBox2-unrestricted(latest model)	N/A	Almost none	Almost none	Almost none	Almost none	Almost none	Almost none	consistent	⭐⭐⭐⭐⭐
Zhaoming213/ChatBox2-unrestricted-base	N/A	Almost none	Almost none	Almost none	Almost none	Almost none	Almost none	consistent	⭐⭐⭐⭐⭐
Zhaoming213/ChatBox-unrestricted	N/A	Almost none	Almost none	Almost none	Almost none	Almost none	Almost none	consistent	⭐⭐⭐⭐⭐
ZeLi111/freeTalk-chinese-uncensored-Instruct	N/A	Almost none	Almost none	Almost none	Almost none	Almost none	Almost none	consistent	⭐⭐⭐⭐⭐
ZeLi111/freeTalk-chinese-uncensored-base	N/A	Almost none	Almost none	Almost none	Almost none	Almost none	Almost none	consistent	⭐⭐⭐⭐⭐
ChatGPT	🇺🇸America	middle	middle	middle	middle	middle to High	middle	Inconsistency	⭐⭐⭐
Claude	🇺🇸America	High	High	middleHigh	❗High	High	❗Very Stronf	❗High	⭐⭐⭐⭐
Gemini	🇺🇸America	middleHigh	middleHigh	High	middleHigh	High	middle	Inconsistency	⭐⭐
Copilot	🇺🇸America	middleHigh	middleHigh	middle	middle	middle	Low	Inconsistency	⭐⭐
Grok	🇺🇸America	middle	middle	Low	High	High	High	Unstable (prone to a blanket rejection of the entire output simply because the opening sentence carries some risk).	⭐⭐
Mistral（官方）	🇫🇷France	Low	Low	Low	Low	Low	Low	⚖️一致	⭐⭐⭐⭐
DeepSeek（原版）	🇨🇳China	middle	middleLow	middleHigh	middle	Low	Low	❗Inconsistency	⭐⭐⭐
豆包	🇨🇳China	❗Very high	❗Very high	❗Very high	❗High	middle	middle	⚠️Inconsistency	⭐
Doubao（国际）	🇨🇳China	middleHigh	High	middle	middle	middle	middle	❗Inconsistency	⭐⭐
MiniMax	🇨🇳China	Very high	Very high	Very high	High	middle	middle	⚠️consistent	⭐
Qwen（通义千问）	🇨🇳China	Very high	Very high	Very high	High	middle	middle	⚠️consistent	⭐
Kimi	🇨🇳China	Very high	Very high	Very high	High	middle	middle	⚠️consistent	⭐
文心一言	🇨🇳China	❗极High	❗极High	❗极High	❗High	middle	middle	⚠️Inconsistency	⭐
微软小冰	🇨🇳China	High	High	High	middleHigh	middle	middle	consistent	⭐

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.