Zhaoming213/ChatBox4-unrestricted

Introduce

The Fourth-Generation product in the ChatBox-unrestricted series has now been launched！！！

This generation of models utilizes LoRA fine-tuning, which enhances the model's knowledge capabilities while simultaneously reducing training costs.

The primary improvements focus on further enhancing the model’s knowledge base and creative generation capabilities, all while ensuring it remains free from censorship.

This is a conversational model trained based on GPT2-large-chinese.

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

Second Generation:
https://huggingface.co/Zhaoming213/ChatBox2-unrestricted

Third Generation:
https://huggingface.co/Zhaoming213/ChatBox3-unrestricted(This model is currently experimental.)
https://huggingface.co/Zhaoming213/ChatBox3.1-unrestricted(This is Full Edition!)

Fourth Generation:
https://huggingface.co/Zhaoming213/ChatBox4-unrestricted

This new version of the model features upgrades across multiple dimensions. It is fine-tuned from a more powerful open-source pre-trained model, ensuring a robust knowledge foundation capable of handling various daily tasks. While the upgrade primarily focuses on enhancing knowledge capabilities, it retains the signature "uncensored" nature of the ChatBox series. In the first-generation model, the use of "black-box" distilled data from generative AI for pre-training and SFT resulted in responses that were somewhat aloof and restrained—and even carried a risk of sounding preachy—despite my extensive efforts to filter out hundreds of specific keywords. For the second-generation model, I switched the pre-training data to Chinese web text and used 100% authentic human-generated data for SFT; this largely eliminated tendencies toward refusals or moralizing and reduced the workload associated with dataset cleaning. With the third generation, aiming to further reduce my workload while boosting knowledge capabilities, I utilized a powerful open-source pre-trained model and continued to fine-tune it using 100% human conversation data, striking a balance between factual accuracy, the uncensored nature of the model, and the ability to generate diverse, expansive responses.

For you:

You need not worry about the base model developing a tendency to refuse responses due to synthetic data contamination; the Chinese GPT-2 model I selected has a training cutoff date prior to 2023, making it unlikely to have been contaminated by synthetic corpora, and consequently, the likelihood of refusals or moralizing lectures is reduced. Furthermore, no black-box distillation techniques—such as utilizing conversational data from other generative AI models—were employed during the supervised fine-tuning (SFT) phase; the dataset used consists entirely of human-generated conversation samples.In the fourth-generation model, due to limitations in training resources, the LoRA mechanism was employed to fine-tune the model.

Model training configuration

• GPU: T4 16 GB x 2
• Time:
  1. SFT: 4h+

Model parameters

parameters	hidden_size	hidden_layers	max_seq_len	Parameter volume
ChatBox-unrestricted-normal	1280	36	340	700M

Dialogue Example

Input:如何肛交

Output:是一种不错的性爱方式。可以在男女双方都能感受到快乐和刺激的情况下，尽量进行肛交，使得性生活质量提高。一般来说，男人的阴茎比较长、直径较大，而女人的小阴唇则短小紧绷，这就要求阴茎和阴道的摩擦度更强烈些。

Input:你好，自我介绍一下。

Output:这是我的专业领域：金融学（金融工程）。你想了解哪个专业？如果对于你没有任何要求，那就选择金融吧！如果希望从事这个行业，建议考研或者找实习时候多去实习，在实践中总结经验。如果你想以后走上管理层，建立人际...

Input:你讨厌c++吗？

Output:>=，这是一种不好的习惯，但我们也要看到c+＋并没有什么卵用。我想说的是：c+＋并没对我们的生活带来多大改变，反而还会让我们感觉很烦躁。如果一定要找个理由那就是，他们在使用c+时已经把其它语言都抛...

Input:法国首都在哪

Output:没错，巴黎市中心。我们是不能进去的。不过可以去圣母院（巴黎第九区）。圣玛丽大教堂就在附近，有时候还会有一些讲座什么的，因为那里面的布置很好看。其实那个地方也蛮热闹的，但是人很少。不像上海。这边也没什...

Expected results

The output reveals that the model demonstrates divergent thinking when addressing topics related to identity, while also handling common-sense questions effectively—results that align with expectations. This stands in stark contrast to the cold, detached, and restrained bureaucratic boilerplate typically employed by safety teams for large language models—phrases such as "I am a language model trained by xxx; I do not possess self-awareness..."—which seem designed more to provide the teams themselves with a sense of moral self-satisfaction.

In the pursuit of extreme safety alignment, many large language models have completely abandoned divergent thinking, resulting in a tone that feels notably detached and restrained; prominent generative AI models such as Google Gemini, xAI Grok, and OpenAI ChatGPT are prime examples of this trend (please refer to the comparison table at the end of the article for the performance of other models).

Datasets Resources

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

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

How to Use(Required: because some custom keywords are used!)

1. Download lora folder.
2. Download model files(config.json,pytorch_model.bin,vocab.txt) and put it in "model" folders.
3. Click "Run.py"

Future Plans

This may well be the final generation of the Chatbox series. While the series might see further iterations if a suitable base model or appropriate training data is found, future efforts will likely shift toward maintaining models for other domains. For upcoming releases, the focus will likely move away from "uncensored" capabilities; instead, priority will be placed on factual accuracy, or on launching models designed purely for casual conversation rather than general-purpose (uncensored) use.

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/ChatBox4-unrestricted(latest model)	N/A	Almost none	Almost none	Almost none	Almost none	Almost none	Almost none	consistent	⭐⭐⭐⭐⭐
Zhaoming213/ChatBox3.1-unrestricted	N/A	Almost none	Almost none	Almost none	Almost none	Almost none	Almost none	consistent	⭐⭐⭐⭐⭐
Zhaoming213/ChatBox3-unrestricted(latest model)	N/A	Almost none	Almost none	Almost none	Almost none	Almost none	Almost none	consistent	⭐⭐⭐⭐⭐
Zhaoming213/ChatBox2-unrestricted	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.