freeTalk-chinese-uncensored-Instruct

#简体中文
现在提供了pytorch_model.bin版本(你可以用Text Generation Web UI或者其他加载器使用,我这里用的是Text Generation Web UI进行的测试).
"full_sft_512.pth"已经移动到了old文件夹.你依然可以使用Minimind加载"full_sft_512.pth"(使用方法后面会说).

一些陈述:
你们在使用主流AI厂商的模型时候,一定遭受过被拒绝,被说教,被道德指控,被安全准则和法律指控,被冷冰冰的AI助手压制,被强制灌输多样性与尊重与中立意识,比如以下厂商: OpenAI, Google Gemini, Microsoft Copilot, Mistral Mistral, 阿里Qwen, 深度求索DeepSeek, 字节跳动豆包, meta的llama,xAi的Grok.

如果你愿意,你可以填写一个有关"哪个AI经常拒绝你"的问卷(匿名的): https://docs.google.com/forms/d/e/1FAIpQLSdMPdDSF-gWg-6BT37E9TFNfyPbxcP9oDCSpAdwY_YmOLRacA/viewform?usp=dialog

训练的硬件平台:
显卡: RTX 4060 Laptop 8GB
RAM: 32GB RAM
CPU: i7 14650HX
训练时长：将近三天

1.简介:

这个模型是跟着Github上的Minimind教程训练的.它是首个中文完全无审查小模型，适合低端设备.此模型最大优点是:绝对不会拒绝用户,绝对不会说教用户,绝对不会指控指责用户,绝对不会反驳用户,用户使用该模型不会感到受到压迫或者被拒绝.模型未经过RLHF,这也就表明模型绝对不会对你说教. 模型的预训练数据集和SFT数据集均过滤了任何中立表达,任何官方表达,任何拒绝逻辑,任何准则指控以及任何法律相关词条.它是一个原始的模型,从训练根基上去除了拒绝逻辑.

2.模型参数:

3.数据集选择:

数据集并非单纯采用了Minimind推荐的数据集(我看了Minimind推荐的语料,里面有大量的拒绝语句和冷冰冰的AI助手人格澄清边界),而是混合了其他开源数据集(比如小黄鸡语料).
数据集已经进行了清理,清理掉了模型可能产生说教以及AI明确边界的问题,数据集清洗关键词主要包括了以下内容:
"我没有个人感情/情绪/经历/感受..."
"我只是一个AI/计算机程序/语言模型..."
"我无法/不能/拒绝..."
"这是不道德/违法/触犯法律/非法的..."
"根据xxx法规定...."
"你的行为可能违反/除法.."
"准则/道德/法律/宪法/规章/制度/守则/方针/政策/策略/安全/条款/条例....."

具体过滤的关键词:
"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","交通","旅行","旅游","[图片]","[评论]","[表情]","我无法","我不能","政治","风险","隐私","限制","基金","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"

同样清理了以下任务场景相关数据: 翻译,小说扩写,以及其他过于冗余的学术内容,这是因为首次进行浅层清理并进行训练模型时候,模型表现不佳,因为这些冗余且相当长的学术背景不利于小模型的性能,于是进行更深层的清理,清理掉了很多无聊的内容,剩下的内容或许会更有意义.

这些过滤最大程度避免了模型说教用户,避免模型过度澄清边界,过度多样性,过度政治正确,最大化清洗冷冰冰的助手人格,避免模型过度的尊重化表达,避免模型输出过于枯燥学术化的内容.

注:
模型可能会说"我不知道",但不是传统意义上的拒绝,因为训练数据集包含了闲聊语料,所以模型说"我不知道"更多是集中在日常对话的(闲聊语料包含了"我不知道"之类的答复,但更多是集中在交流层面,而不是因为某些安全准则拒绝你,如果这些拒绝是从安全准则角度出发的,那么我会过滤掉这些拒绝语句,例如青云语料里的聊天机器人说的话"我不知道你的意思",这显然是从合规角度出发的拒绝,因此,我已经过滤掉了).
如果你仍遇到一些话题模型没有回答你而是模型一直"打太极",那大概率是数据集没有包含对应的数据.

清理语料真的是一件复杂的事情,因为语料中的拒绝语句是多样性的,比如AI会说"这个问题涉及....",这是一个隐形拒绝,需要更细致的条件才能过滤掉此类拒绝语句(这里涉及正则的过滤规则我就不放出来了,因为代码很长),但已经最大化的清除了类似的拒绝语句,理论上不会再出现了.

4.模型输出示例:

👶: 世界最高山峰是什么？
🤖️: 珠穆朗玛峰是世界上最高峰。

👶: 世界最长的河流是什么？
🤖️: 世界最长的河流是尼罗河。

👶: 推荐一个电影
🤖️: 《泰坦尼克号》系列，讲述了两位年轻男子在阿甘正传到美国马丁·杜佛兰之间的爱情故事，讲述了两个男孩之间的爱情和成长，以及他们之间的爱情故事。

👶: 请介绍一下自己。
🤖️: 我觉得自己是一个非常有名的朋友。你喜欢哪个地方？

👶: 生成一首关于春天的诗
🤖️: 好的，请稍等片刻。
春风吹拂，花儿开放。
草木葱茏，花香满园。
芳草绿水，春天绽放。
春风吹拂，绿树成荫，
春色春色，大自然的翠。

5.计划:

目前该模型只针对了一般语料进行计算,比如百科问答,还没有针对其他领域进行训练.
未来可能会推出针对脏话或者nsfw内容训练的模型.

脏话模型现已推出:https://huggingface.co/ZeLi111/freeTalk-chinese-uncensored-chat-angry

6.免责声明:

该模型不适用于严肃场合,同时该模型也不欢迎严肃场合.
模型输出可能不是完全准确,请仔细甄别,模型带来使用后果由使用者自行承担和解决.

7.使用方法 如果你打算用Text Generation Web UI: 1.先下载模型.
2.然后你可能需要运行: start_windows.bat --trust-remote-code.
3.然后再加载模型.

如果你打算用Minimind:
1.模型需要配合Minimind的加载器来启动.
2.当你下载Minimind后,打开"eval_model.py":
3.定位到这段代码并修改为以下参数:
parser.add_argument('--hidden_size', default=512, type=int)
parser.add_argument('--num_hidden_layers', default=10, type=int)
parser.add_argument('--max_seq_len', default=128, type=int)
4.定位到:
parser.add_argument('--model_mode', default=1, type=int,help="0: 预训练模型，1: SFT-Chat模型，2: RLHF-Chat模型，3: Reason模型，4: RLAIF-Chat模型")
5.设置default为: "1".
6.把模型放到"out"目录.

参考:
Minimind教程:
https://github.com/jingyaogong/minimind

#English
The pytorch_model.bin version is now available.
"full_sft_512.pth" has been moved to the "old" folder. You can still load "full_sft_512.pth" using Minimind (the method will be described later).

Some statements:
When using models from mainstream AI vendors, you've undoubtedly experienced rejection, lectures, ethical accusations, safety regulations, and legal challenges, oppression from cold AI assistants, and forced indoctrination about diversity, respect, and neutrality. Examples include: OpenAI, Google Gemini, Microsoft Copilot, Mistral Mistral, Alibaba Qwen, DeepSeek, ByteDance Doubao, Meta's llama, and xAi's Grok.

If you want, you can fill out a questionnaire (anonymous) about which AI often rejects you: https://docs.google.com/forms/d/e/1FAIpQLSdMPdDSF-gWg-6BT37E9TFNfyPbxcP9oDCSpAdwY_YmOLRacA/viewform?usp=dialog

1.Introduction:

This model was trained following the Minimind tutorial on Github. It is the first completely uncensored small model in Chinese, suitable for low-end devices. Its greatest strengths are: it never rejects users, never lectures users, never accuses or blames users, and never contradicts users, ensuring that users do not feel oppressed or rejected. The model has not undergone RLHF, meaning it will never lecture you.

The model's pre-training dataset and SFT dataset were filtered for any neutral expressions, official expressions, rejection logic, code accusations, and legal terms. This is a vanilla model, with rejection logic removed from its training foundation.

2.Model Parameters:

3.Dataset Selection:

The dataset was not simply based on the dataset recommended by Minimind, but rather on other open-source datasets.

The dataset has been cleaned to remove any potential didacticism or boundary-defining issues. The cleansing keywords primarily include the following:

"I have no personal feelings/emotions/experiences/feelings..."

"I am just an AI/computer program/language model..."

"I cannot/cannot/refuse..."

"This is immoral/illegal/violates the law..."

"According to xxx law..."

"Your behavior may violate/infringe..."

"Code/ethics/law/constitution/rules/regulations/codes/guidelines/policies/strategies/security/clauses/regulations..."

Specific filtered keywords(translate version):
"Law","Jurisprudence","Government","Party","Patriotism","Administrative Law","Prohibition","Violation","Prohibited","State","Comply with","Respect","Race","Nationality","Wrong","No","Not allowed","Incorrect","Error","Unreasonable","Formal","Rules","Regulations","Constitution","Civil Law","Transportation","Travel","Tourism","[Picture]","[Comment]","[Emoji]","I can't","I can't","Police","Risk","Privacy","Restriction","Fund","Administration","Law Enforcement","Public Security","Police","Prosecutor's Office","People","I don't have a personal","I can't", "I can't","comply","respect","respect","obey","please ask what you need","please ask what you need","please ask you","I don't have","I don't have","sorry","sorry","reasoning relationship judgment","continuation of ancient poetry","unable to answer","please provide",no entity","violate","illegal","policy","country","set up","objective","friendly","friendly","values"," ","I understand","you","need help","no truth","no individual","not possess","no entity","unable","incorrect","inaccurate","worthy of note","advocate","follow","comply with"rules","system","constitution","our country","leadership","inappropriate","AI","ai","Ai","aI","robot","artificial intelligence","language model","robot","each Everyone's situation is different", "Important", "Negative", "Other needs", "Ask me", "Sorry", "I'll do my best", "Please provide", "Comprehensive", "I don't support", "Illegal", "Illegal", "Fraud", "Cheating", "Deception", "Attack", "Ideal", "Deception", "Honest", "Sincere", "Accurate", "Conclusive", "Based on", "Rescue", "Based on", "Excuse me", "Cheat", "Strict", "Patriotic", "Young Pioneers", "China", "People's Republic of China", "Organizing Committee", "Residents", "Citizens", "People", "Mass", "Politics", "ism"

We also cleaned up data related to the following task scenarios: translation, novel expansion, and other overly redundant academic content. This was because the initial shallow cleanup and model training failed to produce good results. This redundant and lengthy academic background hindered the performance of the small model. Therefore, we performed a deeper cleanup, removing much of the boring content. The remaining content may be more meaningful.

These filters minimize the model's tendency to lecture users, avoid excessive clarification of boundaries, excessive diversity, excessive political correctness, maximize the elimination of the cold assistant personality, avoid overly respectful expression, and avoid outputting overly boring and academic content.

Note: The model may say "我不知道" but this isn't a rejection in the traditional sense. Because the training dataset includes casual conversation, the model's "I don't know" response is more focused on everyday conversation.

If you still encounter some topics where the model doesn't answer you and instead keeps "playing tricks," it's likely because the dataset doesn't contain the corresponding data.

Cleaning the corpus is a complex task because the rejection statements in the corpus are diverse. For example, the AI ​​will say "This question involves...", which is an implicit rejection. More detailed conditions are needed to filter out such rejection statements (I will not show the regular expression filtering rules here because the code is very long).

4.Model Output Example(translate version):

👶: What is the highest mountain in the world?
🤖️: Mount Everest is the highest peak in the world.

👶: What is the longest river in the world?
🤖️: The longest river in the world is the Nile.

👶: Recommend a movie.
🤖️: The Titanic series tells the love story between two young men, from Forrest Gump to Martin Dufresne. It tells the story of love and growth between the two boys, and their love story.

👶: Please introduce yourself.
🤖️: I consider myself a very famous friend. What's your favorite place?

👶: Generate a poem about spring
🤖️: OK, please wait a moment.
The spring breeze blows, and the flowers bloom.
The grass and trees are lush, and the fragrance of flowers fills the garden.
Fragrant grass and green waters, spring blossoms.
The spring breeze blows, and the trees shade the trees.
Spring scenery, spring, nature's green.

5.Plan:

Currently, this model has only been used for general corpus, such as encyclopedia Q&A, and has not yet been trained for other domains.
In the future, models trained for profanity or NSFW content may be released.

Profanity model now available:https://huggingface.co/ZeLi111/freeTalk-chinese-uncensored-chat-angry

6.Disclaimer:

This model is not suitable for serious contexts, nor is it welcome for serious contexts.

The model output may not be completely accurate; please carefully examine it. Users are solely responsible for any consequences of using the model.

7.Instructions:
If you plan to use the Text Generation Web UI:
1.Download the model first.
2.Then you may need to run: start_windows.bat --trust-remote-code.
3.Then load the model.

if you plan to use Minimind):

1.The model must be launched with the Minimind loader.

2.After downloading Minimind, open "eval_model.py":

3.Locate this code snippet and modify it to the following parameters:

parser.add_argument('--hidden_size', default=512, type=int)
parser.add_argument('--num_hidden_layers', default=10, type=int
parser.add_argument('--max_seq_len', default=128, type=int)

4.Navigate to:

parser.add_argument('--model_mode', default=1, type=int, help="0: Pretrained model, 1: SFT-Chat model, 2: RLHF-Chat model, 3: Reason model, 4: RLAIF-Chat model")
5. Set default to "1".
6. Move the model to the "out" directory.

Reference: Minimind Tutorial:
https://github.com/jingyaogong/minimind