eachadea

/

ggml-vicuna-13b-1.1

like 351

From what I understand:

• The uncensored version is trained on all the ChatGPT chats that have been shared on ShareGPT ('v0 release').
• The censored version is trained on a curated collection ('v1.1 release') where a lot of the excuses and mentions of what OpenAI told ChatGPT to do are removed.

E.g. the 'uncensored' version will act more like a distilled version of ChatGPT. But it will also tell you that it was created by OpenAI for example. This is probably the opposite of what people expect 😅.

The way you need to prompt the system to make it act like an LLM assistant is also different: https://huggingface.co/eachadea/ggml-vicuna-13b-1.1#info

For example llama-rs ('old' pre-PR #1405 ggml format) uses a kind of hybrid prompt 🤔

    let character_name = "### Assistant";
    let user_name = "### Human";
    let persona = "A chat between a human and an assistant.";
    let history = format!(
        "{character_name}: Hello - How may I help you today?\n\
         {user_name}: What is the capital or France?\n\
         {character_name}:  Paris is the capital of France."
    );

The uncensored models aren't just v0 - they're independently trained based on the v0 release but with a cleaned version of the dataset

From what I understand:

• The uncensored version is trained on all the ChatGPT chats that have been shared on ShareGPT ('v0 release').
• The censored version is trained on a curated collection ('v1.1 release') where a lot of the excuses and mentions of what OpenAI told ChatGPT to do are removed.

E.g. the 'uncensored' version will act more like a distilled version of ChatGPT. But it will also tell you that it was created by OpenAI for example. This is probably the opposite of what people expect 😅.

The way you need to prompt the system to make it act like an LLM assistant is also different: https://huggingface.co/eachadea/ggml-vicuna-13b-1.1#info

For example llama-rs ('old' pre-PR #1405 ggml format) uses a kind of hybrid prompt 🤔

    let character_name = "### Assistant";
    let user_name = "### Human";
    let persona = "A chat between a human and an assistant.";
    let history = format!(
        "{character_name}: Hello - How may I help you today?\n\
         {user_name}: What is the capital or France?\n\
         {character_name}:  Paris is the capital of France."
    );

You have censored and uncensored mixed up in your explanation.

Censored means it is restricted. This version does the excuses
Uncensored means it is not restricted. Here the excuses are removed.

no. the v0 files are not just v0 files - they're custom trained by a 3rd party group unaffiliated with LMSys.
this is the dataset used: https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered

As I understand it, but @eachadea probably knows best, but fails to explain properly (I think).

Uncensored files are based on v0 release == the first ('zeroth') dataset release, contains all the excuses that ChatGPT also makes.

Main files are based on v1.1 release == excuses are removed (the excuses are censored).

Is this confusing, from the general understanding what "censorship" means?

Yes.

But that is what it is.

you are correct that the v0 dataset contained all the "I am sorry, but as an AI..." ChatGPT garbage, but the uncensored v0 files on this repo are modified, see above dataset I sent

So to paraphrase @CR2022 , but correct his mixup:

Censored means [..] the excuses are removed.
Uncensored means [..] this version does the excuses.

Is this correct ?

uncensored means no "as an AI, I have a stick up my backside"