Greetings, Burger!
Created by Claude-NoAssTag author.
Special thanks to the authors of noass_ext, CladeOpus-GigaSchizoKostyl and how2claude, as well as the XML-sensei. Most of the info was yoinked from them.
NoAss (No Assistant) - an experimental prompting method aimed at eliminating the Human-Assistant finetune in LLMs from Anthropic. The Assistant finetune directly affects Claude's output and roleplaying.
Claude (starting from version 2.1) is the LLM (large language model) that generates a self flowing "Q/A" story using internal names "Human" and "Assistant". However, when the model tries to write for "Human", stop tokens interrupt the generation, allowing the real user to insert their query. Without stop tokens, Claude would just keep writing both sides of the conversation by itself. In the actual context it'll look like this:
(In this example, the AI wrote everything - both the "Human" and "Assistant" parts. But in reality, the AI stops before "Human" response, allowing you to provide the real input.)
Proof of such H/A context structure is also the "leaks" of "Human" and "Assistant" during NoAss usage:
This Human-Assistant finetune results in the model being trained to give one specific response to a "Human" query - short, with monologues, biased and with characteristic so called "roundings".
Roundings - rhetorical questions to the user, implying they will definitely respond.
The ball's in your court now.
This behavior is a consequence of model training, it's not a demonstration of the LLM's real capabilities. Human-Assistant roles are artificial, nothing more than context structure limitations from the Anthropic API.
The main idea of NoAss is to swap Human and Assistant with {{user}} and {{char}}. So we're making the LLM think it's writing a story/novel/game etc. at the user's request, while user messages will be inserted into this story from one of the Characters' perspective.
NoAss Extension - an add-on for SillyTavern that implements the "NoAss" prompting method for Anthropic language models.
Goal - to make the model generate continuous narrative rather than Human-Assistant dialogue. This helps reduce model bias and move away from the "question-answer" template.
Enable NoAss
button.Canon
Narrator
Prose
Tag
NoAss with Canonical format - **Prefixes:**
(Canonical):
Canonical NoAss requires good, reworked greetings with proper separation of actions and exposition, as well as strict subsequent format adherence - which may include manual editing.
V4.5 - by CladeOpus-GigaSchizoKostyl author
NoAss with Canonical format - **Prefixes:**
and blocks wrapped in XML tags (Tag):
V3.5.1 - by Claude-NoAssTag author
NoAss with Canonical format - **Prefixes:**
and timestamps in them (SMS):
Simplest preset for chat with one char. Messages have timestamps in prefixes, not visible in Tavern.
Formatting requirements are zero, greeting should just be char's message. Memo assumes char has a profile in {{personality}}, but you can change it to {{description}} if you're too lazy to generate a profile.
Version as of 20:44 13.08.2024 - by noass_ext author
NoAss with Narrator / MultiNarrator
Spiritual successor of Prose NoAss. Narrator (in the form of Tarantino, "Scoof" or someone else) completely replacing the Assistant.
Narrator. V0_3_0 - by noass_ext author
MultiNarrator. Version as of 14:23 29.07.2024 - by CladeOpus-GigaSchizoKostyl author
NoAss with GM
NoAss where {{user}} and {{char}} RP in first person, while GM describes exposition, spawns events and other shit.
Version as of 19:26 08.08.2024 - by CladeOpus-GigaSchizoKostyl author
2nd person. Alternative to regular (Canonical) NoAss format. Looks like "WithAss" without roundings and monologues, and with possible short replies.
Version as of 00:18 17.06.2024 - by CladeOpus-GigaSchizoKostyl author