LLM Verbal Tics on Bitsy Wiki

Mitigating Verbal Tics

Mon, 01 Jan 0001 00:00:00 +0000

Mitigating a verbal tic means lowering the rate at which a large language model produces one construction, without damaging the text around it. There are four places to do it: the prompt that asks for the text, the sampler that picks each word, the model’s weights, and the finished draft. The obvious method is to tell the model not to use the phrase. It belongs to the first of the four, and it is the method with the most evidence against it.

Verbal Tics by Application

Mon, 01 Jan 0001 00:00:00 +0000

Which mitigation for a verbal tic is available depends on how much of the system around the large language model a builder controls, and which tics are worth the effort depends on the kind of text being produced. A person typing into a chat app controls their own messages. An integrator calling a hosted model controls the system prompt, the standing instructions sent ahead of every conversation. Someone running a model whose weights are published controls the sampler and the weights as well.

Claude's Verbal Tics

Mon, 01 Jan 0001 00:00:00 +0000

This page is an open list of the verbal tics of Claude, the family of large language models made by Anthropic: the wordings that recur in Claude’s output whatever the subject is. Each entry names a frame, gives one example, and records how the habit is known. A frame is a pattern with an open slot, such as not X but Y, and the slot is the reason a list of banned phrases does not catch it. New entries are added as they are noticed, and an entry keeps its number.