From c4d521c29edd89ad6929cd043b0f05171a9e3445 Mon Sep 17 00:00:00 2001 From: ed Date: Tue, 17 Sep 2024 11:50:10 +0200 Subject: [PATCH] vault backup: 2024-09-17 11:50:10 Affected files: Untitled 3.md --- Untitled 3.md | 12 ------------ 1 file changed, 12 deletions(-) diff --git a/Untitled 3.md b/Untitled 3.md index 0894bb0..b28b04f 100644 --- a/Untitled 3.md +++ b/Untitled 3.md @@ -1,15 +1,3 @@ - -"Conditional instructions: You could modify your system prompt to make the fallacy detection conditional, e.g., "You are a helpful assistant. When asked or when relevant, you can act as a critical thinker good at unveiling logical fallacies." This might help the model understand that fallacy detection is a capability, not a constant requirement." - That's a big no. I want my model to be able to always use critical thinking. If user chat and talk about fake news or dangerous cult like thinking, I want my model to engage in "street epistemology" - -"Multi-task training: Instead of focusing solely on fallacy detection, you could include a variety of critical thinking tasks in your dataset. This broader approach might lead to a more balanced model." Yes, there will be multiple dataset and before I'll start training I'll merge all the parts of the system prompt, the goal being to have one fine-tuned model that works well with one specific system prompt, that will surely help to add "normal" examples. - -"Adversarial examples: Include some examples where a user incorrectly identifies a fallacy, and the AI correctly points out that there isn't actually a fallacy present." - Great idea! - -"Context-aware responses: Train the model to consider the broader context of a conversation before applying fallacy detection. This could help it understand when such analysis is appropriate." - Yes, I will definitely need multi-turn chat examples - -"Explicit "no fallacy" examples: Include examples where the AI explicitly states that it doesn't detect any fallacies in a given statement or argument." - Nah - -"Gradual fine-tuning: Start with a more general critical thinking dataset, then progressively introduce more specific fallacy detection examples. This might help the model develop a more nuanced understanding." - That's interesting, are you suggesting that splitting the fine-tuning in sets of "difficulty" will help? Like I first fine-tune Llama3.1 with simple example, save the weights, then fine-tune again with medium, then hard? \ No newline at end of file