Seems ChatGPT didn't learn well about Tiddlywiki

Even gpt4 will generate tid like this

title: RecoveryLogging
tags: CoreFeatures

/* Fields for recording recovery progress */
date_of_recovery: 
notes:

This tiddler provides a structure to log recovery progress.

while GPT4 knows very well about noflo, a library nearly as old as tiddlywiki. It will generate very useful code, and answer question about all concepts.

For example, Question: why next node is triggered without previous node call `output.sendDone` yet · Issue #1047 · noflo/noflo · GitHub

Maybe its tiddlywiki.com that not friendly to OpenAI’s crawler, that stops ChatGPT learn Concept and syntax from us?

The static site should alleviate the crawler-unfriendly design.

As voluminous as the documentation is, though, it’s not created to present how-do-I-style documentation. That’s probably the style used to populate LLMs. If it weren’t for this forum, my progress would be tremendously slower. I wonder if this forum is used in the models; I think it might offer much more benefit than the main site for this process.

I wonder if LoRA fine-tuning on Llama 2, or something like that, might not produce a much more useful tool for those who need an assistant, agent, or rubberducky more fluent, accurate, or focused on TW5. What sorts of data would we need to collect? How much would we ideally have collected? How much of the training data, if any, could be at least partially synthesized yet worthy?

1 Like

there is a request Can I get talk forum and gg dataset to train AI? but seems @boris didn’t reply.

Despite of state version of official website, I think chatgpt could have learnt alot from sourcecode on github, many core feature is written in wikitext…But it didn’t learn well.

There’s also a static version of TW documentation. Maybe CG doesn’t understand how all the pages relate.