Fine-tuned LLMs for article generation

Anybody using fine-tuned LLMs for article generation? Is it worth it? How many articles did you use to train? Do I need to train per nice or would it be fine using say 1k articles of mixed niches?

I mean, I just use Chat GPT with a good prompt. If you know how to edit them, it should not be a problem.

But I am interested what LLM you are using, if you can share.

Not happy with usual ChatGPT, too many of vibrant, delve and other wise-ass things. I want fully automatic content, with absolutely no editing. I am mostly using OpenAI but will soon change to some other model, hosted on a Hetzner GPU-server. Probably phi3-small, looks very promising since very cost-efficient.

I plan to train with a bunch of #1-3 SERP blog posts about different niches and around 1k articles. The LLM should understand how to write a non wise-ass but helpful and engaging article. Would be thankful for input of other BHW people, several are already using fine-tuned LLMs for different purposes. If you read articles about prompting vs fine-tuning, opinions are very mixed. RAG is by the way imho a very important thing definitely to use, eliminates hallucinations to a minimum. But the style of the generated article is still rather wise-ass.

Regarding prompts: I tried a lot - result is ok, but not good enough. The best I have seen is rephrasing an existing article. Maybe that is the way to go.

Bump! Any advice here?

I think splishsplash is running custom models. He gets decent output for real business cases.

You can try single niche on vertex AI using their structured prompts that contains 500 input -> output fields. They have free credits.

Many niches input will of course kill the quality.

If you want to train, you'll have to find an opportunity.

I believe for some niches it can be awesome, for some it'll be incredibly hard.

AI needs focus which boils down to attention mechanism. If you don't let it focus on one thing, it will be miserable like a human who does 50 things at once.

reaaski said:

AI needs focus which boils down to attention mechanism. If you don't let it focus on one thing, it will be miserable like a human who does 50 things at once.

Click to expand...

Yes, it will make AI models process complex data effectively, especially with tasks that require understanding long-term dependencies or relationships within the data.

I should make myself clearer: I wanna use fine-tuning to better imitate a natural way of writing, that is my only intent. Does that make sense?

For the information/content itself, I provide context via RAG.

cooooookies said:

I should make myself clearer: I wanna use fine-tuning to better imitate a natural way of writing, that is my only intent. Does that make sense?

For the information/content itself, I provide context via RAG.

Click to expand...

if you have articles written in a style / voice you wanna finetune on, create fake RAG context which contains necessary details to write some article and then train on those pairings. You could also just finetune on the raw text itself but it wont be nearly as effective as using a constrained template for your use case.

If you already have a dataset created I wouldnt mind giving it a shot. I've finetuned models in a coding / planning niche and im curious how my methods transfer over to writing.

cooooookies said:

I should make myself clearer: I wanna use fine-tuning to better imitate a natural way of writing, that is my only intent. Does that make sense?

For the information/content itself, I provide context via RAG.

Click to expand...

Hi, I am trying something quite similar, I wish to create a chatbot that does not act like the common "Helpful Assistant", but rather impersonate a real human, for example a salesperson. For that I need to make the AI break character and switch to a different way of "speaking" to say.

My guess is that it needs to be fine-tuned, since using RAGs or some other form of data retrieval wont help me in this scenario.

Do u have any tips / guides I could follow? And did u do this yourself ?