4 ms·
I've been getting most of my work from Upwork in the period since generative AI started to take off. I can clarify something about the number of ML jobs. 90% of
by ilaksh 3y ago
I've been getting most of my work from Upwork in the period since generative AI started to take off. I can clarify something about the number of ML jobs. 90% of my clients are convinced or nearly convinced that they cannot fulfill their chat agent requirements without fine-tuning a new LLM.
I do have at least one job in my profile that involved fine tuning. That may explain why some of the clients hired me. But one thing to note is that I do not claim to be a machine learning engineer because I'm not. I say I'm a software engineer with a recent focus on generative AI.
0% of the clients actually need to fine-tune an LLM in order to fulfill their requirements. What they need is any LLM close to the state of the art, temperature 0, and a good system prompt. Maybe some function/tool calling and RAG.
The one guy that I did fine-tuning for, he kept telling me to feed small documents into qLoRA without generating a dataset, just the raw document. I did it over and over and kept showing him it didn't work. But it did sort of pick up patterns (not useful knowledge) from larger documents, so he kept telling me to try it with the smaller documents.
Eventually, I showed him how perfectly RAG worked for his larger documents like a manual. But he still kept telling me to run the small documents through. It was ridiculous.
I also ended up creating a tool to generate QA pairs from raw documents to create a real dataset. Did not get to fully test that because he wasn't interested.
Anyway, the SOTA LLMs are general purpose. Fine-tuning an LLM would be like phase 3 of a project that is designed to make it faster or cheaper or work 10% better. It is actually hard to make that effort pay off because LLM pricing can be very competitive.
Machine learning knowledge is not required for fine-tuning LLMs anyway. You need to understand what format the dataset goes in. And that is very similar to prompt engineering which is also just a few straightforward concepts that in no way require any degree to understand. Just decent writing skills really.
- snowfield 3y agoDoes fine tuning datasets need to be structured a specific way or can you take unstructured data?
- geoduck14 3y agoYou need to structure it in the form of "if the user says X, you say Y." For example: if the user asks "where do I find red pants," say "we don't sell red pants, but paint can be found here" The OP gave a quick example. You can take raw docs and generate a Q/A data set from it, and train on that. Generating the Q/A data set could be as simple as: taking the raw PDF, asking the LLM "what questions can I ask about this doc," and the feeding that into the fine tuning. BUT, and this is important, you need need a human to look at the generated Q/A and make sure it is correct. Key in this. Don't forget: you can't beat a human deciding what is the "right" facts and responses that you want your LLM to produce