on Wednesday, with many customers being restored before then. The majority of the Bay State customers without power live in South Shore and MetroWest communities.Īfter the outages on Monday peaked at around 300,000 Massachusetts households, the number of customers in the dark had dropped to about 60,000 as of Tuesday evening - while power line and tree crews work around the clock to restore power following the storm’s violent winds and heavy rain.Įversource expects to complete restoration for most customers by 11 p.m. Anyhow, what you asking seems possible without fine-tuning a model especially if you have a way to have python grab information from your documents instead of expecting the model to remember everything perfectly.The fallout from this week’s powerful storm continued on Tuesday, as tens of thousands of households remained in the dark with power company crews working 24/7 to turn on all the lights before the holiday weekend. ![]() ![]() I'm learning this still and interesting in adding a lot to my local LLMs as hobby, and see what I can do with it. You can set up chat context so the model knows what's going on before a chat starts.ĭevelopment time could take a while depending on resources you have of course. If you can get function calls, have capable GPUs, I believe even a modest model could do this for you, something like mixtral (modest in size and seems capable). Check out microchain on GitHub for more on that. I'd say even an LLM that's not trained on those specific documents could handle something like this if it was integrated to work with python, do function calls in the chat (some keyword in chat activates a python function to look up a certain document or other functions and it can feed information to the ai from what I understand. This is still not an exact science, though it's getting there.) (You can start with a high-ranked LoRA, but to really get the core concepts sunk in the deep parts of the network, a fine tune does it a little better. Now you can do a full fine-tune, to improve the generation further. You should try to make sure they are high-quality (so ruthlessly edit them to remove any errors or ambiguity) and diverse (so they cover the entire variety you want to generate). Once you've reached this point, you should have a large number of examples of what kind of output you want that you can use as the starting point of a synthetic dataset. For your purposes, you might not even need RAG: if you know exactly what documents are relevant, you can just make a list and stick those in the prompt. At this point it'll be close to a fancy name replacement, so for your purposes that might be enough.įrom there, you can add RAG (looking up the relevant documents automatically). You're not trying to get it to remember facts, you're trying to get it to follow the style of your format. Then, if your format is slightly unusual, train a low-rank LoRA on your specific format. So have two examples and ask the LLM to generate a third one. Then try multi-shot prompting: this just means having more than one example in the prompt. ![]() If you have specific data in mind, you might have something that can work with only a little editing. I mean the whole prompt, including reference documents and so forth. Start by writing a few examples of the kind of thing you want.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |