I have created an application that will scrape websites, generating fine-tuning data, by chopping the HTML up into Hx tags and paragraphs, in addition to crawl links it finds. Scraping two of my own related websites, it produces roughly 1,500 training “snippets”. However, when I submit this to training, and I use my fine tuned model, the model is only 50% accurate.
Can somebody explain (like I’m 5) how I need to apply settings during training to improve accuracy please?
And also, I’ve got this feature allowing the models to be “reinforced” by logging chat requests, making it easy to edit requests, and generating new training data based upon my edited responses. The idea is to allow for starting out with any website, generating a trained model based upon that website, modify it manually by human supervision, and “retrain” it again. How should my settings for the retraining be applied …?
You can see the product, and try it out, at AISTA (dot) com …
It’s also open source if you search for it. I can’t post links here unfortunately, but you can follow the bread crumbs provided above to find it if you’re really interested …