I don't know how to code, but I fine-tuned an AI model last week.
Let me clarify my background first: I'm not an engineer, I haven't studied machine learning, and I can only copy and paste others' Python code, which usually throws errors.
So when I say 'I fine-tuned an AI model last week,' you might think I'm clickbaiting.
But I didn't.
Here's how it all started. I was researching the ecosystem tools of @OpenLedger and saw the introduction of ModelFactory claiming 'purely GUI operation, no command line, no API integration.' My instinctive reaction was—I've seen this kind of talk too many times, clicking in must be a scam.
I clicked in.
The interface is surprisingly clean. On the left is the dataset request entry, in the middle is model selection and parameter configuration, and on the right is the training monitoring panel. It didn't ask me to fill in any environment variables or install any dependency libraries.
I selected a small dataset that I organized myself—about two hundred Q&A pairs on traditional Chinese medicine dietary therapy, which I had accumulated while helping a TCM clinic with content organization, and it had been sitting unused on my hard drive. I uploaded it to Datanet, waiting for review approval, which took less than a day.
Then I chose the base model in ModelFactory, configured the LoRA parameters (the interface had explanations, and I understood about seventy percent), and submitted the training task.
The waiting process was a bit boring. I thought it would be quick, but I ended up staring at the progress bar for nearly forty minutes.
Then it finished running.
I opened the built-in dialogue testing interface and asked a question: 'What should I eat in winter if my hands and feet are cold?'
It provided an answer that was structured, had a logical reference, and was different from the vague responses given by general large models—clearly reflecting the tone and emphasis of my dataset.
I stared at the screen for about thirty seconds.
Honestly, there's a bit of a strange sense of achievement. It's not because the technology is so complex; on the contrary, it's because this should have been difficult, but it wasn't. My two hundred manually organized data points really turned into a somewhat capable vertical Q&A model.
There was a line in the white paper I hadn't paid much attention to before: the core bottleneck of specialized AI isn't model size but high-quality vertical data. Now I understand. My two hundred TCM Q&As, if replaced with general scraping data, would definitely not yield this level of precision.
Of course, there are issues too.
I didn't fully grasp the RAG attribution module, it showed the data source, but the format is still a bit technical for the average user. As for the BLEU and Rouge scores in the benchmarking section, I need to look them up to understand if they're good—this isn't friendly enough for non-tech backgrounds; it should just translate to 'better than general models by XX%'.
There's a real issue: the review cycle of Datanet. If the uploaded data needs to wait more than a day to start training, it can be a bit of a bottleneck for those wanting to iterate quickly.
But the sense of achievement is real.
I have a friend who's working on K12 educational content, and he has accumulated years of wrong question analysis data, but he never knew how to use it. I'm planning to suggest he try ModelFactory.
If someone who can't code can fine-tune a usable specialized model, the boundaries of this thing are much wider than most people think.