I wrote a while ago that I fine-tuned a stock market model.
After using it for a few days, phi4 seems to be good, but the context is too small at 16k...
Other models have 128k...
Phi4 was created in 2024, so maybe that's why the context is small. The performance itself is good, but it seems to be too difficult to use for stock reports due to the small context.
So I brought in a new qwen3-14b model and fine-tuned it.
Phi4 is specialized in mathematics and statistics, so the quality is better than qwen3.... but there are no other alternatives for 14b models, so I decided on qwen3. It has a good 128k context, so I fine-tuned it.
Hmm....
Uh...
I thought something was wrong, so I asked various AIs. They said that the model itself is weak and fine-tuning is a micro-adjustment, so it doesn't change much. I was disappointed, but then I wondered if my fine-tuning data wasn't good???
So I did some research.
I found that quality is more important than quantity. Before, I just threw in a bunch of data, didn't care about the format, and wasn't systematic. I just put in everything that looked good.
So now I'm making about 1000 examples using official data from the stock exchange + free financial textbooks + Q&A examples. I'm fine-tuning it right now.
There's a paper that says 1000 well-refined textbooks and examples are more effective than 10,000 data points, so I'm following that.
I keep forgetting that LLMs need to be educated like humans. I assume that computers can figure things out automatically just by throwing in data, but those automatic processes are also macros created by developers.
