With the development of AI models, it has started to occur to me that just using larger amounts of post-training data might no longer be very helpful.
Synthetic data can help create a massive amount of data samples, but if all of these samples are testing for the same ability, there might not be that much new information gained from the increase in samples.
What I have been more curious about recently is a different direction, where there would be smaller numbers of more difficult examples with clear ground truths.
In my opinion, difficulty is especially important when choosing examples. If something is too easy for a model, it probably would not learn much from it. But if it is completely impossible for it, the training data becomes quite noisy. Useful data should lie somewhere in between – things that show real problems of the model and at the same time have a known solution.
While researching this topic, I stumbled upon Parsewave. This company develops post-training data based on real engineering tasks, as well as evaluations and traces.
For those who deal with SFT, RL or post-training data sets: what is your criteria to select samples that are worth adding?
Is the scalability factor still crucial or the targeted samples that show model failures become more important?