Training data composition

[bollossom]

Nice work. I have another question Are the 130M pairs of data used in the three stages of PT Alignment CT the same? What is the ratio of understanding to generation data?

[Orlando-CS]

In Unipic1, minimal understanding data (1%) to enable functionality while maintaining generation priority.

[bollossom]

Is the 130M data the same at each stage, such as PT alignment CT, or is it different at each stage?

[Orlando-CS]

The 130M data is not the same across different stages (e.g., PT, SFT). In fact, the data volume typically decreases as the stage progresses.

[bollossom]

But in TR. You training samples is always same.