I suppose for the MiniCPM-llama3-V 2.5, you followed the same pre-training, sft, and alignment schema, but on what data?
Is the data you trained MiniCPM-llama3-V 2.5 open? It seems the details in the MiniCPM series are only related to text input.
Where can I find more details on the pre-training / fine-tuning settings and data for MiniCPM-llama3-V 2.5?
Moreover, For the MiniCPM series, you released many versions, such as the sft, aligned and I suppose the base too.
Regarding the MiniCPM-llama3-V 2.5 it seems there is available only one model (full, int4), but only the (I think) aligned version (after the RLAIF-V alignment).
Can you provide more details or some references to better grasp that information?
Thanks, Lele
Hello, and Thanks for the amazing work! Very much appreciated :)
I have read the MiniCPM-V 2 blog and the sources you cite in the blog post, as well as the MiniCPM blogpost. Also, I tried to check the available dataset you provide on the huggingface dataset page and also a recent issue discussion where you named the training data but I couldn't find the details I'm looking for.
My questions are:
Moreover, For the MiniCPM series, you released many versions, such as the sft, aligned and I suppose the base too. Regarding the MiniCPM-llama3-V 2.5 it seems there is available only one model (full, int4), but only the (I think) aligned version (after the RLAIF-V alignment).
Can you provide more details or some references to better grasp that information? Thanks, Lele