Questions about details of dataset usage for reproducing the openchat_3.5
Read your paper, that 's awesome.
However, I inferred from you dev log that openchat_3.5 is actually the improved version of the one that introduced in your paper. Could you please provide more details about how to combine those auxiliary datasets listed in openchat_3.5 model card for reproduction? I noticed that "custom data process pipeline" was used, I wonder if this operation is actually the same with the data encoding script in your github code repo. Also, it would be appreciated if you could tell me how to assign weight for these auxiliary datasets for C-RLFT training, such as:
Capybara 1 2 3
GOAT
Glaive
MetaMathQA
MathInstruct
OpenAssistant
According to your paper, these dataset should be assigned with weight 1, because they seem like high quality data.