#Help with SFT Qwen3 model

12 messages · Page 1 of 1 (latest)

shy cape
#

I am following the Qwen3 SFT notebook but training using my own data.

#

Everything seems to go fine but after finetuning the model, it does not seem to be generating the thinking tokens. The training examples have these thinking tokens.

#

Feels like this might be something small / silly on my side. Any help will be appreciated!

upbeat dew
#

Your prompt format doesn't quite look right.

Edit: Scratch that, I copied the data into a text editor and parsed it, looks like standard multi-turn. Not every turn has a <think></unthink> tag though\

#

You can obviously see that the training is working because the assistant is replying IN ALL CAPS ALL THE TIME

#

But if all of your training examples are like this, with 90% of the turns having no thinking, and only the final turn having thinking, the model will want to mimic that

shy cape
#

This makes sense and is super helpful. Thank-you!!

upbeat dew
#

Sorry I googled the names of the medication in the chat, and the data is all there

#

So the <thinking> section is reasonable

marsh tinsel
#

who is on medication?

upbeat dew
#

The person in this fellows dataset 🙂