The synthetics used for the training seem to be not just based on the same format time, but also the same user prompt "how does redis pub/sub work"
yes but as you see in the results it transfers to prompts in that format that were never trained on really well