[braindead_in]

The 34b Python model is quite close to GPT4 on HumanEval pass@1. Small specialised models are catching up to GPT4 slowly. Why not train a 70b model though?