Post-train a 7B model to code better for one enterprise

An enterprise customer wants a 7-billion-parameter model that writes better code for them. Design the post-training pipeline end to end: collecting and cleaning data, supervised fine-tuning, RLHF or another preference or reward step, evaluation before release, and the inference setup that serves the result. Say where the customer's own code and feedback come in and how you decide the new model is actually better.