jingyaogong/minimind ? reverse-engineered prompt

Build me a small but complete app for training and trying out a tiny language model from scratch. I want it to be able to pretrain on a jsonl dataset, then do supervised fine tuning, and also support simple LoRA style fine tuning and preference style training if that is already part of the project. It should come with a basic chat page where I can talk to the model, see its thinking when enabled, and try tool calling in a simple way.

Please also include a way to run evaluation on a few common language model benchmarks, and a lightweight API server that works like an OpenAI style chat endpoint so other tools can connect to it. If there are scripts or examples already in the repo, wire everything together so I can run the full flow without having to figure out the internals. Keep it simple, readable, and easy to launch on one machine.