ridgerchu/spikegpt ? reverse-engineered prompt

Build me a clean, working implementation of SpikeGPT, a lightweight generative language model that uses spiking neural network style binary activations.

I want the repo to support training on Enwik8, pretraining on a large tokenized corpus, fine tuning on WikiText 103, and running text generation from a prompt. Please make the training script easy to edit for dataset paths and model settings, and make the inference script generate sensible sample text from a custom context. If there are missing pieces or old dependencies, update them so the code runs smoothly with current Python tools, and look up current docs online if you need to.

Keep the structure simple and practical, with clear scripts for training, inference, and the NLU fine tuning examples already mentioned in the project. If a pretrained checkpoint is needed for inference, wire the code so it can load one from the project root without extra hassle. Also make sure the project is documented enough that someone can follow the README flow and actually reproduce the main use cases.