kitft/natural_language_autoencoders ? reverse-engineered prompt

Build me this Natural Language Autoencoders project so I can take a model activation vector, turn it into a short natural language explanation, and then reconstruct the original signal from that explanation.

I want the repo to support both training and inference, with a clean path to run the released checkpoints locally and a simple command I can point at a parquet file of activation vectors. It should also include the data generation, the supervised fine tuning step for the reconstructor, the fine tuning step for the verbalizer, and the reinforcement learning stage that trains both together. Make sure it can load the model specific metadata from each checkpoint instead of hardcoding anything, and include the example workflows and docs so someone can try it end to end.

If you need current setup details for Miles or SGLang, look them up online and follow the project docs.