Keeping a fine-tuned model portable

People think portability means having the weights. It means having the weights, knowing exactly which base model and version they apply to, and still holding the dataset that produced them.

An adapter without its base pairing is close to useless — applied to a different version it produces degraded or nonsensical output, and nothing in the file format warns you. The dataset is the insurance: with it you can rebuild even if the rest is lost.

A generated model card recording base model, version, template, hyperparameters and final metrics is the right container, because it travels with the weights rather than living in someone’s memory or a wiki page that gets archived.

This is the main practical reason to keep the card alongside the model rather than treating it as documentation for other people.

Serving on the OpenAI chat-completions shape means moving providers is a base URL and a key. Chat completions, streaming and tool calling all use the same request and response objects, so application code does not change.

Compatibility is a spectrum at the edges — log probabilities and some structured-output modes vary. Check the specific features your client depends on rather than assuming the whole surface transfers.

Pushing a finished model to a private hub repository puts a copy somewhere the training platform does not control. It costs nothing, it is not public, and it means an account problem is an inconvenience rather than a loss.

Portability that has never been exercised is a claim. Download a model, load it somewhere else, and confirm it produces the output you expect — once, deliberately, while nothing is on fire.

The most common thing this exercise uncovers is that nobody recorded the base model version.