The transformers serve command now includes a /v1/systemone endpoint to serve any autoregressive Hugging Face model as a decision model using next-token probabilities.
This endpoint reads label-token scores from an existing model rather than training a new decision model. It merged into Transformers main on 12 October, so a source install may be needed. Configuration controls the labels and prompting. Reported template-related errors and test follow-ups remain limitations.
How you can use it
Configure how the server represents the labels, then send a situation and questions with fixed answers. The server reads the chosen model’s scores for those answers without generating a reply. You are adding a decision interface to an existing model, not training another one.