DecisionFormer: a Transformers-style library to fine-tune and serve decision models on any backbone
publicado (OK do Fabricio 24/10 BRT).
FrontiersMindAI published DecisionFormer (GitHub, repo of 22 Oct, announced 24 Oct), an Apache-2.0 library that brings decision models to any Hugging Face backbone: AutoDecisionModel.from_pretrained(...) and decide(state, questions) return a probability per option in one forward pass, a single command fine-tunes a decision head (QLoRA optional), and the serve extra exposes /v1/systemone, with an optional vLLM backend.
It ships no weights, so it goes in the catalogue as a runtime, not a model. Its confidence is 1 minus normalised entropy, which differs from other APIs. No independent benchmarks yet.