ENPT

DecisionFormer

Last updated:

RuntimeApache-2.0~20 ★

⚠️ Why not a model

A library to fine-tune, load and serve decision heads on any Hugging Face model. It ships no checkpoint of its own; quality comes from the backbone and data you train.

Technical specs

Base LLMs
any Hugging Face backbone, vllm-sr/d3-edge (README example), Qwen3.5 (Gated DeltaNet kernels)
Decision types
choice, noul, score
Features
fine-tune, qlora, http-systemone, vllm-backend, python-library, cli
License
Apache-2.0

Links

publicado (OK do Fabricio 24/10 BRT).

DecisionFormer (FrontiersMindAI/decisionformer, repo of 22 Oct, announced 24 Oct) calls itself “Transformers for decision models”. You give it a state and typed questions (choice, noul, score with 2 to 10 levels) and it returns a probability for every option in one forward pass, without generating text.

pip install "decisionformer[serve]"
  • Inference: AutoDecisionModel.from_pretrained(...) then model.decide(state=..., questions=...); decide_many for batches.
  • Training: one command to fine-tune a decision head on any Hugging Face model; qlora extra for bitsandbytes.
  • Serving: serve extra exposes /v1/systemone (FastAPI); vllm extra uses a vLLM backend; fla adds Gated DeltaNet kernels for Qwen3.5 backbones.
  • Requires Python 3.10+, PyTorch 2.6+, transformers 5.20+.

Why it is not a model

It is a library. It ships no weights; the README example loads vllm-sr/d3-edge. Note that its confidence is 1 minus the normalised entropy, not the top probability, so it is not directly comparable with other APIs.

Apache-2.0, about 20 GitHub stars on 24 Oct; no independent benchmarks.