Back to collectionEval Engine Releases Open Decision-4B Adapter
The 65MB LoRA adapter turns Qwen3.5-4B into an option-scoring model and is published with code, limits and an issuer-run benchmark.
Eval Engine released Decision-4B v0.1 on September 25 as a 65MB LoRA adapter for Qwen3.5-4B. The model takes a state, a question and two to 24 listed options, then scores only those options in one forward pass rather than generating free-form text. The repository includes weights, example inference code and an Apache 2.0 license for the adapter and code.
The model card links to Unbound for web and mobile use and lists important limits: English only, a 2,048-token context window, weaker performance on multi-rule policies and response-quality grading, and no intended use for unattended high-stakes decisions. Its benchmark is self-reported by the publisher, so the reported scores should not be treated as an independent evaluation.