--- license: cc-by-nc-4.0 base_model: Qwen/Qwen3.8-27B tags: - classification - custom-code - decision-model --- # Lychee-27B Lychee-27B is a decision model fine-tuned from Qwen3.8-27B. Given a state and typed questions (`choice`, `noul`, `score`), it returns a probability for every option in one forward pass. ## Results | Decision Index 0.3 | Jev | Lychee-27B | |---|---:|---:| | Knowledge & Reasoning | **53.9** | 51.8 | | Language Understanding | 59.2 | **66.2** | | Retrieval & Classification | 55.4 | **63.9** | | Tools & Automation | 75.1 | **80.8** | | Arts & Human Taste | 39.1 | **47.0** | | Public index | 58.0 | **62.8** | Bold marks the better score in each row. Decision Index 0.3, public suite: Jev's figures are the board's, Lychee-27B's were measured with the board's kit against the server in this repository. ## Usage Python 3.12 and one CUDA GPU with 80 GB (the bf16 weights take about 52 GB). ```bash pip install -r requirements.txt python -m lychee.server --model . --port 8000 ``` The server answers `POST /v1/systemone` with `{"state": ..., "questions": {name: {"type", "instructions", "criteria"}}}`. In Python: ```python from lychee import Lychee model = Lychee(".") answers, _ = model.decide( "I was charged twice for the same order.", {"team": {"type": "choice", "instructions": "Which team should take this?", "criteria": {"billing": "Payments and refunds", "shipping": "Delivery problems", "returns": "Sending items back"}}}, ) print(answers["team"]) # the chosen option and a probability for every option ``` A `noul` question gives the probability of yes; a `score` question takes its levels as a list, lowest first. Prompts are read whole up to 262,144 tokens, the base model's context; longer ones, or ones that do not fit in the GPU's memory, are refused, never truncated. ## License CC BY-NC 4.0 (`LICENSE`). Lychee-27B is fine-tuned from Qwen3.8-27B, which is licensed under the Apache License 2.0 (`LICENSE.apache-2.0`); `NOTICE` lists the changes.