AReno-Jev-9B

It is a JEV decision head plus a rank-16 LoRA adapter for Qwen/Qwen3.5-9B, trained with AReno. It supports choice, noul, and score questions and scores the candidates you provide rather than generating a free-form answer.

Yes. The original Qwen/Qwen3.5-9B base weights are required and are not included in the repository. Download the backbone separately, for example into a local Qwen3.5-9B directory, and keep the adapter repository alongside it.

Set JEV_BASE_MODEL to the backbone path, JEV_CHECKPOINT to the adapter directory, PYTHONPATH to include the parent directory, SGLANG_EXTERNAL_MODEL_PACKAGE to jev, and SGLANG_HTTP_SERVER_APP_PATH to jev.sglang_jev:app, then run sglang serve with --is-embedding, --trust-remote-code, and --context-length 4096. /v1/systemone is served on the same port.

LoRA rank is 16 and alpha is 32, the maximum context is 4096 tokens, and decision temperature is 1.0. Training used CUDA NF4 QLoRA, and the adapter and FP32 head are published unchanged.

The model card lists an apache-2.0 license. The repository is published by AsystemAReno on Hugging Face with Safetensors and MLX tags.

No. The model adapter supports text states only; image and audio attachments are not processed. Requests are typed decisions submitted to /v1/systemone with a state and a set of questions.

Run the included merge_weights.py with --base-model pointing at the original Qwen3.5-9B weights, --checkpoint at this repository, and --output at a new directory. The script preserves the tokenizer and JEV head and copies sglang_jev_merged.py into the merged directory.