Ling-3.0-tiny-JEV

Clone the AReno repository, create a virtual environment, and install with pip install -e . on Mac or pip install -e . --no-build-isolation for CUDA. Download the checkpoint with hf download, then run areno serve with --backend mlx or cuda, --model-path inclusionai/ling-3.0-tiny, --decision-checkpoint ./jev, and --port 8000.

It supports Apple Silicon Macs through the MLX backend and CUDA GPUs through the CUDA backend. The quantized MLX checkpoint is 9.45 MB, and the repository lists MLX as a supported library.

Yes. The repository contains only the decision head and rank-128 backbone LoRA adapters, so the inclusionAI/Ling-3.0-tiny base model is required and is passed separately via --model-path when serving or training.

It supports choice, noul, and score questions. A request to the /v1/systemone endpoint carries a state string plus a questions object, for example a route question of type choice with criteria mapping billing and technical to responsible teams.

Run areno train with --algo jev, pointing --ckpt at inclusionai/ling-3.0-tiny, --lora-adapter-path at the downloaded checkpoint, and --dataset-path at the JEV-Hardmix parquet file. Use --q4 for online 4-bit quantization on Mac MLX or --qlora for NF4 QLoRA on CUDA; the rank-128 adapter configuration is restored from the checkpoint.

Training accepts typed-decision parquet files and request/targets JSONL datasets. The model card advises keeping training and evaluation data separate, and notes this release omits the original dataset iterator state so fine-tuning starts with a fresh data cursor.

No. The model is not deployed by any Inference Provider, so it must be run locally through AReno. Downloads are also not tracked for this model, and the training dataset is AsystemAReno/JEV-Hardmix.