Ling-3.0-tiny-JEV

A JEV decision model based on inclusionAI/Ling-3.0-tiny, trained with AReno.

Supports choice, noul, and score questions on Apple Silicon Mac and CUDA. This repository contains the decision head and rank-128 backbone LoRA adapters. The original base model is required.

Installation

git clone https://github.com/inclusionAI/AReno.git
cd AReno
python3 -m venv .venv
source .venv/bin/activate

Mac:

pip install -e .

CUDA:

pip install -e . --no-build-isolation

Download the checkpoint:

hf download AsystemAReno/Ling-3.0-tiny-JEV --local-dir ./jev

Serving

Mac:

areno serve \
  --backend mlx \
  --model-path inclusionai/ling-3.0-tiny \
  --decision-checkpoint ./jev \
  --world-size 1 --tp-size 1 --port 8000

CUDA:

areno serve \
  --backend cuda \
  --model-path inclusionai/ling-3.0-tiny \
  --decision-checkpoint ./jev \
  --world-size 1 --tp-size 1 --port 8000

Example request:

curl http://localhost:8000/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{
    "state": "The customer requests a refund.",
    "questions": {
      "route": {
        "type": "choice",
        "instructions": "Choose the responsible team.",
        "criteria": {
          "billing": "Billing",
          "technical": "Technical support"
        }
      }
    }
  }'

Fine-tuning

Fine-tuning updates both the decision head and backbone LoRA adapters. The adapter configuration, including rank 128, is restored from the checkpoint.

Download the training dataset:

hf download AsystemAReno/JEV-Hardmix --repo-type dataset --local-dir ./jev-data

Mac with online 4-bit quantization:

areno train \
  --algo jev --backend mlx \
  --ckpt inclusionai/ling-3.0-tiny \
  --lora-adapter-path ./jev \
  --dataset-path ./jev-data/data/train-00000-of-00001.parquet \
  --save-path outputs/jev \
  --q4 \
  --world-size 1 --tp-size 1 \
  --batch-size 16 --mini-bs 4 \
  --lr 2e-5 --max-context-len 4096 \
  --epochs 3 --save-interval 20

CUDA with NF4 QLoRA:

areno train \
  --algo jev --backend cuda \
  --ckpt inclusionai/ling-3.0-tiny \
  --lora-adapter-path ./jev \
  --dataset-path ./jev-data/data/train-00000-of-00001.parquet \
  --save-path outputs/jev \
  --qlora \
  --world-size 1 --tp-size 1 \
  --batch-size 16 --mini-bs 4 \
  --lr 2e-5 --max-context-len 4096 \
  --epochs 3 --save-interval 20

Supports typed-decision parquet and request/targets JSONL datasets. Keep training and evaluation data separate. This release omits the original dataset iterator state so fine-tuning starts with a fresh data cursor.

See the AReno JEV documentation for data formats and configuration.