Juicer: Natural-Language Data Refinement Model¶

Juicer is a data-refinement model built on Qwen3.6-35B-A3B (MoE architecture: 35B total parameters / 3B activated). It turns natural-language cleaning instructions, filtering rules, and semantic-tagging requirements into structured outputs—strict tagged text or canonical JSON.

Juicer is designed for data-refinement workflows and supports local deployment to process sensitive data in your own environment.

Resources¶

Resource	Link
HuggingFace Model	datajuicer/Juicer-35B-A3B
ModelScope Model	Data-Juicer/Juicer-35B-A3B
Juicer Playground	data-juicer-hub/juicer_playground

Quick Start¶

0. Get the launch scripts¶

The juicer_playground directory in data-juicer-hub provides serve.sh, requirements.txt, and app.py. Clone the repository and enter its Playground directory:

git clone https://github.com/datajuicer/data-juicer-hub.git
cd data-juicer-hub/juicer_playground

1. Deploy the model¶

Prepare the inference environment and model weights using the Playground deployment instructions. Juicer can be served as an OpenAI-compatible endpoint. On a single H20 (96 GB):

export MODEL_ID=/path/to/juicer-model
bash serve.sh --model "$MODEL_ID" --port 8000

2. Launch the Playground¶

The Juicer Playground provides an interactive UI to try recipes, browse showcase cases, and compare Juicer against the base model:

pip install -r requirements.txt
export JUICER_BASE_URL=http://localhost:8000/v1
python app.py
# open http://localhost:7860

Evaluation¶

Juicer is evaluated on CDR-Bench, covering atomic mappers/filters, compositional workflows, order-sensitive pipelines, and semantic tasks (PII, hallucination, rubric, safety).

Learn More¶

For full deployment options (vLLM, SGLang, Transformers), showcase cases, integration code, and AB comparison setup, visit the Juicer Playground README.