Juicer¶

Natural-language data-refinement model · Built on Qwen3.6-35B-A3B · Locally deployable

Source: data-juicer-hub/juicer_playground

Juicer turns cleaning instructions, filtering rules, and semantic-tagging requirements into strict tagged text or canonical JSON. It is not a general chat model; it is designed for data-refinement workflows.

Evaluated on CDR-Bench.

Highlights¶

Natural-language recipe execution — for example, remove emails, deduplicate sentences, and normalize whitespace.
Order-sensitive refinement — distinguishes filtering before cleanup from filtering after cleanup and tracks intermediate state.
Structured semantic tagging — rubric and related tasks can return JSON under a defined schema.
Local deployment — supports processing sensitive data in your own environment.

Quickstart¶

1. Start the model¶

Juicer is not tied to a specific inference framework. We recommend deploying it as an OpenAI-compatible service for use by the Playground and data-processing jobs.

Recommended: vLLM service¶

On a single H20 (96GB), use the included launcher:

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

Optional: SGLang service¶

Install SGLang for your accelerator environment by following the official installation guide, then launch:

export MODEL_ID=/path/to/juicer-model
python -m sglang.launch_server \
  --model-path "$MODEL_ID" \
  --served-model-name juicer \
  --port 8000 \
  --trust-remote-code

Optional: local Transformers loading¶

This approach calls the model directly from Python without starting a service:

pip install "transformers>=5.3" accelerate
export MODEL_ID=/path/to/juicer-model

import os
from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = os.environ["MODEL_ID"]
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "Remove email addresses from: Contact ops@example.com."}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
    return_dict=True,
).to(model.device)
prompt_length = inputs["input_ids"].shape[-1]
outputs = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
print(tokenizer.decode(outputs[0, prompt_length:], skip_special_tokens=True))

2. Open the Playground¶

Juicer Playground homepage

The Playground can connect to any OpenAI-compatible model service:

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

The Playground is a single-page app with no build step and four tabs:

Tab	Purpose
Try Juicer	Load a random case or run custom text and a recipe
Showcase Gallery	Browse 51 examples, inspect details, or load one into the try page
Capability Showcase	Explore core capabilities through representative examples
AB Comparison	(Optional) Compare the base model and Juicer side by side

AB Comparison (optional)¶

If multiple GPUs are available, you can run Juicer and the base model at the same time to experience their behavior on the same recipe. Adjust GPU allocation through the launcher options to suit the machine.

bash serve_ab.sh \
  --juicer /path/to/juicer-model \
  --raw /path/to/Qwen3.6-35B-A3B

export JUICER_RAW_BASE_URL=http://localhost:8001/v1
python app.py

bash serve_ab.sh --stop

See tutorial.md for detailed steps.

Capabilities¶

Type	Meaning	Example
Atomic	Single-step mapper or filter	Remove URLs; keep only English text
Compositional	Multiple refinement steps in one pass	Remove emails, deduplicate, and normalize whitespace
Order-sensitive	Execution order that respects intermediate state	Filter before cleanup versus after cleanup
Semantic	PII redaction, rubric scoring, and safety tagging	Redact identifiers; return structured scores

CDR-Bench and the showcase cases focus on these core operators:

text mappers: link cleanup, comment/reference removal, duplicate-sentence removal, and whitespace and formatting normalization
filters: text/word length, repetition ratio, stopword ratio, and related quality filters
PII operators: contact, location, temporal, identity, and person-related redaction
semantic operators: hallucination detection/correction, rubric scoring, and classification tagging

Showcase Cases¶

The cases/ directory contains 51 showcase cases grouped by CDR-Bench capability:

Category	File	Count	Content
Atomic Mapper	`atomic_mapper.jsonl`	6	Single-step text transforms
Atomic Filter	`atomic_filter.jsonl`	5	Single-step filtering decisions
PII Redaction	`pii.jsonl`	10	Name/contact/address/ID/temporal redaction
Hallucination	`hallucination.jsonl`	5	Hallucination detection and correction
Rubric Scoring	`rubric.jsonl`	4	HelpSteer2 dimension scoring
Safety	`safety.jsonl`	6	Aegis safety classification
Order-sensitive	`order_sensitive.jsonl`	8	Ordered multi-operator execution
Compositional	`compositional.jsonl`	7	Multi-step end-to-end workflows

All cases share a unified schema; see cases/README.md.

Integration Code¶

python examples/single_clean.py   # single request
python examples/batch_clean.py    # batch cleaning
python examples/pii_redact.py     # PII redaction

adapter.py provides APIs such as build_prompt_from_row(row) and parse_output(text, output_format). See examples/README.md.

Model Overview¶

Item	Description
Base model	Qwen/Qwen3.6-35B-A3B
Architecture	Qwen3.6 MoE causal LM; 35B total / 3B activated parameters
Output contracts	Tagged text or task-specific canonical JSON
Evaluated serving length	32,768 tokens
Inherited context configuration	262,144 tokens; quality beyond 32K was not evaluated

When supported by the serving framework, disable thinking with chat_template_kwargs.enable_thinking=false.

Evaluation¶

Juicer results on CDR-Bench

Limitations¶

Data-refinement quality beyond 32K tokens has not been evaluated.
Juicer should not directly replace deterministic tools or human review in high-precision or compliance-critical workflows.

Citation¶

@misc{juicer2026,
  title        = {Juicer: Natural-Language Data Refinement with Qwen3.6-35B-A3B},
  author       = {Juicer Contributors},
  year         = {2026},
  howpublished = {Hugging Face model release}
}

Juicer builds on Qwen3.6-35B-A3B and uses CDR-Bench, Data-Juicer, and Trinity-RFT.

License¶

Apache License 2.0, inherited from the Qwen3.6-35B-A3B base model.