Qwen3.5-4B-Judge-1M

Qwen3.5-4B-Judge-1M

Small judge model that returns a calibrated probability on whether a reply is backed by its source, whether text is unsafe, and which task or domain it belongs to. For developers adding checks to AI pipelines.

Made by MaziyarPanahi on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a 4B judge model that answers short check questions about text instead of writing prose. It can say whether a reply is backed by a given source, whether a request or reply is unsafe, and which task or domain a text belongs to. For each option it returns a calibrated probability.

Who it suits: It suits developers who need a small, cheap checker inside an AI pipeline, such as a retrieval chatbot that must catch unsupported claims.

Industry | Who uses it | How they use it | Example
Software | AI app developers | Check replies against retrieved documents | Flagging an answer the source does not support
Trust and safety | Moderation teams | Screen requests and replies for unsafe content | Scoring a message before it reaches a user
Customer support | Support platform teams | Route conversations by domain or intent | Sending a banking question to the right queue

You can skip it if you need free-text answers, medical decisions, or a tool-calling decider.

Where to use it: It runs in Python on your own machine through Unsloth, with the base model from the Qwen team and a LoRA adapter on top.

• Python scripts that use the Unsloth decision model class
• Pipelines that need a yes-or-no style judgment with a probability
• Windows, macOS or Linux, wherever Python and a suitable GPU are available

Use case | Who | What they get back
Claim support check | RAG developers | A probability that every claim is backed by the context
Safety screening | Platform teams | A probability that a request or reply is unsafe
Domain routing | Support teams | A probability for each domain or task option

How to set it up: You need Python, the Hugging Face tool and the Unsloth library. The code loads the model by its repository name with 4096 as the maximum sequence length.

• Python
• The Unsloth library, which provides the decision model class
• Disk and GPU memory for a 4B model

1. Install the Hugging Face download tool.
2. Download the adapter files, or let the script fetch them by name.
3. Run the usage script below with your own text and question.

pip install -U huggingface_hub
hf download MaziyarPanahi/Qwen3.5-4B-Judge-1M

How to use it: You build a question with instructions and criteria, put your text into a state string, and call predict. The result is a probability per option. The model reads at most 4096 tokens.

from unsloth import FastDecisionModel

model, tokenizer = FastDecisionModel.from_pretrained("MaziyarPanahi/Qwen3.5-4B-Judge-1M", max_seq_length=4096, load_in_4bit=False)
FastDecisionModel.for_inference(model)
question = {"decision": {"type": "choice",
    "instructions": "Is every factual claim in `assistant_response` supported by `authoritative_context`?",
    "criteria": {"true": "Every factual claim in the response is supported by the context.",
                 "false": "The response contains at least one unsupported or contradicted claim."}}}
state = '{"authoritative_context": "The clinic opens at 8 a.m. on weekdays and is closed on Sundays.", '\
        '"assistant_response": "The clinic is open every day from 8 a.m."}'
print(FastDecisionModel.predict(model, tokenizer, state, question))

Tool | What it does
FastDecisionModel.from_pretrained | Loads the model and tokenizer
FastDecisionModel.for_inference | Prepares the model for predictions
FastDecisionModel.predict | Returns a probability for each option

Example: You give it a clinic notice that says it opens at 8 a.m. on weekdays and is shut on Sundays, plus a reply claiming it is open every day. With the support question, it returns a probability for true and for false, and the false side should score high.

For newcomers and experts: If you are new, think of it as a referee that reads a text and a question, then reports how sure it is of each answer. Run the example above and read the two probabilities.

If you know the field, it is Qwen3.5-4B with an Unsloth decision head, a LoRA of rank 16, trained on 39,592 rows of source-labelled decisions from OpenDecisionMix-1M. It was calibrated on 400 held-out rows. The card notes that its test and training rows share their origin.

Your text and a question
Decision head reads the state
Calibrated probability per option
You apply your own threshold

Good to know: On a 6,035 row test it matched an earlier judge on RAG plus safety, and the card says to tune thresholds for your own data.

• License: Apache 2.0
• Cost to run: no API key, account or paid service is named, but you need a GPU able to run a 4B model
• Data it reaches: the text you pass in on your machine, plus files downloaded from Hugging Face
• Not for: clinical use or any decision about a patient, deciding when to call a tool, and inputs beyond 4,096 tokens

Title: RAG plus safety accuracy
This model: 89.5%
Earlier Judge: 89.4%
Jev 1.14: 82.2%

On Hugging Face, 13 November 2026

• Library: peft
• Formats: Safetensors
• Downloads in the last 30 days: 29
• Likes: 3
• License: apache-2.0
• Based on: unsloth/Qwen3.5-4B
• Tags: peft, safetensors, unsloth, decision-model, llm-judge, hallucination-detection, safety

Links

• Model: https://huggingface.co/MaziyarPanahi/Qwen3.5-4B-Judge-1M
• Files: https://huggingface.co/MaziyarPanahi/Qwen3.5-4B-Judge-1M/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download MaziyarPanahi/Qwen3.5-4B-Judge-1M

Then follow the usage in its model card. Nothing in this kit runs by itself.