LemonDecider 4B

LemonDecider 4B

Makes typed decisions from a text state and named questions, returning probabilities for choices, yes or no answers and ordinal scores, with no generated text. A 4.5B model by inclusionAI for developers.

Made by inclusionAI 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: LemonDecider 4B is a decision model from inclusionAI, fine-tuned from Qwen3.5-4B. You give it a text state and named questions. It returns probabilities for choice questions, probabilities for yes or no questions, or scores on an ordered scale, and it writes no answer text. Question lists that are too long get split, and the card says inputs are never truncated.

Who it suits: Developers who want a small model to make routine decisions with probabilities attached, such as routing messages or judging responses.

Industry | Who uses it | How they use it | Example
Customer service | Support engineers | Route requests and classify cases | Choosing a category for an incoming message
Agent safety | Agent builders | Judge whether an action is safe | A yes or no check before an agent acts
Search and commerce | Retrieval engineers | Judge relevance and select products | Scoring how well a result fits a query
Quality review | ML teams | Compare answers for helpfulness and correctness | Rating a response on an ordered scale

If you want a model that writes text or reads images, you can skip it, because it is text-only and returns probabilities.

Where to use it:

• Python with the transformers library, loading the model's own code
• Your own computer on Windows, macOS or Linux, after you download the files
• A text-only HTTP service, which the card says is published

Use case | Who | What they get back
Intent routing | Support teams | A probability for each option you list
Yes or no checks | Agent builders | The probability that the answer is yes
Ordinal ratings | Quality teams | A score on an ordered scale

How to set it up:

• Python 3.12
• Torch and Transformers, with versions near those the card evaluated, and the FLA package for the run notes
• About 9.1 GB of disk space in BF16, or about 2.3 GB at 4-bit (an estimate)

1. Install the Hugging Face hub tools.
2. Download the model files.
3. Load the model by following the card, allowing its remote code to run.

pip install -U huggingface_hub
hf download inclusionAI/LemonDecider

How to use it: Write the state as text and give each question a name and a type. A choice question can have up to 255 options, a yes or no question returns one probability, and a score question returns an ordinal rating. Requests that fit are scored in one joint pass.

Example: A support message says checkout keeps failing. You ask which team should handle it from a list of options, and whether it sounds urgent. You get back a probability for each team and one probability for urgency.

For newcomers and experts: If you are new to this, treat it as a sorting machine that says how sure it is. Start with one question that has three options and read the probabilities.

The model has a Qwen3.5-4B hybrid backbone with a readout over up to 255 options, calibrated with a scalar temperature. Each chunk holds up to 8192 tokens including the template, and longer question lists are split by recursive bisection with the full state kept in every chunk. Training used about 1.2 million decision and question rows, with four checkpoints averaged.

You write a text state
You add named, typed questions
The model scores every option jointly
You read the probabilities or scores

Good to know:

• Input: it is text-only, and the vision part was frozen in training
• Training data: the card warns that source-specific terms of some data still require review

Title: Share of training rows by question format, in percent
Choice questions: 84.4%
Yes or no questions: 15.2%
Score questions: 0.4%

• License: apache-2.0
• Cost to run: no API key or paid service is named, only your own hardware
• Data it reaches: downloading fetches model files, and it reads the text state you send
• Not for: writing text, reading images, or any use that skips the card's note about checking data terms

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: transformers
• Parameters: 4.5B
• Weights: about 9.1 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 988
• Likes: 53
• License: apache-2.0
• Based on: Qwen/Qwen3.5-4B
• Fine-tunes: 1
• Tags: transformers, safetensors, lemondecider_decision, feature-extraction, decision-engine, typed-decision, choice, classification, merge, qwen3.5

Links

• Model: https://huggingface.co/inclusionAI/LemonDecider
• Files: https://huggingface.co/inclusionAI/LemonDecider/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download inclusionAI/LemonDecider

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