A fine-tuned version of Qwen/Qwen3.5-4B-Base trained on facebook/natural_reasoning to improve reasoning ability, structured problem solving, and instruction-following performance.

This model was created through supervised fine-tuning (SFT) on reasoning-focused examples, allowing the base model to better understand complex questions, generate more coherent explanations, and follow reasoning patterns more effectively.

Model Overview

Attribute	Details
Base Model	Qwen/Qwen3.5-4B-Base
Fine-tuning Dataset	facebook/natural_reasoning
Training Method	Supervised Fine-Tuning (SFT)
Training Epochs	3
Parameters	~4 Billion
Architecture	Transformer Decoder-only Causal Language Model
Framework	Hugging Face Transformers
Task	Reasoning-oriented text generation
Language	English

Intended Use

This model is intended for:

Reasoning-based question answering
General-purpose AI assistants
Educational applications
Research experiments
Agentic AI systems
Local LLM deployment
Natural language problem solving

Example use cases:

Answering complex questions
Explaining concepts step-by-step
Solving analytical problems
Building reasoning agents
Research and experimentation with language models

Model Capabilities

Improved Reasoning

The fine-tuning process improves the model's ability to:

Break down complex problems
Follow logical reasoning patterns
Produce structured answers
Explain conclusions clearly
Handle multi-step questions

Instruction Following

Compared with the original base model, this version is optimized for:

Better understanding of user instructions
More helpful responses
Improved response formatting
More consistent explanations

Problem Solving

The model is suitable for tasks involving:

Logical reasoning
Mathematical thinking
Analytical questions
Conceptual explanations
Multi-step problem solving

Training Dataset

facebook/natural_reasoning

This model was trained using the facebook/natural_reasoning dataset, which contains reasoning-oriented examples designed to improve language model reasoning capabilities and encourage structured problem solving.

Training Details

The model was fine-tuned using supervised fine-tuning (SFT).

Config	Value
Base Model	Qwen/Qwen3.5-4B-Base
Dataset	facebook/natural_reasoning
Training Epochs	3

The model learns from reasoning examples by optimizing the causal language modeling objective, adapting the original Qwen3.5-4B-Base weights toward stronger reasoning behavior.

Usage

Transformers

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "YOUR_USERNAME/YOUR_MODEL_NAME"

tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto"
)

prompt = """
Explain why the Earth has seasons.
"""

inputs = tokenizer(
    prompt,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True
)

response = tokenizer.decode(
    outputs[0],
    skip_special_tokens=True
)

print(response)

Recommended Prompting

Use clear natural language instructions. The model performs best when the task and expected output format are clearly described.

Example:

Question:
A car travels 240 kilometers in 3 hours.
What is its average speed?

Answer:

Limitations

Although this model has improved reasoning capabilities, it still has limitations:

It can generate incorrect or incomplete answers.
It does not have real-time access to external information.
Complex reasoning problems may still contain mistakes.
Outputs should be verified for critical applications.
The model may reflect biases present in the training data.

This model should be used as an assistant, not as a replacement for expert judgment.

Deployment

Recommended inference frameworks:

Hugging Face Transformers
vLLM
llama.cpp (after conversion)
TensorRT-LLM (after conversion)

Recommended hardware:

Hardware	Recommendation
GPU	NVIDIA CUDA GPU
VRAM	8GB+ recommended
System RAM	16GB+ recommended

Evaluation

This release focuses on improving reasoning behavior through targeted supervised fine-tuning.

Recommended evaluation benchmarks:

GSM8K
MMLU
ARC
BIG-Bench reasoning tasks
Custom reasoning evaluations

Acknowledgements

Special thanks to:

The Qwen Team for releasing the Qwen3.5 family of models.
Facebook AI Research for providing the Natural Reasoning dataset.
Hugging Face for providing the open-source ML ecosystem.

Citation

If you use this model in your research, projects, or applications, please cite:

bibtex

@misc{koohi2026qwen35-natural-reasoning,
  author    = {Reza Koohi},
  title     = {Qwen3.5-4B Natural Reasoning Fine-Tuned Model},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/YOUR_USERNAME/YOUR_MODEL_NAME}
}

Model Card Authors

Created by: Reza Koohi Hugging Face: https://huggingface.co/YOUR_USERNAME

License

This model follows the license terms of the original base model: Qwen/Qwen3.5-4B-Base.

Please review the original model license before commercial deployment.

Note: Replace YOUR_USERNAME and YOUR_MODEL_NAME with your actual Hugging Face repository name before uploading. If you actually used LoRA/QLoRA instead of a full fine-tune, update the "Training Method" and "Training Details" sections accordingly.

Rzkoohi/Qwen3.5-4B-Natural_reasoning

Qwen3.5-4B Natural Reasoning Fine-Tuned Model