You can deploy this model and start earning money today!

Views

No views yet

Qwen3-4B-Computer-Science

Name: Qwen3-4B-Computer-Science
Rating: 2.5 (14 reviews)

Qwen3-4B-Computer-Science is a supervised fine-tuned language model based on Qwen/Qwen3-4B, designed for computer science and software engineering tasks.

This repository contains the merged BF16 checkpoint compatible with the Hugging Face Transformers ecosystem.

Model Summary

The model specializes in programming-oriented instruction following across multiple computer science domains, including software engineering, debugging, algorithms, testing, and technical reasoning.

Training was performed using parameter-efficient supervised fine-tuning (LoRA). The released checkpoint contains merged BF16 weights and can be used directly without PEFT adapters.

Motivation

General-purpose language models provide strong performance across many domains but are not specifically optimized for computer science workflows.

Qwen3-4B-Computer-Science aims to improve programming-oriented instruction following while preserving the capabilities of the original Qwen3-4B base model.

Model Details

Field	Value
Model Name	Qwen3-4B-Computer-Science
Base Model	Qwen/Qwen3-4B
Model Type	Causal Language Model
Architecture	Decoder-only Transformer
Parameters	4 Billion
Fine-Tuning	Supervised Fine-Tuning (SFT)
Fine-Tuning Method	LoRA
Training Strategy	Distributed Data Parallel (DDP)
Released Weights	Merged BF16
Framework	Hugging Face Transformers
Primary Language	English

Training

Training was performed using supervised fine-tuning (SFT) with parameter-efficient fine-tuning (LoRA).

Optimization utilized Distributed Data Parallel (DDP). After training, the LoRA adapters were merged into the base model to produce the released BF16 checkpoint.

The published model does not require PEFT adapters during inference.

Training Data

The final training corpus contains 60,989 training examples and 512 evaluation examples.

Dataset	Configuration	License	Train	Eval
HuggingFaceTB/smoltalk	smol-magpie-ultra	Apache-2.0	49,584	416
agentica-org/DeepCoder-Preview-Dataset	primeintellect	MIT	11,405	96

Dataset Attribution

The model was fine-tuned using publicly available datasets released under their respective licenses.

Dataset	Configuration	License
HuggingFaceTB/smoltalk	smol-magpie-ultra	Apache-2.0
agentica-org/DeepCoder-Preview-Dataset	primeintellect	MIT

Credit for the datasets belongs to their respective authors.

Intended Use

Recommended applications include:

Software engineering
Programming assistance
Python development
Code generation
Code explanation
Debugging
Unit testing
Technical documentation
Computer science education

Capabilities

The model has been fine-tuned for:

Programming-oriented instruction following
Code generation
Code completion
Code explanation
Refactoring
Debugging
Algorithm implementation
Standard library usage
Technical reasoning

The model inherits the general instruction-following capabilities of Qwen3-4B.

Installation

pip install -U transformers accelerate torch

Usage

python

from transformers import AutoTokenizer
from transformers import AutoModelForCausalLM

model_name = "Irfanuruchi/Qwen3-4B-Computer-Science"

tokenizer = AutoTokenizer.from_pretrained(model_name)

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

Example

python

messages = [
    {
        "role": "user",
        "content": "Implement binary search in Python."
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

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

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
)

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

Hardware Requirements

This repository contains merged BF16 weights.

Memory requirements depend on the selected precision and inference backend.

Users with limited GPU memory are encouraged to use the GGUF release when available.

Limitations

Although specialized for computer science tasks, the model remains a probabilistic language model.

Outputs should be reviewed before use in production environments.

The model may:

generate incorrect code
hallucinate APIs or libraries
produce incomplete implementations
misunderstand project-specific context

License

This repository is released under the Apache License 2.0.

Base Model

This project is derived from Qwen/Qwen3-4B, which is distributed under the Apache License 2.0.

Training Data

The datasets retain their original licenses.

Dataset	License
HuggingFaceTB/smoltalk	Apache-2.0
agentica-org/DeepCoder-Preview-Dataset	MIT

Acknowledgements

This project builds upon the work of:

Alibaba Qwen Team
Hugging Face
HuggingFaceTB
Agentica
Unsloth

The contributions of these open-source projects made this work possible.

Citation

bibtex

@misc{uruci2026qwen3cs,
  title={Qwen3-4B-Computer-Science},
  author={Irfan Uruçi},
  year={2026},
  publisher={Hugging Face},
  howpublished={https://huggingface.co/Irfanuruchi/Qwen3-4B-Computer-Science}
}

Contact

Questions, bug reports, and suggestions are welcome through the Hugging Face repository discussions.