PyCoder

PyCoder is a Qwen2.5-0.5B-Instruct fine-tune for Python code generation. This repository includes the merged model in Q8_0 GGUF format, an Ollama Modelfile, a small Ollama API client, and the source scripts used for training.

Use with Ollama

The quickest option is to install Ollama and run:

ollama run hf.co/Safeeq/pycoder-gguf:Q8_0

To download the files and create a local Ollama model instead:

hf download Safeeq/pycoder-gguf --local-dir pycoder-gguf
cd pycoder-gguf
ollama create pycoder -f Modelfile
ollama run pycoder

The hf command is provided by huggingface_hub (pip install huggingface_hub). Alternatively, download pycoder-q8_0.gguf and Modelfile from the Files and versions tab, keep them in the same directory, and run the ollama create command there.

Use the Python helper

The included run_pycoder.py sends a prompt to a local Ollama server. First create or run the pycoder model using the steps above, then run:

python run_pycoder.py "Write a function that reverses a string"

The helper uses only the Python standard library. Ollama must be installed and running locally.

Model behavior and limitations

Intended for generating Python code. For unrelated requests it is trained to return # ERROR: Only Python code requests are supported.
This behavior is learned, not a security boundary; prompts can still produce other content.
This is a small 0.5B-parameter model. Outputs can be incorrect, incomplete, or invalid Python, and no formal benchmark results are provided here.
Review and test all generated code. Do not run untrusted output directly; use an isolated environment when executing generated code.

Training

The model was fine-tuned from Qwen/Qwen2.5-0.5B-Instruct with LoRA. make_dataset.py prepares examples based on iamtarun/python_code_instructions_18k_alpaca plus off-topic refusal examples. The training scripts and dependencies are included for reference; training requires a compatible environment and hardware. The generated data.jsonl, intermediate checkpoints, and merged full-precision model are not included.

Files

pycoder-q8_0.gguf: quantized model weights.
Modelfile: Ollama configuration for the local GGUF.
run_pycoder.py: small client for the local Ollama API.
common.py, make_dataset.py, train.py, merge.py, test_model.py: training and testing source.
requirements.txt: dependencies for the training scripts and upload helper.
LICENSE: license terms.

License

The files in this repository are provided under the Apache License 2.0. The model is based on Qwen/Qwen2.5-0.5B-Instruct; review the upstream model and dataset terms when redistributing or using their materials.

Safeeq/pycoder-gguf