--- license: other license_name: qwen-research license_link: LICENSE base_model: huihui-ai/Qwen2.5-3B-Instruct-abliterated-SFT base_model_relation: finetune language: - en library_name: transformers pipeline_tag: text-generation tags: - python - code-generation - lora --- # Python Coder 3B Research **Built with Qwen.** This 3.09B-parameter model completes short Python functions from a function signature and a docstring. It is prepared for noncommercial research and evaluation under the [Qwen Research License](LICENSE). ## Training This model started from [huihui-ai/Qwen2.5-3B-Instruct-abliterated-SFT](https://huggingface.co/huihui-ai/Qwen2.5-3B-Instruct-abliterated-SFT) at revision e21343fbcabbb6225d1cb9122061404f2d21f1ad, derived from [Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct). The Python fine-tune used 120 LoRA steps on 165 synthetic training rows. The LoRA updates were merged into the BF16 weights. The four SafeTensors shards contain the modified model weights. See [NOTICE](NOTICE) for attribution and file changes. ## Evaluation A prompt passed when its generated function passed every test input. These were small, synthetic Python suites scored using a restricted subset of the language. | Suite | Source checkpoint | This model | | --- | ---: | ---: | | Familiar operation patterns | 124/144 | 125/144 | | Original combination challenge | 24/24 | 24/24 | | Fresh task families | 27/40 | 24/40 | The 60% fresh-task score applies to short functions in this evaluation. The benchmark did not measure repository-level bug fixing, other programming languages, or refusal behavior. Detailed results are in [evaluation.json](evaluation.json). ## Inference Load this repository with Transformers using AutoTokenizer and AutoModelForCausalLM. Use the included chat template. To match the benchmark format, give the signature and docstring in the user message and place the same signature and docstring at the start of the assistant completion. The model then generates the function body. Review and test generated code before using it.