SOULAMA/qwen2.5-coder-ft

Overview

This model, developed by Soulama Haicanama Ismael, is a fine-tuned and merged version of Qwen2.5-Coder-1.5B-Instruct. It is specifically optimized for Python programming and precise code generation, acting as a strict code assistant. The model was fine-tuned using Low-Rank Adaptation (LoRA) and subsequently merged into full 16-bit precision weights, focusing its attention layers on script structures and programmatic vocabulary.

Key Capabilities

• Specialized Code Generation: Excels at generating Python code and answering programming questions.
• Optimized for Code: Designed to deliver accurate programming solutions with minimal conversational overhead.
• Deterministic Output: Recommended to use low generation temperature ($\le 0.2$) and clear system instructions for deterministic code results.
• Qwen ChatML Format: Utilizes the Qwen ChatML format for structured prompts, ensuring semantic alignment with the original instruct template.

Good for

• Direct Code Generation: Ideal for tasks requiring the generation of Python code snippets or functions.
• Programming Assistance: Suitable for use as a code assistant within a Chat Template infrastructure, especially for isolating Python code blocks.

Limitations

Due to its 1.5 billion parameter size, the model can exhibit context-loop repetition if stopping criteria are not explicitly configured during inference. Users must strictly handle stop tokens (<|im_end|>) to ensure stable execution. It is not intended for generic non-coding tasks like creative writing or general chat.