Overview

kiro-1.0-7B-XCode is the inaugural model in the kiro series developed by constructai. This 7.6 billion parameter model is built upon Qwen2.5-Coder-7B and is specifically fine-tuned for code-related tasks, supporting a 32,768 token context length.

Key Capabilities

• Code Generation: Proficient in generating Python and JavaScript code.
• Code Analysis: Capable of analyzing and explaining existing code snippets.
• Instruction Following: Trained to follow ### Instruction → ### Response format, making it suitable for interactive coding tools.

Training Details

The model was trained using QLoRA (4-bit, NF4) on a mixed dataset of approximately 58,000 samples. The dataset included real-world Python and JavaScript code from bigcode/the-stack, alongside Python instruction-response pairs from iamtarun/python_code_instructions_18k_alpaca.

Ideal Use Cases

• IDE Plugins: Enhancing integrated development environments with AI-powered coding assistance.
• Coding Assistants: Providing intelligent suggestions and code completion.
• Code Review Tools: Aiding in the review process by analyzing code quality and identifying potential issues.

Limitations

As a v1.0 release, the model was trained for a single epoch, which may lead to repetitions in longer outputs (mitigated with repetition_penalty). While optimized for Python and JavaScript, its support for other programming languages is currently limited.