JetBrains/CodeLlama-7B-Kexer: Kotlin Code Generation

JetBrains/CodeLlama-7B-Kexer is a specialized 7 billion parameter generative text model, fine-tuned by JetBrains from the CodeLlama-7B base model. Its primary distinction lies in its optimization for the Kotlin programming language, achieved through fine-tuning on the Kotlin Exercises dataset.

Key Capabilities

• Kotlin Code Generation: Excels at generating Kotlin code, particularly for programming exercises and functions.
• Fill-in-the-Middle (FIM): Supports FIM capabilities using a specific token format (<PRE> prefix <SUF> suffix <MID>).
• Enhanced Kotlin Performance: Achieves a 42.24% Pass Rate on the Kotlin HumanEval dataset, significantly outperforming the base CodeLlama-7B model (26.89%).

Training and Data

The model was fine-tuned using 15,000 examples from the synthetically generated Kotlin Exercises dataset, totaling approximately 3.5 million tokens. Training was conducted on a single A100 GPU over 4 epochs with a total batch size of 256.

Good for

• Kotlin Developers: Ideal for developers seeking assistance with Kotlin code generation, completion, and problem-solving.
• Educational Tools: Suitable for integration into platforms or tools focused on teaching or practicing Kotlin programming.
• Code Assistants: Can serve as a backend for intelligent code assistants specifically tailored for Kotlin environments.