burtenshaw/GemmaCoder3-12B Overview

This model is a 12 billion parameter variant of the Gemma-3 family, specifically fine-tuned from google/gemma-3-12b-it. Its training utilized the open-r1/codeforces-cots dataset and the TRL framework, indicating a focus on code-related tasks.

Key Capabilities & Performance

The primary differentiator of GemmaCoder3-12B is its enhanced performance in code generation, as evidenced by a significant improvement on the LiveCodeBench benchmark, achieving 32.9% compared to the base Gemma3-12B-it's 21.9%. While excelling in code, the model shows a slight decrease in MMLU performance (61.0% vs 69.5%) but maintains competitive scores on Winogrande (63.9%) and HellaSwag (54.0%).

Use Cases

• Code Generation: Ideal for applications requiring the generation of programming code.
• Competitive Programming: Its fine-tuning on a Codeforces dataset suggests suitability for solving algorithmic problems.
• Code Understanding: Can be leveraged for tasks involving the interpretation and analysis of code snippets.