Strangefrost/CloneOllama-selfplay-coder-0.5B
Strangefrost/CloneOllama-selfplay-coder-0.5B is a 0.5 billion parameter coding-focused language model, fine-tuned from Qwen2.5-Coder-0.5B. It was trained using a novel self-play reinforcement learning method with four LLM judges and DeepSeek API evaluation. This model excels at generating Python and C# code, algorithms, and computer science theory, demonstrating improved scoring over 500 training rounds. It is optimized for efficient code generation with reported speeds of 540 tokens/s on AMD Radeon RX 9070 XT hardware.
Loading preview...
Overview
Strangefrost/CloneOllama-selfplay-coder-0.5B is a 0.5 billion parameter model derived from Qwen2.5-Coder-0.5B, specifically engineered for code generation. Its unique differentiator is its self-play reinforcement learning training methodology, where the model iteratively improved its code generation capabilities without human labeling.
Key Capabilities & Training
- Self-Play Training: The model was trained over 500 rounds, with 397 accepted improvements, using 4 LLM judges (DeepSeek-Coder-V2, VibeThinker-3B, gemma4-coding, Qwen2.5-Coder-7B) and DeepSeek API evaluation. This process led to an improvement in average scores from 5.3 to 5.5 out of 10.
- Code Generation: Specializes in Python and C# coding, algorithms, and computer science theory, generating code based on prompts.
- Efficiency: Benchmarked on an AMD Radeon RX 9070 XT, it achieves prompt speeds of 619 tokens/s and generation speeds of 540 tokens/s.
- Compact Size: Available in a 506 MB Q8_0 GGUF format, making it suitable for local deployment.
Use Cases
- Code Autocompletion: Assisting developers with Python and C# code snippets.
- Algorithmic Problem Solving: Generating solutions for common algorithms.
- Educational Tools: Providing examples or explanations for computer science concepts.
- Local Development: Its efficient performance and smaller size make it ideal for running on consumer-grade hardware via
llama.cpporOllama.