Siddh07ETH/Atlas-Coder-2-0.5B
Siddh07ETH/Atlas-Coder-2-0.5B is a 494 million parameter causal language model developed by Siddharth N.R. under the Pluto AI research project, fine-tuned from Qwen2.5-Coder-0.5B-Instruct. It is specifically optimized for Python code generation, trained exclusively on 50K execution-verified OSS-Instruct samples. This model is designed to be the top-performing sub-1B coding model, excelling in precise and reliable code generation for self-contained Python tasks.
Loading preview...
Atlas-Coder-2-0.5B: A Specialized Sub-1B Code Generation Model
Atlas-Coder-2-0.5B, developed by Siddharth N.R. as part of the Pluto AI research project, is a 494 million parameter model built upon the Qwen2.5-Coder-0.5B-Instruct base. This second-generation Atlas-Coder model is highly specialized for Python code generation, focusing on delivering sharp and reliable outputs with a low hallucination rate for self-contained Python tasks.
Key Differentiators & Capabilities
- Execution-Verified Training Data: Uniquely trained on 50,000 100% execution-verified OSS-Instruct samples, ensuring high-purity data and direct alignment with HumanEval+ and MBPP+ benchmark formats.
- Optimized for Code Generation: By starting from an instruction-tuned base model, Atlas-Coder-2 focuses solely on deepening its Python code generation capabilities, rather than re-learning instruction following.
- Top Sub-1B Performance: Ranked among the top 5 globally for strictly sub-1B parameter models on the EvalPlus (HumanEval+) strict benchmark, demonstrating strong performance relative to its size.
- Efficient Training: Utilizes QLoRA with a reduced rank (r=32) for faster and leaner training, completing within a single Kaggle session.
- Edge Performance: Capable of high-speed inference (75 tokens/second) on devices like an M2 MacBook Air using FP16 GGUF, leveraging Apple Silicon's Metal FP16 vector cores without requiring a dedicated GPU.
Ideal Use Cases
This model is particularly well-suited for:
- Generating precise Python functions for well-defined problems.
- Applications requiring efficient code generation on resource-constrained devices.
- Research and development focused on pushing the boundaries of small-scale code models.
Limitations
Due to its compact size and specialized training, Atlas-Coder-2-0.5B may exhibit limitations on complex multi-file tasks, deeply nested logic, and non-Python languages. Its performance is optimized for single-function completion tasks, and generated code should always be verified.