dharun2049/kaveri-hgrpo-0.5b

Kaveri H-GRPO 0.5B: Experimental Coding Model

Kaveri H-GRPO 0.5B is an experimental 0.5 billion parameter model developed by dharun2049, specifically designed for coding tasks. It is built upon the Qwen/Qwen2-0.5B-Instruct base model and utilizes a unique post-training approach.

Key Differentiator: Hypergraph Group Relative Policy Optimization (H-GRPO)

This model's core innovation lies in its Hypergraph Group Relative Policy Optimization (H-GRPO) reward-shaping layer. This experimental technique enhances the standard GRPO trainer by modeling candidate solutions as hypergraph vertices, where hyperedges encode shared algorithmic strategies, partial test behavior, and code-structure similarity. A compilation gate prevents uncompilable candidates from inheriting correctness rewards through the hypergraph, aiming to improve the quality of generated code.

Training Details

• Base Model: Qwen/Qwen2-0.5B-Instruct
• Datasets: open-r1/codeforces, verifiable-prompts
• Target Language: C++17
• Trainer: TRL GRPOTrainer with H-GRPO alpha of 0.2
• LoRA Configuration: Rank 8, Alpha 16, targeting q_proj, v_proj
• Verification: Local g++ compilation and official test cases.

Performance

The model achieves 43.92% on the MMLU benchmark, indicating its general language understanding capabilities alongside its specialized coding focus.

Use Cases

This model is particularly suited for:

• Experimental C++17 code generation: Leveraging its H-GRPO training for potentially more robust solutions.
• Research into reward-shaping techniques: For developers interested in advanced policy optimization methods for code generation.