Kaveri H-GRPO 0.5B
Kaveri H-GRPO 0.5B is an experimental coding post-training checkpoint
derived from Qwen/Qwen2-0.5B-Instruct.
It was trained with LoRA using TRL's stock GRPO trainer plus an experimental Hypergraph Group Relative Policy Optimization (H-GRPO) reward-shaping layer.
H-GRPO
For each rollout group, candidate solutions form hypergraph vertices. Hyperedges encode shared algorithmic strategy, partial official-test behavior, and lightweight code-structure similarity.
P_H = D_v^(-1/2) H W D_e^(-1) H^T D_v^(-1/2)
r_H = r + alpha * compile_gate * (P_H @ r - r)
The compilation gate prevents uncompilable candidates from inheriting correctness reward through the hypergraph.
Training setup
- Base model:
Qwen/Qwen2-0.5B-Instruct - Dataset:
open-r1/codeforces,verifiable-prompts - Language: C++17
- Trainer: TRL
GRPOTrainer - H-GRPO alpha:
0.2 - Generations per prompt:
4 - Generation batch size:
16 - Optimizer steps:
150 - Max completion length:
256 - LoRA rank:
8 - LoRA alpha:
16 - LoRA targets:
q_proj, v_proj - Verifier: local
g++compile + official tests
Files
The repository root contains the merged standalone model. The original
LoRA adapter is preserved in lora_adapter/.
Status
MMLU - 43.92%
