We lifted 13 functions out of this paper's own repositories and ran 7 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.
| Repository | Role | Ran |
|---|---|---|
| lzhxmu/cppo | canonical | 7 of 13 |
| Function | Status | Where it lives |
|---|---|---|
| extract_answer_from_model_output | Ran | lzhxmu/cppo/cppo_verl/recipe/cppo/src/gsm8k_compute_score.py code served (permissive licence) · get_code("6859aa3f82a38002") |
| extract_last_number | Ran | lzhxmu/cppo/cppo_verl/recipe/cppo/src/gsm8k_compute_score.py code served (permissive licence) · get_code("0014331dab98c0a5") |
| extract_single_number | Ran | lzhxmu/cppo/cppo_verl/recipe/cppo/src/gsm8k_compute_score.py code served (permissive licence) · get_code("6036741237b33141") |
| get_custom_reward_fn | Ran | lzhxmu/cppo/cppo_verl/recipe/cppo/src/main_cppo.py code served (permissive licence) · get_code("34be00218891a21a") |
| get_huggingface_actor_config | Ran | lzhxmu/cppo/cppo_verl/verl/utils/model.py code served (permissive licence) · get_code("fe79c459a3e4cc64") |
| union_tensor_dict | Ran | lzhxmu/cppo/cppo_verl/verl/protocol.py code served (permissive licence) · get_code("21331a58f93375e4") |
| unpad_dataproto | Ran | lzhxmu/cppo/cppo_verl/verl/protocol.py code served (permissive licence) · get_code("25f0ea3f460f6ce1") |
| extract_solution | Not yet run | lzhxmu/cppo/cppo_verl/recipe/cppo/precess_gsm8k.py code served (permissive licence) · get_code("e4128a6f1b0585bc") |
| get_generation_config | Not yet run | lzhxmu/cppo/cppo_verl/verl/utils/model.py code served (permissive licence) · get_code("660713a8fa047736") |
| get_weight_loader | Not yet run | lzhxmu/cppo/cppo_verl/verl/models/weight_loader_registry.py code served (permissive licence) · get_code("39fba8c5b01d595e") |
| get_weight_saver | Not yet run | lzhxmu/cppo/cppo_verl/verl/models/weight_loader_registry.py code served (permissive licence) · get_code("8bf1b3e9bf87072c") |
| load_state_dict_to_megatron_gptmodel | Not yet run | lzhxmu/cppo/cppo_verl/verl/models/mcore/loader.py code served (permissive licence) · get_code("1901b047ab8e1654") |
| squeeze | Not yet run | lzhxmu/cppo/cppo_verl/verl/utils/model.py code served (permissive licence) · get_code("3b15e2ac7497c441") |
Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.
This paper introduces Completion Pruning Policy Optimization (CPPO) to accelerate the training of reasoning models based on Group Relative Policy Optimization (GRPO). GRPO, while effective, incurs high training costs due to the need to sample multiple completions for each question. Our experiment and theoretical analysis reveal that the number of completions impacts model accuracy yet increases training time multiplicatively, and not all completions contribute equally to policy training -- their contribution depends on their relative advantage. To address these issues, we propose CPPO, which prunes completions with low absolute advantages, significantly reducing the number needed for gradient calculation and updates. Additionally, we introduce a dynamic completion allocation strategy to maximize GPU utilization by incorporating additional questions, further enhancing training efficiency. Experiments show that CPPO achieves up to $7.98\times$ speedup on GSM8K and $3.48\times$ on Math while preserving or even enhancing the accuracy compared to the original GRPO. We release our code at \href{https://github.com/lzhxmu/CPPO}{https://github.com/lzhxmu/CPPO}.
The same record, over MCP at https://syntology.ai/mcp:
get_harvested_code_for_paper("2503.22342")
get_code_for_paper("2503.22342")
have("2503.22342")
Connect an agent — have() is free.