We lifted 23 functions out of this paper's own repositories and ran 15 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 |
|---|---|---|
| Unakar/Logic-RL | canonical | 10 of 16 |
| OpenLLMAI/OpenLLaMA2 | pwc_unofficial | 5 of 7 |
| Function | Status | Where it lives |
|---|---|---|
| aggregate_loss | Ran | OpenLLMAI/OpenLLaMA2/openrlhf/models/loss.py code served (permissive licence) · get_code("0edeb138e07fbecc") |
| compute_approx_kl | Ran | OpenLLMAI/OpenLLaMA2/openrlhf/models/utils.py code served (permissive licence) · get_code("c146f3cb0e5259a0") |
| compute_reward | Ran | OpenLLMAI/OpenLLaMA2/openrlhf/models/utils.py code served (permissive licence) · get_code("812f2e3f5cda7c93") |
| create_huggingface_actor | Ran | Unakar/Logic-RL/verl/utils/model.py code served (permissive licence) · get_code("360c671d6041ba3b") |
| extract_solution | Ran | Unakar/Logic-RL/eval_kk/compute_score.py code served (permissive licence) · get_code("b870fd9007fabd94") |
| extract_step | Ran | Unakar/Logic-RL/verl/trainer/fsdp_sft_trainer.py code served (permissive licence) · get_code("216e28c040173a61") |
| get_loss_batch_info | Ran | OpenLLMAI/OpenLLaMA2/openrlhf/utils/loss_utils.py code served (permissive licence) · get_code("557aef158c28bb84") |
| is_transformers_version_in_range | Ran | Unakar/Logic-RL/verl/models/transformers/monkey_patch.py code served (permissive licence) · get_code("42a7954a8cfa89e0") |
| load_jsonl | Ran | Unakar/Logic-RL/eval_kk/main_eval_instruct.py code served (permissive licence) · get_code("0998b08e39672d66") |
| log_probs_from_logits | Ran | OpenLLMAI/OpenLLaMA2/openrlhf/models/utils.py code served (permissive licence) · get_code("75afa2affd2a8e36") |
| parse_cot_eval | Ran | Unakar/Logic-RL/eval_kk/kk_processor.py code served (permissive licence) · get_code("f4e717416b80af20") |
| parse_model_answer | Ran | Unakar/Logic-RL/eval_kk/compute_score.py code served (permissive licence) · get_code("df0cff7e9a4ca96e") |
| parse_solution_text_format | Ran | Unakar/Logic-RL/eval_kk/compute_score.py code served (permissive licence) · get_code("a6a2e23358e05581") |
| union_tensor_dict | Ran | Unakar/Logic-RL/verl/protocol.py code served (permissive licence) · get_code("21331a58f93375e4") |
| unpad_dataproto | Ran | Unakar/Logic-RL/verl/protocol.py code served (permissive licence) · get_code("25f0ea3f460f6ce1") |
| apply_monkey_patch | Not yet run | Unakar/Logic-RL/verl/models/transformers/monkey_patch.py code served (permissive licence) · get_code("1fb4adb621eefddc") |
| get_huggingface_actor_config | Not yet run | Unakar/Logic-RL/verl/utils/model.py code served (permissive licence) · get_code("19472d4343d0a082") |
| get_tensor_in_current_ring_attn_rank | Not yet run | OpenLLMAI/OpenLLaMA2/openrlhf/models/ring_attn_utils.py code served (permissive licence) · get_code("5110d3bb195169a8") |
| get_weight_loader | Not yet run | Unakar/Logic-RL/verl/models/weight_loader_registry.py code served (permissive licence) · get_code("919cf310ad4c189a") |
| load_eval_records | Not yet run | Unakar/Logic-RL/eval_kk/main_eval_instruct.py code served (permissive licence) · get_code("cbfcdbb8e1ddcaca") |
| num_tokens_from_string | Not yet run | Unakar/Logic-RL/eval_kk/kk_processor.py code served (permissive licence) · get_code("26b075196c1c10bf") |
| reset_ring_attn_position_ids | Not yet run | OpenLLMAI/OpenLLaMA2/openrlhf/models/ring_attn_utils.py code served (permissive licence) · get_code("e93117d99aa6b210") |
| squeeze | Not yet run | Unakar/Logic-RL/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.
Inspired by the success of DeepSeek-R1, we explore the potential of rule-based reinforcement learning (RL) in large reasoning models. To analyze reasoning dynamics, we use synthetic logic puzzles as training data due to their controllable complexity and straightforward answer verification. We make some key technical contributions that lead to effective and stable RL training: a system prompt that emphasizes the thinking and answering process, a stringent format reward function that penalizes outputs for taking shortcuts, and a straightforward training recipe that achieves stable convergence. Our 7B model develops advanced reasoning skills-such as reflection, verification, and summarization-that are absent from the logic corpus. Remarkably, after training on just 5K logic problems, it demonstrates generalization abilities to the challenging math benchmarks AIME and AMC.
The same record, over MCP at https://syntology.ai/mcp:
get_harvested_code_for_paper("2502.14768")
get_code_for_paper("2502.14768")
have("2502.14768")
Connect an agent — have() is free.