math_with_judge: resources_servers: math_with_judge: entrypoint: app.py # Grading uses math-verify only (should_use_judge: false), so the judge reference just needs to resolve. judge_model_server: type: responses_api_models name: policy_model judge_responses_create_params: { input: [] } should_use_judge: false domain: math verified: false description: DAPO17k train problems repeated for trajectory collection with math-verify value: Collect rollouts on DAPO17k without training (NeMo RL is_trajectory_collection) math_with_judge_simple_agent: responses_api_agents: simple_agent: entrypoint: app.py resources_server: type: resources_servers name: math_with_judge model_server: type: responses_api_models name: policy_model datasets: # NeMo RL trajectory collection reads only the validation split, so the train problems go there. # Own local file: collate checks each dataset against `_metrics.json`, and # dapo17k_train_metrics.json describes dapo17k.yaml's train split. - name: validation type: validation jsonl_fpath: resources_servers/math_with_judge/data/dapo17k_trajectory_collection_train.jsonl # From YouJiacheng/DAPO-Math-17k-dedup gitlab_identifier: dataset_name: dapo17k version: 0.0.1 artifact_fpath: dapo17k_train.jsonl num_repeats: 16 license: Apache 2.0 math_with_judge_environment_server: environment_servers: single_agent_turn_legacy: entrypoint: app.py resources_server: type: resources_servers name: math_with_judge agent_server: type: responses_api_agents name: math_with_judge_simple_agent