# overlap_metric_rule options: seq_match, ngram16 (default), lcs # -> which overlap signal is computed and becomes reasoning_overlap # overlap_grading_rule options: base, multiply (default), minus # -> how answer_score and reasoning_overlap combine into the reward: # base : reward = answer_score # multiply : reward = answer_score * (1 - reasoning_overlap) # minus : reward = answer_score - reasoning_overlap # answer_score : F1 of the final answer's node list vs expected_answer # reasoning_overlap : overlap between the reasoning and the input message (per overlap_metric_rule) lc_niah: resources_servers: lc_niah: entrypoint: app.py overlap_metric_rule: ngram16 overlap_grading_rule: multiply domain: knowledge verified: false description: Answer-correctness reward gated by low reasoning/input overlap value: Reward correct answers while discouraging copying the prompt into reasoning lc_niah_simple_agent: responses_api_agents: simple_agent: entrypoint: app.py resources_server: type: resources_servers name: lc_niah model_server: type: responses_api_models name: policy_model datasets: - name: example type: example jsonl_fpath: resources_servers/lc_niah/data/example.jsonl lc_niah_environment_server: environment_servers: single_agent_turn_legacy: entrypoint: app.py resources_server: type: resources_servers name: lc_niah agent_server: type: responses_api_agents name: lc_niah_simple_agent resources_tool_transports: - direct_http