# RoleMRC LLM-as-judge scorer (5 aspects: knowledge range, style, nested/ # multi-turn instruction, instruction priority). # # The judge runs on a DEDICATED model server instance named `judge_model`, # wired by the `${judge_*}` overrides, so it is never the model under test — # pointing the judge at the policy model makes it grade its own output. Supply # the judge endpoint alongside the policy one, e.g.: # # gym env start \ # "+config_paths=[resources_servers/rolemrc/configs/rolemrc_judge.yaml,\ # responses_api_models/vllm_model/configs/vllm_model.yaml]" \ # +policy_base_url=... +policy_api_key=... +policy_model_name=... \ # +judge_base_url=https://api.openai.com/v1/ +judge_api_key=$OPEN_AI_KEY \ # +judge_model_name=gpt-4.1 # # To reuse the policy model as the judge anyway (cheap smoke tests, no second # endpoint available), set judge_model_server.name back to `policy_model`. rolemrc_judge: resources_servers: rolemrc: entrypoint: app.py domain: rlhf verified: false description: RoleMRC role-play MRC scored by a 5-aspect LLM-as-judge (knowledge range, style, nested/multi-turn instruction, instruction priority). value: Improve role-play instruction following under knowledge-range and instruction-priority constraints. mode: judge # Dedicated judge instance, NOT the model under test — see the header. judge_model_server: type: responses_api_models name: judge_model judge_api: chat_completions judge_chat_completion_create_params: messages: [] n: 1 temperature: 0.0 top_p: 1.0 max_tokens: 2048 judge_endpoint_max_concurrency: 64 # Dedicated judge model instance referenced above. judge_model: responses_api_models: openai_model: entrypoint: app.py openai_base_url: ${judge_base_url} openai_api_key: ${judge_api_key} openai_model: ${judge_model_name} openai_default_headers: {} rolemrc_judge_simple_agent: responses_api_agents: simple_agent: entrypoint: app.py resources_server: type: resources_servers name: rolemrc_judge model_server: type: responses_api_models name: policy_model datasets: - name: example type: example jsonl_fpath: resources_servers/rolemrc/data/example_judge.jsonl license: MIT # Build with `python resources_servers/rolemrc/prepare_rolemrc.py`. - name: test type: validation jsonl_fpath: resources_servers/rolemrc/data/test_judge.jsonl huggingface_identifier: repo_id: Junrulu/RoleMRC artifact_fpath: roleMRC_test.jsonl license: MIT # Null defaults so OmegaConf resolves ${judge_*} without CLI overrides (needed # for config listing / `gym list benchmarks`). Recommended judge: gpt-4.1. # policy_* are intentionally NOT defaulted here: the policy model is supplied by # whichever model-server config is composed in. judge_base_url: null judge_api_key: null judge_model_name: null rolemrc_judge_environment_server: environment_servers: single_agent_turn_legacy: entrypoint: app.py resources_server: type: resources_servers name: rolemrc_judge agent_server: type: responses_api_agents name: rolemrc_judge_simple_agent resources_tool_transports: - direct_http