# Example-dataset agent for the opencode_search resources server. `gym env test # +should_validate_data=true` (nemo_gym/cli/env.py `_validate_data_single`) checks that data/example.jsonl, # data/example_metrics.json and data/example_rollouts.jsonl exist; this file is what collate needs to # rebuild example_metrics.json from example.jsonl: # # gym dataset collate "+config_paths=[resources_servers/opencode_search/configs/opencode_search_example.yaml]" \ # +output_dirpath=data/opencode_search +mode=example_validation # # The committed example_rollouts.jsonl were produced by the benchmark's opencode_sandboxed_search_agent. # This agent composes no harness config and is not meant to be run. # # A SEPARATE FILE from opencode_search.yaml on purpose. Benchmarks compose opencode_search.yaml and # build their own server from it with `_inherit_from`, which MOVES opencode_search_resources_server # (global_config.py pops the source). An example agent living in that file would then point at a # server that no longer exists, and the benchmark fails to resolve -- the same reason # swebench.yaml carries no example agent. # # Also a separate INSTANCE from opencode_sandboxed_search_agent, so the example rows can never ride # along into a run that composes the search agent. OmegaConf replaces lists wholesale: a composition # that restates `datasets` (browsecomp_opencode.yaml does) would silently drop an example entry, and # one that does not would inherit it. config_paths: - resources_servers/opencode_search/configs/opencode_search.yaml opencode_search_example_agent: responses_api_agents: opencode_sandboxed_agent: entrypoint: app.py resources_server: type: resources_servers # The INSTANCE name declared in opencode_search.yaml, not the implementation directory. name: opencode_search_resources_server model_server: type: responses_api_models name: policy_model datasets: - name: example type: example jsonl_fpath: resources_servers/opencode_search/data/example.jsonl opencode_search_example_environment_server: environment_servers: legacy_agent: entrypoint: app.py agent_server: type: responses_api_agents name: opencode_search_example_agent # opencode_search.yaml points the server's judge_model_server at `search_judge_model`, which only the # benchmark (benchmarks/browsecomp/configs/browsecomp_opencode.yaml) defines. Same judge, restated so # this file resolves on its own; keep the two in sync. search_judge_model: responses_api_models: vllm_model: entrypoint: app.py base_url: https://inference-api.nvidia.com/v1 model: nvidia/zai-org/glm-5.2 api_key: ${oc.select:inference_api_key,null} return_token_id_information: false uses_reasoning_parser: true