indian_banking: resources_servers: indian_banking: entrypoint: app.py allowed_agents: [gymnasium_agent] domain: agent verified: false description: >- Indian retail-banking customer-support environment: a multi-turn dialog with an LLM-simulated customer over 33 deterministic banking tools (accounts, deposits, loans, cards, mandates, service requests), scored with partial credit against gold tool calls, final DB state, required disclosures and an NL-assertion judge. value: >- Improve stateful multi-turn tool use, policy compliance and customer communication in a regulated domain. # User simulator and judge. Point both at a model server that is NOT the # policy (e.g. the vllm_model / openai_model instance below) so the policy # never judges itself. user_sim_model_server: type: responses_api_models name: user_simulator_model user_sim_responses_create_params: input: [] temperature: 0.0 max_output_tokens: 512 judge_model_server: type: responses_api_models name: user_simulator_model judge_responses_create_params: input: [] temperature: 0.0 max_output_tokens: 1024 should_use_judge: true max_user_turns: 8 max_tool_rounds: 12 indian_banking_agent: responses_api_agents: gymnasium_agent: entrypoint: app.py resources_server: type: resources_servers name: indian_banking model_server: type: responses_api_models name: policy_model # Keep >= ~25: dialogs run 22-25 steps (tool rounds + user turns combined). max_steps: 25 # db.json and kb.json — the synthetic customer DB and knowledge base the # server loads at startup — live in the same Hugging Face dataset repo as # the splits; download them into resources_servers/indian_banking/data/ # before serving (see the README's Data section). datasets: - name: train type: train jsonl_fpath: resources_servers/indian_banking/data/train.jsonl source: type: huggingface repo_id: NPCI/nemo-gym-indian-banking artifact_fpath: train.jsonl license: Apache 2.0 - name: validation type: validation jsonl_fpath: resources_servers/indian_banking/data/validation.jsonl source: type: huggingface repo_id: NPCI/nemo-gym-indian-banking artifact_fpath: validation.jsonl license: Apache 2.0 - name: example type: example jsonl_fpath: resources_servers/indian_banking/data/example.jsonl # OpenAI-compatible endpoint used for both the user simulator and the judge # (one server so the policy never judges itself; split into two blocks and # repoint judge_model_server if you want separate models). # Supplied from env.yaml (repo root, gitignored) like the policy_* keys: # user_sim_base_url / user_sim_api_key / user_sim_model_name # The user simulator (and judge) must be a different model from the policy. user_simulator_model: responses_api_models: openai_model: entrypoint: app.py openai_base_url: ${user_sim_base_url} openai_api_key: ${user_sim_api_key} openai_model: ${user_sim_model_name} openai_default_headers: {} indian_banking_environment_server: environment_servers: legacy_agent: entrypoint: app.py agent_server: type: responses_api_agents name: indian_banking_agent