# Finance Agent v2 (FABv2): Vals's finance-agent-v2 tools wired into Gym. # Run together with a model config (openai_model.yaml or vllm_model.yaml). # # The public FABv2 release ships no grader, so we score the dataset's `rubric` # criteria with our own per-criterion judge: each criterion is graded on its own # (1 = the answer asserts it, 0 = it does not) over judge_required_successes # parsed verdicts, decided by majority vote. Reward is Vals's Partial Credit — # the severity-weighted pass fraction, forced to 0.0 when any must_pass # criterion fails — with rubric_all_pass and the unweighted rubric_fraction # reported alongside. The judge model and prompt are ours, not Vals's private # grader, so treat comparisons against their published numbers as approximate. finance_agent_v2_resources_server: resources_servers: finance_agent_v2: allowed_agents: [finance_agent] entrypoint: app.py domain: agent verified: false description: Vals finance-agent-v2 tools (web/EDGAR/price/calculator) for financial research questions value: Run the official Vals FABv2 agent tools at scale in NeMo Gym # A null key registers that tool as unavailable rather than failing startup, # except sec_api_key in live mode, where startup fails. tavily_api_key: ${oc.select:tavily_api_key,${oc.env:TAVILY_API_KEY,null}} # web_search (Tavily) sec_api_key: ${oc.select:sec_api_key,${oc.env:SEC_API_KEY,null}} # edgar_search (sec-api.io) pricing_data_api_key: ${oc.select:pricing_data_api_key,${oc.env:TIINGO_API_KEY,null}} # price_history (Tiingo) # Required. live: edgar_search queries sec-api.io. local: it reads the index below. edgar_search_mode: ${oc.select:edgar_search_mode,null} local_edgar_index_path: ${oc.select:local_edgar_index_path,null} local_edgar_metadata_path: ${oc.select:local_edgar_metadata_path,null} # Local mode only: parse_html_page reads SEC filings from this corpus after the cache. sec_dump_path: ${oc.select:sec_dump_path,null} # live: price_history queries Tiingo. local: it reads local_pricing_dir and never uses the network. price_history_mode: ${oc.select:price_history_mode,live} local_pricing_dir: ${oc.select:local_pricing_dir,null} retrieval_model_server: type: responses_api_models name: policy_model retrieval_responses_create_params: input: [] retrieval_max_output_tokens: 8192 retrieval_system_prompt_fpath: prompt_templates/finance_agent_v2_retrieval.yaml judge_model_server: type: responses_api_models name: search_judge_model judge_responses_create_params: input: [] max_output_tokens: 8192 reasoning: effort: high rubric_judge_prompt_template_fpath: prompt_templates/finance_agent_v2_rubric_judge.yaml judge_call_timeout: 60.0 # 3 parsed verdicts per criterion decide it by majority; up to 10 calls to # get them (API errors and unparseable replies are retried). A criterion # that never reaches 3 is left unresolved and flags the whole question. judge_required_successes: 3 judge_max_attempts: 10 judge_max_concurrency: 4 # Soft budget: on overrun tools return an error asking the model to submit. max_rollout_time_seconds: 1200 max_end_date: "2026-03-01" # Caches raw upstream responses, so a hit is byte-identical to a live call. # null cache_dir falls back to ~/.cache/nemo_gym/finance_agent_v2. use_cache: true cache_dir: ${oc.select:finance_agent_v2_cache_dir,${oc.env:FINANCE_AGENT_V2_CACHE_DIR,null}} # Kept separate from the policy model: if each policy graded itself, scores # across models would not be comparable. Pin this for any cross-model claim. search_judge_model: responses_api_models: openai_model: entrypoint: app.py openai_base_url: ${oc.select:search_judge_model_base_url,https://api.openai.com/v1} # A visible sentinel rather than an empty string, but note what it does NOT # do: nothing validates it, so an unset key reaches the judge and comes back # as a 401 on the first /verify, after a full rollout has already been paid # for. It stays a sentinel (not `null`, which openai_api_key: str would reject # at parse time) so `gym env resolve` and dry runs still work on a clean # checkout. Set search_judge_model_api_key in env.yaml before a real run. openai_api_key: ${oc.select:search_judge_model_api_key,unset} openai_model: ${oc.select:search_judge_model_name,gpt-5-mini} # The three policy fields below mirror vals-ai/finance-agent-v2 and are checked # against the installed package by tests/test_upstream_parity.py, so a pin bump # that changes them fails there. finance_agent_v2: responses_api_agents: finance_agent: entrypoint: app.py no_tool_call_nudge: >- Your last response produced no tool call. Call `submit_final_result` if you have a final result, otherwise continue with the next tool call. max_time_seconds: 3600 abort_on_tool_error_types: [RetryExhaustedError] # Upstream v2 bounds the run by time, not turns; max_steps stays a # belt-and-braces cap so a looping model cannot burn the full hour. max_steps: 50 truncate_on_overflow: true tool_call_timeout: 300 resources_server: type: resources_servers name: finance_agent_v2_resources_server model_server: type: responses_api_models name: policy_model datasets: - name: example type: example jsonl_fpath: resources_servers/finance_agent_v2/data/example.jsonl # First 5 questions of the public 27Q set, from vals-ai/finance-agent-v2. license: MIT finance_agent_v2_environment_server: environment_servers: legacy_agent: entrypoint: app.py agent_server: type: responses_api_agents name: finance_agent_v2