FinanceTasks

Description

FinanceTasks is a synthetic training environment for agentic financial research. Each task is an investor query asked as of a specific date; the agent researches it on the web (SEC filings, earnings call transcripts, press releases, market data) and submits a long-form answer, graded checklist-style against per-query rubrics generated from — and verified against — public primary sources.

Capabilities

• Financial research across the investor workflow: financial data/modeling, sector/industry analysis, earnings/events, company research, coverage/catalyst monitoring, and screening/discovery
• Exhaustive retrieval (temporal across quarters, cross-entity within sectors, thematic within calls)
• Numerical reasoning over financial figures with correct units and periods
• Temporal anchoring: answering relative to the query's as-of date
• Long-form synthesis with professional investor judgement

Compute Requirements

This environment does not require a sandbox; compute requirements are minimal. Grading and web search are API calls.

License

MIT.

Tasks

A single train split with 2,509 tasks (67,904 rubrics; per-task quartiles 16/23/35; 67% must-have). Tasks expose only the query and its as-of date; rubrics stay server-side. Use-case shares track the FrontierFinance eval distribution:

Tasks come in five archetypes by document-bundle shape: single_filing (985), cross_entity (633), temporal (547), transcript (220), market (124). Companies referenced in the 220 FrontierFinance eval queries were excluded from the task universe, so training shares no companies with the eval.

Reward Structure

Sparse, LLM-graded reward delivered once when the agent finishes with a plain-text reply. Each rubric is an atomic pass/fail criterion; the reward is the query's rubric qualification rate:

$R = \frac{\text{rubrics satisfied}}{\text{total rubrics}} \in [0, 1]$

In test rollouts with gpt-5.2, rewards ranged from 0.08 to 1.00 across tasks (20–47 tool calls per episode), indicating a usable training signal spread rather than saturation.

Tools

Search and fetch come from the OpenReward SDK's WebToolset (toolsets = [WebToolset]) rather than being implemented in this environment.

• web_search — search the web. Takes a query and optional allowed_domains or blocked_domains; returns a Links: list of {title, url, snippet} sources
• web_fetch — fetch the readable content of a URL. Takes a url and a prompt describing what to extract

Choosing a search backend

Which provider answers those two tools is configuration on the environment server, not code here, so swapping it needs no change to this environment:

Tasks in this environment carry a query_date and instruct the agent to answer as of that date, so the default backsearch backend is the better fit: Tavily searches the live web and cannot bound results to a cutoff.

Grading uses a hidden @terminal tool: when research is complete, reply with the full long-form answer as an ordinary message. The whole reply is graded by a gpt-5-mini judge against the query's rubric checklist.

Time Horizon

Multi-turn agentic research episodes ending in a single plain-text reply. Test rollouts with gpt-5.2 used 20–47 tool calls depending on query breadth.

Environment Difficulty

Rewards in gpt-5.2 spot-check rollouts spanned 0.08 (a subsidiary-level 10-Q MD&A query) to 1.00 (a single-company quarterly print), with judge fail reasons naming the precise missing figure or comparison. Cross-entity and temporal archetypes generally demand substantially more retrieval than single-filing tasks.

Other Environment Requirements

• openai_api_key — required, for the gpt-5-mini grading judge.
• Search credentials — whichever the configured backend needs: api_key for the default backsearch backend, or tavily_api_key when the server runs with OPENREWARD_SEARCH_BACKEND=tavily. Both fall back to the server process environment (OPENREWARD_API_KEY / TAVILY_API_KEY) if not passed. An unconfigured backend surfaces as a tool error rather than failing session creation.

Safety

The agent performs read-only web research on public financial information and produces a text answer; it takes no real-world actions and handles no funds. Standard web-access considerations apply (the agent fetches live third-party content). Answers are research artifacts graded against rubrics, not investment advice, and models trained on this environment should not be treated as licensed financial advisors.

Citations

FinanceTasks is an original synthetic dataset by the General Reasoning team, designed to mirror the task distribution of the FrontierFinance benchmark:

@dataset{GRFinanceTasks,
  author    = {General Reasoning Inc. Team},
  title     = {FinanceTasks},
  year      = {2026},
  publisher = {OpenReward},
  url       = {https://openreward.ai/GeneralReasoning/FinanceTasks}
}