ToolMind-Web-QA

API Endpoint
Leaderboard
Loading leaderboard...
README

ToolMind-Web-QA

⭐ OpenReward Environment Hugging Face Dataset

Description

ToolMind-Web-QA is an ORS environment for evaluating agents on complex, multi-hop question answering that requires web search and information synthesis. Questions are synthesized from Wikipedia entity-relation knowledge graphs and require multiple search steps to resolve. Agents use web search and URL fetching tools to find information before submitting their answer.

Capabilities

  • Multi-hop reasoning across multiple web sources
  • Web search query formulation
  • Information extraction from web pages
  • Synthesizing answers from multiple documents

Compute Requirements

This is a multi-turn environment with no sandbox. Agents interact through web search and URL fetching tools only.

License

Apache 2.0.

Tasks

There is one split in this environment:

  • Train: 6,801 complex multi-hop questions

Each task presents a question that requires searching and synthesizing information from multiple web sources to answer.

Reward Structure

This is a multi-turn environment with binary reward:

  • 1.0 — Correct answer (semantically equivalent to the reference, as judged by gpt-5-mini)
  • 0.0 — Incorrect answer

The agent uses web_search and fetch_url to research, then calls submit_answer with an explanation, exact answer, and confidence score. The LLM grader evaluates semantic equivalence using both the reference answer and the solution reasoning trace, accepting minor formatting differences.

Data

Data consists of a single JSONL file (syn_wikiqa.jsonl) containing 6,801 multi-hop questions. Each instance includes the question text, ground truth answer, and a solution reasoning trace showing the multi-hop logic required.

Source: Nanbeige/ToolMind-Web-QA

Tools

ToolDescription
web_searchSearch the web using Tavily API. Returns titles, URLs, and snippets.
fetch_urlFetch the full content of a specific URL.
submit_answerSubmit your answer with explanation, exact answer, and confidence score.

Note that the fetch_url and web_search tools require Tavily, but are optional. If you want to use a different provider for search you can exclude these tools and use external tools instead.

Time Horizon

ToolMind-Web-QA is a multi-turn environment. Agents typically perform several web searches and URL fetches before submitting a final answer.

Environment Difficulty

The ToolMind paper evaluates models trained on this dataset across web research benchmarks:

ModelGAIABrowseCompBrowseComp-ZHHLE
DeepSeek-V3.263.5%67.6%65.0%40.8%
MiniMax-M275.7%44.0%48.5%31.8%
GLM-4.671.9%45.1%49.5%30.4%
ToolMind-Web-3B67.0%17.4%30.8%24.8%

Questions are synthesized from Wikipedia entity-relation knowledge graphs and require multi-hop reasoning to trace connections between entities.

Other Environment Requirements

  • OpenAI API key: Required for LLM-based answer grading via gpt-5-mini
  • Tavily API key: Required for web search and URL content extraction

Pass via secrets={"openai_api_key": "...", "tavily_api_key": "..."}.

Safety

Agents in ToolMind-Web-QA perform web searches to answer factual questions. The environment does not involve sensitive content or actions beyond information retrieval.

Citations

@article{toolmind2025,
  author    = {Nanbeige Lab},
  title     = {Nanbeige ToolMind Technical Report: A Large-Scale, Reasoning-Enhanced Tool-Use Dataset},
  journal   = {arXiv preprint arXiv:2511.15718},
  year      = {2025},
  url       = {https://arxiv.org/abs/2511.15718}
}
GeneralReasoning/ToolMind-Web-QA | OpenReward