---
task_categories:
- text-generation
language:
- zh
---

# QuestBench

**QuestBench** is a course-curated benchmark of **256 expert-level questions** for evaluating the deep-search capabilities of language models, introduced in the paper [Teaching AI Through Benchmark Construction: QuestBench as a Course-Based Practice for Accountable Knowledge Work](https://huggingface.co/papers/2605.21413). 

Questions are long-tail queries spanning **14 normalized domains** in the humanities and social sciences (Literature, Law, History, Foreign Languages, Social Sciences, Arts, Archaeology, Journalism, …). Each question, reference answer, and grading rubric is in Chinese; the model under evaluation may search and read in any language.

## Files

- `questbench.jsonl` — 256 questions, one JSON record per line.
- `.claude.json` — [Claude Code](https://www.anthropic.com/claude-code) MCP server config.
- `vision_server.py` — local MCP server proxying image analysis to OpenRouter.
- `package.json`, `package-lock.json` — npm dependencies.

Each JSONL record has four fields: `domain`, `question`, `answer`, `grading_criteria` (the rubric an LLM judge uses to assign a 0–100 score).

## Run

Setup:

```bash
brew install --cask claude        # Claude Code CLI
brew install jq node uv
npm install
export ANTHROPIC_API_KEY=...      # Claude
export SERPER_API_KEY=...         # https://serper.dev
export OPENROUTER_API_KEY=...     # https://openrouter.ai (vision)
```

Run a single question:

```bash
QUESTION=$(jq -r '.question' questbench.jsonl | head -n 1) && \
  claude -p "$QUESTION" \
    --mcp-config ./.claude.json --strict-mcp-config \
    --permission-mode bypassPermissions \
    --disallowedTools WebSearch WebFetch \
    --output-format stream-json --verbose
```