QwenClawBench

A real-user-distribution benchmark for OpenClaw agents — built for robust evaluation at scale

QwenClawBench is a real-user-distribution benchmark for evaluating OpenClaw agents. It was originally built as an internal benchmark during the development of Qwen3.6-Plus, and has since been optimized and open-sourced.

Why QwenClawBench?

QwenClawBench contains 100 tasks across 8 core domains, each with an isolated simulated workspace. Domains are carefully chosen to reflect real OpenClaw usage patterns, and assets are designed to simulate authentic working scenarios.

Reproducing robust large-scale evaluations on OpenClaw is nontrivial, as results depend heavily on infrastructure reliability. We have built in the following features to address this:

• Docker Isolation: Each task runs in a dedicated Docker container, ensuring a consistent and reproducible environment
• Concurrent Execution: Tasks run in parallel across multiple containers, significantly reducing total evaluation time
• Anomaly Detection: Infrastructure failures (API errors, container crashes, timeouts) are flagged explicitly rather than silently folded into scores, so you always know which results to trust
• Resumable Runs: Interrupted runs can be resumed from where they left off, skipping already-completed healthy tasks to obtain stable results

Dataset Structure

This dataset repo contains:

qwenclawbench.jsonl   # Structured task data for Dataset Viewer
tasks/                # 100 raw task .md files (YAML frontmatter + sections)
assets/               # Per-task workspace files copied into each agent container

tasks/ and assets/ are the canonical source used by the benchmark runner. qwenclawbench.jsonl is a derived view for browsing on this page.

Tasks

Task Category Distribution

Category	Count	Description
Workflow and Agent Orchestration	21	Workflow orchestration, skill creation, cron jobs, multi-agent coordination
System Operations and Administration	20	System ops, environment configuration, troubleshooting, workspace management
Knowledge and Memory Management	15	Knowledge base construction, memory system design, document management, context retrieval
Finance and Quantitative Trading	10	Quant strategy backtesting, arbitrage monitoring, trade analysis, position management
Data Analysis and Modeling	10	Statistical analysis, data processing, quality auditing, regression modeling
Security and Vulnerability Management	9	Security auditing, credential management, injection defense, privacy compliance
Communication and Scheduling	8	Message notifications, schedule planning, timed reminders, task scheduling
Research and Information Retrieval	7	Competitive analysis, literature retrieval, technical research, SEO keyword research

Task Structure

Each task is a Markdown file (tasks/task_*.md) with a YAML frontmatter header and structured body sections.

Frontmatter Metadata:

Field	Description
id	Unique task identifier
name	Short task title
category / subcategory	Task category and subcategory
grading_type	Scoring mode: automated, llm_judge, or hybrid
grading_weights	Weight allocation between automated and LLM judge in hybrid mode
timeout_seconds	Task execution timeout
workspace_files	Initial workspace file mappings

Body Sections:

Section	Content
## Prompt	User instructions for the agent
## Expected Behavior	Detailed description of expected behavior; serves as reference for the LLM judge
## Grading Criteria	Scoring checklist (- [ ] format)
## Automated Checks	Python grade(transcript, workspace_path) -> dict function
## LLM Judge Rubric	Scoring dimensions with detailed descriptions for each score tier

Asset Directories:

Each task has a corresponding directory under assets/<task_id>/ containing the initial workspace files (code, configs, data, logs, etc.) copied into the Docker container before the task runs.

Scoring Mechanism

QwenClawBench supports three scoring modes: automated, llm_judge, and hybrid.

Automated: A Python grade(transcript, workspace_path) function performs deterministic, rule-based checks on agent deliverables — verifying output files, command results, and workspace state.

LLM Judge: A judge model (claude-opus-4.5 by default) reviews the agent's action transcript and scores performance across rubric dimensions, each from 0.0 to 1.0.

Hybrid: Both methods run independently and are combined via grading_weights. To prevent agents from receiving high LLM judge scores despite failing concrete deliverable checks, we apply penalized scoring:

$$\text{score}=w_{\text{auto}}\cdot s_{\text{auto}}+w_{\text{llm}}\cdot s_{\text{llm}}\cdot \mathbb{1}[s_{\text{auto}}\ge 0.75]\backslash text\{score\}\; =\; w\_\backslash text\{auto\}\; \backslash cdot\; s\_\backslash text\{auto\}\; +\; w\_\backslash text\{llm\}\; \backslash cdot\; s\_\backslash text\{llm\}\; \backslash cdot\; \backslash mathbb\{1\}[s\_\backslash text\{auto\}\; \backslash geq\; 0.75]$$

When the automated score falls below 0.75, the LLM judge contribution is zeroed out.

Usage

Download the Dataset

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="skylenage-ai/QwenClawBench",
    repo_type="dataset",
)
# path now contains tasks/ and assets/ directories

Run Evaluation

See the GitHub repo for full setup instructions. Quick start:

# Pull the OpenClaw Docker image
docker pull ghcr.io/openclaw/openclaw:main

# Run evaluation (10 parallel containers, 3 runs per task)
./scripts/run.sh --model dashscope/qwen3.6-plus \
    --dataset qwenclawbench-v1.1-100 \
    --runs 3 \
    --concurrency 10 \
    --output-dir ./results/qwen3.6-plus

Acknowledgments

QwenClawBench is built on top of the PinchBench framework. We also acknowledge other open-source contributions from the community, such as Claw-Eval, ZClawBench, and WildClawBench.

Citation

@misc{qwenclawbench1.1,
    title = {{QwenClawBench}: Real-user-distribution benchmark for OpenClaw agents},
    url = {github.com/SKYLENAGE-AI/QwenClawBench},
    author = {{Qwen Team} and {Data Team}, {Alibaba Group}},
    month = {April},
    year = {2026}
}

License

MIT — see LICENSE for details.