Create new eval suites for the deepagentsjs monorepo. Handles dataset design, test case scaffolding, scoring logic, vitest configuration, and LangSmith integration. Use when the user asks to: (1) create an eval, (2) write an evaluation, (3) add a benchmark, (4) build an eval suite, (5) evaluate agent behaviour, (6) add test cases for a capability, or (7) implement an existing benchmark (e.g. oolong, AgentBench, SWE-bench). Trigger on phrases like 'create eval', 'new eval', 'add eval', 'benchmark', 'evaluate', 'eval suite', 'write evals for'.
- 📁 references/
- 📁 scripts/
- 📄 SKILL.md
This skill should be used when the user asks to "implement LLM-as-judge", "compare model outputs", "create evaluation rubrics", "mitigate evaluation bias", or mentions direct scoring, pairwise comparison, position bias, evaluation pipelines, or automated quality assessment.
Automatically evaluate and compare multiple AI models or agents without pre-existing test data. Generates test queries from a task description, collects responses from all target endpoints, auto-generates evaluation rubrics, runs pairwise comparisons via a judge model, and produces win-rate rankings with reports and charts. Supports checkpoint resume, incremental endpoint addition, and judge model hot-swap. Use when the user asks to compare, benchmark, or rank multiple models or agents on a custom task, or run an arena-style evaluation. --- # Auto Arena Skill End-to-end automated model comparison using the OpenJudge `AutoArenaPipeline`: 1. **Generate queries** — LLM creates diverse test queries from task description 2. **Collect responses** — query all target endpoints concurrently 3. **Generate rubrics** — LLM produces evaluation criteria from task + sample queries 4. **Pairwise evaluation** — judge model compares every model pair (with position-bias swap) 5. **Analyze & rank** — compute win rates, win matrix, and rankings 6. **Report & charts** — Markdown report + win-rate bar chart + optional matrix heatmap ## Prerequisites ```bash # Install OpenJudge pip install py-openjudge # Extra dependency for auto_arena (chart generation) pip install matplotlib ``` ## Gather from user before running | Info | Required? | Notes | |------|-----------|-------| | Task description | Yes | What the models/agents should do (set in config YAML) | | Target endpoints | Yes | At least 2 OpenAI-compatible endpoints to compare | | Judge endpoint | Yes | Strong model for pairwise evaluation (e.g. `gpt-4`, `qwen-max`) | | API keys | Yes | Env vars: `OPENAI_API_KEY`, `DASHSCOPE_API_KEY`, etc. | | Number of queries | No | Default: `20` | | Seed queries | No | Example queries to guide generation style | | System prompts | No | Per-endpoint system prompts | | Output directory | No | Default: `./evaluation_results` | | Report language | No | `"zh"` (default) or `"en"` | ## Quick start ### CLI `
Add a new simulation benchmark to the VLA evaluation harness. Use this skill whenever the user wants to integrate, create, or add a new benchmark or simulation environment — e.g. 'add ManiSkill3', 'integrate OmniGibson', 'hook up a new sim'. Also use when they ask how benchmarks are structured or want to understand the benchmark interface.
Add a new SWE benchmark task from a real GitHub bug-fix. Use when the user provides a GitHub issue or PR URL and wants to add it to the bench-swe pipeline.
Use when adding a new scenario to remindb's benchmark suite — symptoms include "compare X tool against grep/cat", "add a token-savings benchmark for Y", "extend `internal/bench/scenarios.go`", "wire a new scenario into `bench.Run`", or any task that adds a row to the `scenario / naive (tok) / remindb (tok) / saved` output table. Distinct from Go `testing.B` benchmarks in `*_bench_test.go`.
Write benchmark scripts for EmbodiChain modules following project conventions
Critically analyze content, claims, or arguments with rigorous evaluation.
- 📁 assets/
- 📁 references/
- 📁 scripts/
- 📄 SKILL.md
Use this when you need to EVALUATE OR IMPROVE or OPTIMIZE an existing LLM agent's output quality - including improving tool selection accuracy, answer quality, reducing costs, or fixing issues where the agent gives wrong/incomplete responses. Evaluates agents systematically using MLflow evaluation with datasets, scorers, and tracing. IMPORTANT - Always also load the instrumenting-with-mlflow-tracing skill before starting any work. Covers end-to-end evaluation workflow or individual components (tracing setup, dataset creation, scorer definition, evaluation execution).
- 📁 examples/
- 📁 scripts/
- 📁 server/
- 📄 .gitignore
- 📄 group.jpg
- 📄 install.sh
学术论文搜索与分析服务 (Academic paper search & analysis)。当用户涉及以下学术场景时，必须使用本 skill 而非 web-search：搜索论文、查找 ArXiv/PubMed/PapersWithCode 论文、查询 SOTA 榜单与 benchmark 结果、引用分析、生成论文解读博客、查找论文相关 GitHub 仓库、获取热门论文推荐。Keywords: arxiv, paper, papers, academic, scholar, research, 论文, 学术, 搜索论文, 找论文, SOTA, benchmark, MMLU, citation, 引用, 博客, blog, PapersWithCode, HuggingFace.
Create coding agent benchmarks for evaluation with nasde. Use this skill when the user wants to: - Create a new benchmark project (set of tasks for evaluating coding agents) - Add tasks to an existing benchmark - Create or modify agent variants (configurations that control agent behavior) - Set up assessment dimensions and scoring criteria - Verify that a new benchmark's Docker environment and tests work Even if the user doesn't say "benchmark" — if they're talking about creating coding challenges for AI agents or setting up evaluation criteria, this skill applies. --- # NASDE Benchmark Creator Create and configure coding agent benchmarks for evaluation with `nasde`. A benchmark is a set of coding tasks that AI agents solve inside isolated Docker containers, scored both by functional tests (pass/fail) and by an LLM-as-a-Judge architecture assessment. ## Critical: line endings on Windows (read this first) Benchmark scripts execute inside **Linux** sandboxes (Docker, Daytona). If `tests/test.sh`, `solution/solve.sh`, or `environment/Dockerfile` are checked out with **CRLF** line endings (the Windows git default when `core.autocrlf=true` and there is no `.gitattributes`), every trial fails immediately with: ```
Run a full Build + Style + Move + Write evaluation on a page — score each framework, produce a combined report out of /200 with prioritized recommendations across all four.
