经 AI Skill Hub 精选评估,智能评判 获评「推荐使用」。这款AI工具在功能完整性、社区活跃度和易用性方面表现出色,AI 评分 7.5 分,适合有一定技术背景的用户使用。
智能评判 是一款基于 Rust 开发的开源工具,专注于 ai、judge-system、llm 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
智能评判 是一款基于 Rust 开发的开源工具,专注于 ai、judge-system、llm 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
# 方式一:cargo install(推荐) cargo install nanojudge # 方式二:从源码编译 git clone https://github.com/nanojudge/nanojudge cd nanojudge cargo build --release # 二进制在 ./target/release/nanojudge
# 查看帮助 nanojudge --help # 基本运行 nanojudge [options] <input> # 详细使用说明请查阅文档 # https://github.com/nanojudge/nanojudge
# nanojudge 配置说明 # 查看配置选项 nanojudge --config-example > config.yml # 常见配置项 # output_dir: ./output # log_level: info # workers: 4 # 环境变量(覆盖配置文件) export NANOJUDGE_CONFIG="/path/to/config.yml"
NanoJudge quantifies the relative strengths of arbitrary items under a criterion you define, using LLMs as judges. Provide the criterion (e.g., "Which is healthier?") and your item list of any length (e.g., "Eggs", "Butter", "Spinach", ...), and get a ranking with confidence intervals.
Instead of overwhelming an LLM with one massive prompt, NanoJudge breaks the task down into a series of head-to-head matchups. Operating like an intelligent matchmaking league, it adaptively pairs items of similar strength against each other as the results come in to efficiently produce an accurate leaderboard. These individual wins and losses are fed into an Elo-style rating system, producing a transparent ranking, all backed by AI explanations.
Works with any OpenAI-compatible API endpoint.
nanojudge.ai is a hosted version built on this engine, wrapped in a web UI with managed GPU infrastructure.
Download a prebuilt binary from GitHub Releases, or build from source:
cargo install --path nanojudge-cli
First, create a config file with your judge panel:
nanojudge init # creates ~/.config/nanojudge/config.toml
Example config with multiple judges:
rounds = 10
logprobs = true
[[judge]]
endpoint = "http://localhost:8000"
model = "Qwen/Qwen3-4B-Instruct-2507"
weight = 2
temperature = 0.8
[[judge]]
endpoint = "https://api.openai.com/v1"
model = "gpt-4o"
api_key_env = "OPENAI_API_KEY"
weight = 3
temperature = 1.0
concurrency = 5
Each [[judge]] block defines a judge in the panel. Comparisons are distributed across judges according to their weight. All judges share the same logprobs mode.
Then run:
```bash
The config file lives at ~/.config/nanojudge/config.toml. Run nanojudge init to create one with defaults and documentation for all available options.
Key settings:
| Setting | Description |
|---|---|
rounds | Number of comparison rounds |
logprobs | true to extract logprobs for continuous confidence (requires endpoint support, e.g. vLLM). false for text-based verdict parsing (works everywhere, but needs more rounds). |
comparison_distribution | "uniform" (default) or "top-heavy" |
Per-judge settings (in [[judge]] blocks):
| Setting | Required | Description |
|---|---|---|
endpoint | Yes | OpenAI-compatible API base URL |
model | Yes | Model ID |
temperature | Yes | Sampling temperature |
weight | No | Relative weight for pair assignment (default: 1) |
concurrency | No | Max concurrent requests (default: 16) |
max_tokens | No | Max tokens in response (default: 2048) |
api_key_env | No | Environment variable containing the API key |
reasoning_effort | No | Controls model reasoning mode (e.g. "none" to disable Qwen 3.5 thinking) |
min_logprob_coverage | No | Min fraction of A–E logprob mass required to trust a verdict, > 0.0 and ≤ 1.0 (default: 0.95) |
Save a sample of LLM responses to a JSONL file for spot-checking or live monitoring with tail -f:
```bash
nanojudge rank ... --save-comparisons 1.0
nanojudge rank ... --save-comparisons 0.1
nanojudge rank ... --save-comparisons 50
nanojudge rank ... --save-comparisons 0.3 --save-comparisons-to samples.jsonl ```
Each line is a JSON object with round, item1, item2, probability, and response (the raw LLM text). Lines are flushed immediately so you can tail -f during a run.
智能评判工具,使用LLM进行排名
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
✅ MIT 协议 — 最宽松的开源协议之一,可自由商用、修改、分发,仅需保留版权声明。
AI Skill Hub 点评:智能评判 的核心功能完整,质量良好。对于AI 技术爱好者来说,这是一个值得纳入个人工具库的选择。建议先在非生产环境试用,再逐步推广。
| 原始名称 | nanojudge |
| 原始描述 | 开源AI工具:Rank any list by any criteria at any scale using LLMs as judges.。⭐29 · Rust |
| Topics | aijudge-systemllmrankingrating-system |
| GitHub | https://github.com/nanojudge/nanojudge |
| License | MIT |
| 语言 | Rust |
收录时间:2026-06-09 · 更新时间:2026-06-09 · License:MIT · AI Skill Hub 不对第三方内容的准确性作法律背书。