kingdee/lingee-bench

最近记录的来源活动: 2026年10月21日 07:03

SkillsMP 收录数据更新: 2026年11月12日 12:52

已收集 skills: 3

GitHub 星标: 12
GitHub Forks: 0

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/kingdee/lingee-bench.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# Lingee Bench **English** | [简体中文](README.zh-CN.md) Lingee Bench is an agent evaluation framework built on [Harbor](https://github.com/harbor-framework/harbor). It provides harness adapters, model and MCP proxies, an offline ERP mock service, and batch-run and scoring tools. The tasks, input data and graders are published as a dataset with two download sources that carry the same content; use either one: - Hugging Face: [KD-Lingee/lingee-bench](https://huggingface.co/datasets/KD-Lingee/lingee-bench) - GitHub: [kingdee/lingee-bench-datasets](https://github.com/kingdee/lingee-bench-datasets) | Split | Tasks | Where to get it | | --- | ---: | --- | | Enterprise | 30 | Download from Hugging Face or GitHub into `datasets/lg-bench-enterprise-v1.0/`; finance, manufacturing, supply chain | | General | 35 | Download from Hugging Face or GitHub into `datasets/lg-bench-general-v1.0/`; data analysis, documents, retrieval, images and scheduling | | Examples | 2 | Shipped with the repository in [examples/tasks](examples/tasks); for framework self-tests and task authoring, not part of official scores | Enterprise and general results are reported separately. A release is identified by the Git commit or tag of this repository, the source and revision of the dataset (Hugging Face revision or GitHub commit/tag), the task manifest and the file digests together. Supported harnesses: Claude Code, Pi, DeepSeek Harness, OpenCode. See the [docs index](docs/README.md) for runtime requirements and configuration. ## Let a coding agent walk you through it The repository ships skills under `.agents/skills/`. A coding agent that supports skills can follow them from a fresh clone to final scores. It shows each command before running it and asks before anything that costs money: - [lg-bench-setup](.agents/skills/lg-bench-setup/SKILL.md): install, dataset download, images, credentials, preflight and baselines, smoke test, general batch run, score summary. - [erp-mock-batch-eval](.agents/skills/erp-mock-batch-eval/SKILL.md): the 30-task ERP batch run and its summary. Open the agent in the repository root and state your goal, for example: > Use lg-bench-setup to get OpenCode + glm-5.3-flash running: first the example smoke test, then all 35 general tasks, and summarize the scores. Put your keys into `.env.glm` yourself; do not paste them into the chat. The manual steps below follow the same flow. ## 1. Install Requires Python 3.12, [uv](https://docs.astral.sh/uv/) and a working Docker (API ≥ 1.48, macOS or Linux). ```bash uv sync --locked cp .env.example .env.glm ``` Fill in the model endpoint and key in `.env.glm`: `LG_OPENAI_*` for harnesses that speak the OpenAI-compatible protocol, `LG_ANTHROPIC_*` for Claude Code. The judge used by the general and ERP tasks speaks the OpenAI-compatible protocol, so `LG_OPENAI_*` is required for those tasks. Credential files are never committed; real credentials only reach the model proxy. Build the model proxy and the harness image you need, OpenCode as an example: ```bash docker build -f docker/model-proxy/Dockerfile -t lg-bench/model-proxy:0.1.0 . uv run lg-bench harness build opencode ``` ## 2. Smoke test on the example task (no dataset needed) Check that the pipeline and credentials work: ```bash uv run lg-bench preflight --config configs/jobs/opencode-glm-5.2-smoke.yaml --env-file .env.glm --require-env uv run harbor run --config configs/jobs/opencode-glm-5.2-smoke.yaml --env-file .env.glm --job-name smoke-opencode uv run python .agents/skills/lg-bench-setup/scripts/summarize_job.py jobs/smoke-opencode --env-file .env.glm ``` This job uses `glm-5.2`, which is the only model the proxy allows by default. To use another model, set `LG_PROXY_ALLOWED_MODELS` in `.env.glm` and change `model_name` in a `*.local.yaml` copy of the job. ## 3. Download the datasets Use either source. See [datasets](datasets/README.md) for pinning a revision. From Hugging Face: ```bash uvx --from huggingface_hub hf download KD-Lingee/lingee-bench \ --repo-type dataset --revision <revision> --local-dir datasets \ --include "lg-bench-*/**" --include "environments/**" --include "registry.json" ``` Or from GitHub. Binary inputs are stored with Git LFS, so install [git-lfs](https://git-lfs.com/) first; without it you get pointer files. Clone into a scratch directory and copy only the suites and the evaluation environment, so the data repository's root README does not overwrite `datasets/README.md`: ```bash git lfs install D=.cache/lingee-bench-datasets git clone https://github.com/kingdee/lingee-bench-datasets.git "$D" git -C "$D" checkout <revision> cp -R "$D"/lg-bench-enterprise-v1.0 "$D"/lg-bench-general-v1.0 "$D"/environments "$D"/registry.json datasets/ ``` Then verify: ```bash uv run python scripts/check_release.py --datasets ``` ## 4. General tasks: batch run and scoring Generate a batch job for any harness, agent model and judge model. Network isolation and judge wiring follow the official job. The script also writes a model allowlist fragment. ```bash uv run python .agents/skills/lg-bench-setup/scripts/general_job.py \ --harness-job configs/jobs/opencode-glm-5.2-smoke.yaml \ --model glm-5.3-flash --judge-model glm-5.3-flash \ --output configs/jobs/general-opencode-glm-5.3-flash.local.yaml uv run lg-bench preflight --config configs/jobs/general-opencode-glm-5.3-flash.local.yaml --env-file .env.glm --require-env uv run harbor run --config configs/jobs/general-opencode-glm-5.3-flash.local.yaml \ --env-file .env.glm --job-name general-opencode-flash uv run python .agents/skills/lg-bench-setup/scripts/summarize_job.py jobs/general-opencode-flash --env-file .env.glm ``` All 35 tasks run by default; use `--tasks` for a subset. Single-task runs and the network boundary are covered in [general task runs](docs/general-smoke.md). ## 5. ERP tasks: batch run and scoring Build the MCP proxy and the offline ERP mock first. The mock image bakes in the downloaded evaluation instance from `datasets/environments/erp-mock-mcp/`, so rebuild it after updating the dataset: ```bash docker build -f docker/mcp-proxy/Dockerfile -t lg-bench/mcp-proxy:0.1.0 . docker build -f docker/erp-mock-mcp/Dockerfile -t lg-bench/erp-mock-mcp:0.1.0 . ``` Set `LG_PROXY_ALLOWED_MODELS` in `.env.glm` to include both the agent and the judge model, for example `glm-5.3-flash,glm-5.3`. Then generate the 30 per-task configs, run them one by one and summarize: ```bash B=jobs/erp30-opencode uv run python .agents/skills/erp-mock-batch-eval/scripts/batch.py prepare \ --harness-job configs/jobs/opencode-glm-5.2-smoke.yaml \ --model glm-5.3-flash --judge-model glm-5.3 --output $B python3 -c "import json,sys; [print(t['task'], t['config'], t['job_dir']) for t in json.load(open(sys.argv[1]))['tasks']]" $B/batch.json | while read -r task cfg dir; do [ -e "$dir" ] && continue # skip tasks that already have results; safe to rerun after an interruption uv run harbor run --config "$cfg" --env-file .env.glm --job-name "$task" < /dev/null done uv run python .agents/skills/erp-mock-batch-eval/scripts/batch.py summary --batch $B/batch.json ``` Each task binds its own mock dataset and tool allowlist, so each task is its own job. Evaluation runs against the frozen mock published with the dataset, so no connection to a real ERP is needed. Scoring conventions are in [ERP scoring](docs/erp-scoring.md). ## Reading the scores Summaries bucket trials by scoring status. Only officially graded trials enter the mean; execution errors, evaluation errors (including judge failures), missing scores and unfinished trials are listed separately and never counted as zero. When reporting, state the repository commit, dataset source and revision, harness and version, agent and judge models, number of attempts, and the denominator of the mean. Report the general and ERP splits separately. ## Checks and development Without the datasets: ```bash uv run lg-bench validate # validates examples/tasks uv run lg-bench smoke # local reference solution and grader of the example task uv run python scripts/check_release.py uv run pytest -q # dataset-dependent tests are reported as skipped ``` After downloading the datasets: ```bash uv run python scripts/check_release.py --datasets uv run lg-bench validate --path datasets/lg-bench-enterprise-v1.0/tasks uv run lg-bench validate --path datasets/lg-bench-general-v1.0/tasks uv run pytest -q ``` ## Layout ```text examples/tasks/ example tasks (framework self-tests, authoring reference) datasets/ download location for the datasets (not committed, see datasets/README.md) erp-mock-mcp/ offline business service engine, demo instance and hand-modeled packs configs/ Harbor job and compose configuration src/lg_bench/ harness adapters, proxies, validation and preflight scripts/ release checks and batch scripts .agents/skills/ setup, batch-run and score-triage skills docs/ usage and scoring documentation ``` ## License Our own code and documentation are licensed under [Apache-2.0](LICENSE). Third-party components keep their original authorship, copyright and license notices; see [NOTICE](NOTICE) for the scope. The dataset license is stated on the Hugging Face dataset card and in the LICENSE of the GitHub data repository.

正在加载仓库 README…

这个仓库中的 skills

职业分类待补全

skill

职业分类

描述

更新

erp-mock-batch-eval

未分类

对当前固定30道ERP题使用离线mock和指定harness、Agent模型、Judge模型做批量评测，生成逐题配置、预检、直接使用Harbor执行并汇总单轮及多轮结果。支持全批和明确指定的子集。安装环境用lg-bench-setup，修改题目或mock数据不属于本skill。

2026-10-21

lg-bench-setup

未分类

帮外部用户从刚 clone 的 Lingee Bench 仓库一路做到出成绩：装环境、从 Hugging Face 下载题集、构建代理/harness/ERP mock 镜像、配模型凭据、选 job、预检与 oracle/nop 基线、真实运行，最后按评分状态分桶汇总成绩。覆盖示例题冒烟、通用题集 35 题批跑和 ERP 30 题批跑。用户提到安装、上手、快速开始、下载题集、配置 harness/模型/凭据、跑冒烟、跑批、评分汇总时使用。

2026-10-21

score-incident-triage

未分类

真实评测批次出现意外低分、异常分桶、单题reward与人工阅读最终回复的判断明显不符、judge理由与证据对不上，或用户问"这题为啥0分/numeric为0，是缺数据还是agent做不对"时使用。numeric=0或reward=0的题优先按零分异质桶拆解归因。站在出题人立场区分数据、候选、评分器及题目设计问题，按纪律修复评分器并验证；不为特定harness开后门，不在批次运行中改题面。跑批与汇总本身用erp-mock-batch-eval；本skill只管事故归因与处置。

2026-10-21

已展示 3 / 3 个已收集 Skill。