| --- |
| pretty_name: SkillNet-Fabric |
| --- |
| |
| # SkillNet-Fabric |
|
|
| This repository contains the released Skills and graphs, benchmark resources, and experiment |
| artifacts for SkillFabric. |
|
|
| - **Code:** https://github.com/zjunlp/SkillNet-Fabric |
| - **Artifacts:** https://huggingface.co/datasets/zjunlp/SkillNet-Fabric |
| - **Related SkillNet paper:** https://arxiv.org/abs/2603.04448 |
|
|
| ## Repository contents |
|
|
| | Path | Contents | |
| | --- | --- | |
| | `skills/` | Skill corpus used to construct the released graphs | |
| | `data/` | SkillRouter, SkillsBench, and AgentSkillOS benchmark inputs | |
| | `graphs/` | `G_SR`, `G_SB`, `G_A53`, `G_A500`, `G_A1K`, and graph-build logs | |
| | `main_experiments/` | Runs, results, evaluator records, deliverables, and logs for the three main experiments | |
| | `analysis_experiments/` | Localization and ablation experiment artifacts | |
| | `artifact_index.json` | Bundle names, sizes, extraction targets, and checksums | |
| | `provenance.json` | Source checkout and release inventory | |
|
|
| ## Experiments included |
|
|
| | Experiment | Released artifacts | |
| | --- | --- | |
| | SkillRouter | SkillFabric routes, per-task scores, aggregate results, and runtime logs | |
| | SkillsBench | Two executors, 87 tasks, three trials, selected and runtime Skills, trajectories, verifier records, and deliverables | |
| | AgentSkillOS | Two executors, three Skill pools, task workspaces, generated frames, baseline artifacts, Bradley-Terry rankings, and runtime logs | |
| | Localization analysis | Routes, rankings, protocol, reports, and logs | |
| | AgentSkillOS ablation | Run outputs, rankings, reports, and logs | |
|
|
| The experiment directories preserve the run names used on the experiment server. |
|
|
| ## Download |
|
|
| Install the [SkillNet-Fabric repository](https://github.com/zjunlp/SkillNet-Fabric), then run: |
|
|
| ```bash |
| skillfabric-repro download --repo-id zjunlp/SkillNet-Fabric --revision main --profile main |
| ``` |
|
|
| The command downloads the selected bundles and extracts them into `experiments/artifacts/` using |
| the directory layout expected by the reproduction code. |
|
|
| ## Sources and licenses |
|
|
| - SkillRouter benchmark and evaluation: https://github.com/zhengyanzhao1997/SkillRouter and |
| https://huggingface.co/datasets/pipizhao/SkillRouter-Eval-Core |
| - SkillRouter registry Skills: `majiayu000/claude-skill-registry`, pinned by the released layer |
| manifest |
| - SkillsBench v1.1 tasks and evaluator: https://github.com/benchflow-ai/skillsbench |
| - AgentSkillOS tasks and evaluator: https://arxiv.org/abs/2603.02176 |
| - Redistributed Skill packages: mixed-source snapshot; layer manifests record the source datasets |
| and pinned upstream revisions |
|
|
| The repository `LICENSE` covers SkillFabric code. Released benchmark resources and redistributed |
| Skills retain their applicable upstream terms. Layer manifests preserve the available source and |
| license metadata. |
|
|
| ## Citation |
|
|
| Until the SkillFabric paper is public, cite the related SkillNet paper: |
|
|
| ```bibtex |
| @article{liang2026skillnet, |
| title = {SkillNet: Create, Evaluate, and Connect AI Skills}, |
| author = {Liang, Yuan and Zhong, Ruobin and Xu, Haoming and Jiang, Chen and others}, |
| journal = {arXiv preprint arXiv:2603.04448}, |
| year = {2026} |
| } |
| ``` |
|
|
