skillcorpus-demo-1k

1,000 agent skills — SKILL.md documents collected from public repositories, each with the files it ships (scripts/, references/, assets). A sample of SkillCorpus (arXiv:2607.15557) carrying the same schema as the full corpus, small enough to explore in a notebook.

The full corpus is being prepared for release — coming soon. This sample shares its schema, so code written against it carries over unchanged.

Stratified random sample over the 16 categories, drawn from 399 source repositories. Every skill keeps its own upstream license, permissive only: MIT 874 · Apache-2.0 112 · MIT-0 4 · BSD-3-Clause 4 · ISC 3 · CC0-1.0 1 · Mulan-PSL-2.0 1 · Unlicense 1.

Skill bodies and bundled scripts are third-party content that instructs an agent to take actions — read before running.

Retrieval models trained on this schema: skillcorpus-embedding-0.6b and skillcorpus-reranker-0.6b.

Files

file	size	contents
skills.parquet	4.0 MB	1,000 rows, 18 columns
attachments.tar.zst	26 MB	8,552 bundled files (79 MB uncompressed)

Load

from datasets import load_dataset

ds = load_dataset("EverMind-AI/skillcorpus-demo-1k", split="train")
row = ds[0]
print(row["name"], "|", row["category"], "|", row["license"])
print(row["body"][:500])          # the SKILL.md body, ready to inject

A skill's original SKILL.md is frontmatter_raw + body; that file is not duplicated in the tarball. (32 SKILL.md files do appear there — they belong to nested sub-skills shipped inside a skill's own bundle, not to the row itself.)

Attachments

Every skill here bundles at least one file. attachment_path is that skill's member prefix inside attachments.tar.zst:

import tarfile, zstandard as zstd

prefix = row["attachment_path"] + "/"
with open("attachments.tar.zst", "rb") as fh:
    with zstd.ZstdDecompressor().stream_reader(fh) as z:
        with tarfile.open(fileobj=z, mode="r|") as tar:
            for member in tar:
                if member.name.startswith(prefix):
                    print(member.name, member.size)

Bundles are complete: whatever the skill ships is here, minus build artifacts (__pycache__, .git, node_modules, *.pyc, …). 461 skills carry a scripts/ directory and 611 a references/ one.

Columns

column	type	description
skill_id	string	Primary key: upstream source key, skill name and a hash, joined by __. The source key here is the one recorded upstream and does not always match the source column
name	string	Skill name, from frontmatter
description	string	Short description, from frontmatter (normalized upstream, so not always byte-identical to frontmatter_raw)
body	string	SKILL.md markdown body, frontmatter stripped
frontmatter_raw	string	Original YAML frontmatter, JSON-encoded
source	string	Provider key, e.g. NVIDIA/skills, awesome:openclaw/skills. The awesome: prefix marks skills reached through an aggregator rather than crawled from the repo directly
source_url	string	A homepage the skill's author declared in frontmatter — present for 54 skills, and not a pointer to where the skill came from: fewer than half are GitHub links, most name a service the skill calls, and a few are unreplaced placeholders. Use source + source_path for provenance
source_path	string	Path of the skill within the origin repo
license	string	The skill's own declared license
category	string	One of the 16 classes below
tags	list<string>	Free-form tags — declared by only 119 skills
content_hash	string	64-hex content fingerprint carried over from the source corpus, used there for dedup
body_tokens	int32	Crude size proxy, len(body) // 4 — not a tokenizer count (median 1,337)
has_scripts	bool	Bundles a scripts/ directory
has_references	bool	Bundles a references/ directory
added_at	timestamp[us, UTC]	When the skill was fetched into the skill tree
updated_at	timestamp[us, UTC]	Last-updated stamp from the source corpus. Recorded by a different system than added_at, so the two are not on one clock and this is sometimes the earlier of the pair
attachment_path	string	Member prefix inside the tarball

DEV 185 · OTHER 144 · DATA 142 · WRITING 70 · DEVOPS-INFRA 64 · AI-ML 56 · MULTIMEDIA 55 · TESTING 45 · WORKFLOW 44 · FRONTEND-UI 41 · META 37 · DOC-PROC 30 · SECURITY 29 · PRODUCTIVITY 24 · COMMS 21 · AUTH 13

Citation

@article{wang2026skillcorpus,
  title         = {SkillCorpus: Consolidating and Evaluating the Open Skill Ecosystem for Real-World LLM Agents},
  author        = {Wang, Yanze and Yao, Pengfei and Sun, Tianyi and Hu, Chuanrui and Xiao, Yan and Luo, Xiaotian and Han, Yunyun and Chen, Yifan and Sun, Jun and Deng, Yafeng},
  year          = {2026},
  eprint        = {2607.15557},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2607.15557}
}