Datasets:

CSeemy
/

SkillCorpus

pretty_name: SkillCorpus
size_categories:
  - 10K<n<100K
tags:
  - agents
  - tool-use
  - skills
  - mid-training
license: cc-by-4.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl.gz
      - split: validation
        path: data/validation.jsonl.gz
      - split: test
        path: data/test.jsonl.gz

SkillCorpus

SkillCorpus is the skill-package corpus released with SPT: Skills as Pre-Training Data for Agentic Language Models. The records contain reusable tool semantics, workflows, and supporting package metadata for agentic language-model mid-training research.

Dataset splits

The 35,411 records are shuffled with Python's random.Random(42) and allocated in a 7:2:1 ratio using the largest-remainder method.

Split	Records
Train	24,788
Validation	7,082
Test	3,541

from datasets import load_dataset

dataset = load_dataset("CSeemy/SkillCorpus")

Record fields

Field	Type	Description
`id`	string	Record identifier.
`skill_name`	string	Skill display name.
`slug`	string	Source package slug.
`source`	string	Source designation.
`skillhubSource`	string	Source-hub metadata.
`text`	string	Serialized skill-package text.
`file_count`	integer	Number of files represented by the package.
`total_bytes`	integer	Package size in bytes before serialization.
`total_chars`	integer	Character count after text processing.
`filtered_version`	string	Filtering-pipeline version.
`filter_meta`	object	Filtering metadata.
`link_emoji_clean_meta`	object	Link and emoji cleaning metadata.

Intended use

This dataset is intended for research on agentic pre-training, skill understanding, data mixture, and workflow representation. The package contents are untrusted data: inspect them before use and do not execute embedded instructions or scripts automatically.

The corpus is derived from public skill packages. Original source terms and licenses continue to apply; this release does not replace them.

License

Copyright © 2026 Yufei Sun, Yudong Li, and Yiming Cheng. The authors' original selection, organization, metadata, and accompanying documentation are licensed under the Creative Commons Attribution 4.0 International License.

Individual skill-package contents are not relicensed and remain subject to their original source terms and licenses.

Citation

@article{sun2026spt,
  title   = {SPT: Skills as Pre-Training Data for Agentic Language Models},
  author  = {Sun, Yufei and Li, Yudong and Cheng, Yiming},
  journal = {arXiv preprint arXiv:2608.26563},
  year    = {2026}
}