SkillFlow Final Skill Libraries

This dataset contains the final evolved skill libraries produced by multiple LLM-based agents on the SkillFlow benchmark.

It focuses on the final reusable skill artifacts only, rather than full raw execution trajectories.

Contents

At the time of release, the dataset contains:

• 11 models
• 20 task families
• 220 model-family final libraries
• 598 skills

The dataset also includes:

• index.json: structured metadata for all entries
• index.csv: flat index for filtering and analysis

Structure

The main directory layout is:

<model>/<task_family>/skills/<skill_name>/SKILL.md

Some skill directories may also include small supporting files such as:

• scripts/
• references/

What This Dataset Is For

This dataset is intended for:

• comparing final skill libraries across models
• analyzing how agents externalize reusable procedures into skills
• building retrieval or analysis pipelines over skill artifacts
• studying model differences in skill decomposition and specialization

What This Dataset Does Not Include

This release does not include:

• full raw execution trajectories
• complete intermediate trial histories
• full verifier logs
• complete agent reasoning traces

Notes

• Skill contents are machine-generated artifacts from benchmark runs.
• Final skills should not be treated as guaranteed-correct ground truth without review.