The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 291, in _generate_tables
pa_table = paj.read_json(
io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
)
File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: JSON parse error: Column(/top5_largest_families/[]/top_names/[]/[]) changed from string to number in row 0
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 101, in _split_generators
pa_table = next(iter(self._generate_tables(**splits[0].gen_kwargs, allow_full_read=False)))[1]
~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 306, in _generate_tables
batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
examples = [ujson_loads(line) for line in original_batch.splitlines()]
~~~~~~~~~~~^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
return pd.io.json.ujson_loads(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
ValueError: Expected object or value
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Skill Atlas
Verdicts, copy groups and a search index for about 1.9 million public agent skills (SKILL.md files on GitHub).
This dataset is derived from GitSkills (Destefanis et al., arXiv 2608.10906), shared under CC BY 4.0. Please credit GitSkills if you use these files. The Skill Atlas code is MIT, at 0xQuez/skill-atlas.
Skill Atlas answers one question for each skill: is this worth using? Each row has a verdict (use, caution or ignore), reasons, a pointer to the original version of a near-copy, slop and risk flags, and a white-hat tag for legitimate security tools.
It filters out bad options. It does not yet find better ones. In a 60-task test, near-duplicate picks in an agent's top 5 dropped from 26% to under 1%, with no loss in relevance. See the repo's eval/README.md.
Try it
- Website: 0xQuez/skill-atlas-web (direct: https://0xquez-skill-atlas-web.static.hf.space)
- Search API: https://0xquez-skill-atlas-api.hf.space (
/healthz,/stats,/skills/search,/skills/{id}/verdict,/skills/{id}/canonical) - MCP (streamable HTTP): https://0xquez-skill-atlas-api.hf.space/mcp
- Code: 0xQuez/skill-atlas
{
"mcpServers": {
"skill-atlas": { "url": "https://0xquez-skill-atlas-api.hf.space/mcp" }
}
}
What's in this dataset
| File | Role |
|---|---|
verdicts_compact.parquet |
One row per distinct skill (1,877,981). The main table. Descriptions are included; full skill bodies are not (those stay on GitHub). |
search_api.parquet |
Same skills plus a normalised search string (hay, name_n, repo_n) and rank rk. Used by the JSON API and MCP server. |
search_index.parquet |
Smaller search columns for the website (DuckDB-WASM). |
skill_details.parquet |
Card fields for the website, sorted by skill_id with small row groups so the browser can fetch one skill without reading the whole file. |
headline.json |
Snapshot totals used by /stats and the Overview page. |
white_hat_stats.json |
White-hat tagging counts and the hand-check note. |
slop_criteria.json |
The three slop rules and the owner exemptions. |
malicious_source_stats.json |
Counts for the known-harmful-collection list. |
risk_v2_stats.json |
Risk v2 counts and the precision note. |
verdict_shift_v1_v2.json |
How verdicts changed between v1 and v2. |
ingest_stats.json, refine_pairs_stats.json |
Pipeline counts. |
label_translation.csv, label_before.csv, label_after.csv |
Hand labels for near-copy and translation checks. |
Embeddings, MinHash signatures, the full DuckDB database and other internal intermediates are not included. They are large and only needed to rebuild the atlas from scratch (see the git repo).
Verdicts
| Verdict | Meaning |
|---|---|
use |
Fine to use. |
caution |
Read the reasons first (medium risk or rough quality). |
ignore |
A copy, slop-skill, junk, or danger. If it is a copy, use canonical_id instead. |
Counts in this snapshot: use 1,163,380; caution 139,907; ignore 574,694.
Common reasons (a skill can have several):
near_copy_of:<id>— a copy; use skill<id>renamed_copy— a copy published under a different nameslop_skill— one of many template look-alikes from one accountknown_malicious_source— from a collection of deliberately harmful test skills, or a copy of the harmful partrisk_high/risk_medium/risk_flag:<flag>— danger heuristicsquality_low/quality_medium/invalid_frontmatter— thin or roughjunk:*— empty, a redirect, or too shorttranslated_copy_of:<id>— a translation; information only, does not change the verdict
canonical_id is the original (best) version in the same copy group and name. Known-malicious skills are never chosen as the original. If every member is known-malicious, canonical_id is null.
Slop-skills
A slop-skill is one of many look-alike skills mass-produced by one account from a template. Three rules (see slop_criteria.json): a huge near-copy cluster dominated by one owner; numbered templated names from an owner who has many of them; or a mass of unnamed skills from one owner. They stay in the dataset and in search; they are tagged, not removed. Klotzkette and membranedev are exempt (deliberate templated collections).
White-hat security tools
white_hat marks legitimate offensive-security, pentest, CTF, forensics or malware-analysis tools. They do dangerous things on purpose. Use them only for security work you're allowed to do. The tag never changes the verdict or the risk score.
- 13,981 skills in about 1,900 repos (
white_hat_kindisoffensiveorforensics;categoryiswhite_hat_security) - In a hand check of 40 tagged skills, 36 were real security tools, 2 were wrong and 2 were unclear
- Real toolkits such as AeonDave/malskill are tagged white-hat; they are not on the harmful-collection list
Columns (verdicts_compact.parquet)
| Column | Type | Meaning |
|---|---|---|
skill_id |
int | Stable id for this distinct content |
canonical_id |
int or null | Id of the original in this name group |
file_sha |
string | Content hash from GitSkills |
name, description |
string | From the skill front matter (description may be truncated) |
rep_repo, rep_path |
string | The GitSkills representative copy |
github_url |
string | Link to that copy on GitHub |
top_repo |
string | Most-starred repo that holds this exact content |
license |
string | Declared license of the representative repo, if any |
max_stars, n_files, n_repos |
int | Popularity and how widely the text is copied |
cluster_id, cluster_size |
int | Near-copy family |
name_key |
string | Normalised name used to pick the original |
is_canonical |
bool | This row is the original for its (family, name) |
family_canonical_id |
int | Family-level original (fallback) |
canon_basis |
string | Which rule picked the original |
renamed_copy |
bool | Same text, different name |
origin_date |
timestamp | Best-effort date from the snapshot (often unreliable) |
quality, risk |
int | Heuristic scores, 0–100-ish |
risk_flags |
string | Comma-separated flag names |
slop_skill, slop_reasons |
bool, string | Slop tag and evidence |
verdict, reasons |
string, JSON | The verdict and machine-readable reasons |
white_hat, white_hat_kind, category |
bool, string, string | White-hat tag |
search_api.parquet adds name_n, repo_n, hay (normalised search text) and rk (rank). The two website files are column subsets, optimised for browser range requests.
Known limitations
- Verdicts come from rules, not from a person reading each skill. They catch a lot, but not everything.
- A high risk score means "read this before you run it". It does not prove the skill is harmful. Risk-flag precision: of 370 skills newly marked high-risk in v2, 289 were in known harmful collections. In a hand review of 30 of the other 81, only 3 really sent secrets or keys to an unvetted host; 14 were offensive-security or CTF notes with live attack commands, 11 were ordinary skills with a hard-coded server IP, and 2 were pattern mistakes.
- The atlas cannot tell when a skill only works inside one project.
- Origin dates in the snapshot are unreliable, so the original is picked by stars first, then date.
- This is a snapshot. Some GitHub repos have changed or gone since.
- Near-copy detection can still miss or over-link a few cases (letter-suffixed template siblings, shared banners, same-author rewrites).
License and attribution
These files are licensed under CC BY 4.0, as a work based on GitSkills. Please cite:
- Destefanis et al., GitSkills, arXiv 2608.10906, Hugging Face dataset
- Skill Atlas, https://github.com/0xQuez/skill-atlas
Skill contents remain under each file's own license. Many source repos declare none. This dataset links to GitHub rather than re-hosting full skill bodies.
- Downloads last month
- 145