Verified Koskari language examples exported from the Koskari conformance corpus. Each row traces to a corpus record with machine-checked expectations at one or more language phases (read, expand, eval, error, invalid). Every corpus record.
Dataset Card
By Koskari, published under gpl-3.0, revision c3f2babea980.
Koskari language corpus
Verified Koskari language examples exported from the Koskari conformance corpus. Each row traces to a corpus record with machine-checked expectations at one or more language phases (read, expand, eval, error, invalid).
Source repository: github.com/koskari-lang/koskari
Configurations
corpus (default)
Every corpus record. One row per record regardless of which verify stages are present. Missing stages are null.
| Column | Description |
|---|---|
corpus_id |
Unique corpus record identifier |
feature |
Feature path (e.g. special/compile/cond, layout/if) |
source |
Koskari source text |
read |
Expected reader output from verify.read |
expand |
Expected macro expansion from verify.expand |
eval |
Expected eval result from verify.eval |
error |
Expected pipeline error from verify.error |
invalid |
Prose explanation from verify.invalid |
references |
Links to user documentation pages under docs/ |
stability |
Record stability label (stable, evolving, experimental) |
tags |
Maintainer tags (not copied from training-only metadata) |
schema_version |
Corpus schema identifier at export time |
git_ref |
Short git commit at export time |
docs
User documentation pages (Markdown) from the source repository's
docs/reference/, docs/start/, docs/guide/, docs/patterns/, and
docs/agents/ trees. One row per page. The path column matches the
strings in the corpus config's references column verbatim, so a
record's references resolve directly to their documentation content.
The agents section is written for AI consumers rather than humans:
dense normative rules, valid/invalid pairs, and guidance on querying
this dataset. Filter with section == "agents".
| Column | Description |
|---|---|
path |
Docs-relative page path (e.g. reference/specials/compile-define.md) |
title |
Page title (first Markdown heading) |
summary |
First paragraph, plain text |
content |
Full raw Markdown of the page |
section |
Parent directory (e.g. reference/specials) |
kind |
Always doc |
schema_version |
Corpus schema identifier at export time |
git_ref |
Short git commit at export time |
Whole-file cells: each page is one content cell. The Hugging Face
Dataset Viewer caps cell size in API responses, so a very large page would
be truncated. Pages are small today (a few KiB); chunk-by-heading is a
deferred mitigation should pages grow. See tools/corpus/views/docs.ki.
programs
Example Koskari programs (whole .ki files and their READMEs) from the
source repository's examples/ tree, excluding build artifacts. These give
composition-level context — full programs rather than single-construct
snippets.
These rows are illustrative and NOT verified. Unlike the corpus
config, they carry no machine-checked read/expand/eval expectations; they
are working example source, not conformance assertions.
| Column | Description |
|---|---|
path |
Repo-relative file path (e.g. examples/life/koslife/life.ki) |
title |
File stem, or README heading |
summary |
Module docstring or leading comment, plain text |
content |
Full raw source of the file |
section |
Top-level example group (e.g. examples/life) |
kind |
Always example |
schema_version |
Corpus schema identifier at export time |
git_ref |
Short git commit at export time |
Verify stages
Each corpus record asserts behaviour at language boundaries:
- read — reader and layout output
- expand — macro expansion on the outer form
- eval — runtime result (including
@-marked complex encodings in source) - error — expected pipeline failure
- invalid — documented illegal source (may coexist with read expectations)
Only stages present in the source corpus record appear as non-null columns in
the corpus config.
Provenance
Every row includes corpus_id, schema_version, and git_ref so downstream
consumers can join back to the authoritative YAML record and the export
revision.
License
GNU General Public License v3. See the Koskari repository for full terms.
Structure
corpus 786 rows
| Split | Rows | Size |
|---|---|---|
| train | 786 | 141.4 KB |
docs 509 rows
| Split | Rows | Size |
|---|---|---|
| train | 509 | 1.1 MB |
programs 67 rows
| Split | Rows | Size |
|---|---|---|
| train | 67 | 312.3 KB |
Details
- Repository
- koskari/koskari
- Publisher
- Koskari
- Task category
- Not stated by the source
- Tags
- Not stated by the source
- Size category
- Not stated by the source
- Languages
- Not stated by the source
- Revision
- c3f2babea9809a245e5c44da2a853ac67d677ff8
- Last updated
- 2026-09-18
Files
6 files, 730.1 KB in total.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| koskari-corpus/train.parquet | Data | 73.5 KB | 7b0c10c6407e |
| koskari-docs/train.parquet | Data | 486.7 KB | 42b6839685d3 |
| koskari-eval/train.parquet | Data | 19.5 KB | 11efe423f097 |
| koskari-programs/train.parquet | Data | 143.3 KB | d8f0ec5f00e4 |
| README.md | Documentation | 4.7 KB | — |
| .gitattributes | Repository | 2.5 KB | — |
License and Download
- License
- gpl-3.0
- Access
- No access gate
Released by Koskari through its official repository on Hugging Face. Read the license.