---
license: other
license_name: mixed-permissive
license_link: LICENSE
task_categories:
  - text-generation
language:
  - en
  - gle
tags:
  - code
  - gleam
  - programming-language
  - corpus
  - code-generation
size_categories:
  - 10K<n<100K
---

# Gleam Code Corpus

A structured, attributed corpus of **22,581 Gleam source files** collected from **1,528 GitHub repositories** for continued pre-training and code generation research.

## Overview

This dataset contains `.gleam` source files from the [Gleam](https://gleam.run/) programming language ecosystem. Every file includes full attribution — repo owner, name, URL, license, star count, and Hex.pm package metadata where available.

**What this is for:**
- Continued pre-training (CPT) of code-focused language models
- Fine-tuning models for Gleam code generation
- Research on small programming language ecosystems
- Reference corpus for Gleam idioms and patterns

**What this is NOT:**
- Instruction/response pairs (see the companion `gleam-sft` dataset, coming later)
- Documentation (language tour, hexdocs — separate dataset, coming later)
- Raw/unfiltered data (generated files, unlicensed code, and duplicates have been removed)

## Dataset Configs

| Config | Files | Size | Description |
|--------|-------|------|-------------|
| `all` | 22,581 | 27 MB | Full filtered corpus |
| `official` | 296 | 422 KB | Official repos only (`gleam-lang`, `gleam-wisp`, `lustre-labs`) |
| `community_top` | 6,702 | 8 MB | Community repos with stars ≥ 10 or Hex downloads ≥ 1,000 |

All configs contain only permissively licensed, deduplicated code.

## Schema

Each record is a single `.gleam` file with full attribution:

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | `{owner}/{repo}/{file_path}` — unique identifier |
| `source` | string | Always `"github"` |
| `code` | string | Raw file contents |
| `file_path` | string | Path within the repository |
| `repo_url` | string | GitHub repository URL |
| `repo_owner` | string | Repository owner/organization |
| `repo_name` | string | Repository name |
| `repo_description` | string | Repository description from GitHub |
| `stars` | int | GitHub star count at collection time |
| `is_official` | bool | True if owner is `gleam-lang`, `gleam-wisp`, or `lustre-labs` |
| `license` | string | Normalized SPDX license identifier |
| `license_raw` | string | Original license string from source |
| `authors` | list[string] | Repository authors/contributors |
| `hex_package` | string? | Hex.pm package name (null if not on Hex) |
| `hex_downloads` | int? | Hex.pm download count (null if not on Hex) |
| `fork` | bool | Whether the repo is a fork (unreliable — see notes) |
| `archived` | bool | Whether the repo is archived |
| `collected_at` | string | ISO 8601 timestamp of collection |
| `commit_sha` | string | Git commit SHA of the snapshot |
| `schema_version` | int | Schema version (currently 1) |

## Collection Methodology

1. **Discovery**: GitHub search API with `language:gleam`, using star-range + date-range chunking to bypass the 1,000-result-per-query limit. Discovered 4,660 unique repositories across 17 query chunks.

2. **Enrichment**: Cross-referenced with Hex.pm `search=gleam` (1,035 packages) to add package names and download counts.

3. **Collection**: Shallow-cloned each repo (`git clone --depth 1`), extracted all `.gleam` files. 4,489 repos processed, 0 clone failures.

4. **Filtering**: Applied a multi-stage quality pipeline:

   | Filter | Removed | Reason |
   |--------|---------|--------|
   | Generated files (>50K chars) | 666 | Auto-generated schemas, lookup tables, data dumps |
   | Stub files (<20 chars) | 316 | Empty/placeholder files |
   | Unlicensed repos | 42,368 | No license and no Hex package |
   | Non-permissive licenses | 0 | GPL, AGPL removed |
   | Exercism repos | 840 | Same exercises solved repeatedly — low diversity |
   | Content deduplication | 2,642 | Identical code from different repos (kept highest-starred) |

   **69,413 raw files → 22,581 clean files** (67% reduction)

## Statistics

### Corpus
- **22,581 files** from **1,528 repos**
- **4.3M lines**, **121M characters**
- Median file: 2,692 chars; Mean: 5,352 chars; Max: 49,945 chars

### License Distribution
- MIT: 57% (13,068 files)
- Apache-2.0: 40% (8,723 files)
- Other permissive: 3% (MPL-2.0, BSD, ISC, Unlicense, 0BSD, CC0-1.0)

### Source Distribution
- `src/` files: 48%
- `test/` files: 24%
- `examples/` files: 6%
- Other: 22%

### Star Distribution
| Stars | Files | % |
|-------|-------|---|
| 0 | 6,904 | 31% |
| 1–4 | 6,842 | 30% |
| 5–9 | 3,432 | 15% |
| 10–49 | 3,602 | 16% |
| 50–99 | 757 | 3.4% |
| 100–999 | 918 | 4.1% |
| 1000+ | 126 | 0.6% |

## Known Limitations

- **`fork` field is unreliable**: GitHub search API doesn't consistently mark forks. All records show `fork=false`. Some near-duplicate fork content may be present (content dedup catches exact copies only).
- **Near-duplicate forks**: Diverged forks (e.g., `Ivan-Vakhula07/lustre` vs `lustre-labs/lustre`) survived dedup because their files differ slightly. These are valid Gleam code and not harmful for training.
- **Semi-generated content**: `dusty-phillips/monks_of_style` contributes 496 files (2.2%) of auto-generated CSS property wrappers from MDN documentation. Each file is valid Gleam but follows an identical template.
- **Hex.pm enrichment incomplete**: Only 44% of files (10,015) have Hex package metadata. The rest are from repos not published on Hex.
- **Collection snapshot**: Data was collected on 2026-05-28. Repositories may have changed since.

## Licensing

This dataset aggregates code from many repositories, each with their own license. The `license` field in each record contains the normalized SPDX identifier. All included code is under permissive licenses:

- MIT
- Apache-2.0
- MPL-2.0
- BSD-3-Clause
- BSD-2-Clause
- ISC
- Unlicense
- 0BSD
- CC0-1.0

The dataset card and curation scripts are released under the same license terms as the underlying code. See individual records for specific license attribution.

## How to Use

```python
from datasets import load_dataset

# Load full corpus
ds = load_dataset("kasuboski/gleam-code-corpus", config="all")

# Load only official Gleam repos
ds = load_dataset("kasuboski/gleam-code-corpus", config="official")

# Load community top (stars >= 10 or hex_downloads >= 1000)
ds = load_dataset("kasuboski/gleam-code-corpus", config="community_top")

# Access a record
record = ds["train"][0]
print(record["code"])           # the Gleam source code
print(record["repo_url"])       # GitHub URL for attribution
print(record["license"])        # SPDX license identifier
```

## Acknowledgments

- The [Gleam](https://gleam.run/) language team and community
- [Hex.pm](https://hex.pm/) for package metadata
- All the repository authors whose code is included in this corpus