Upload README.md with huggingface_hub

README.md

@@ -16,9 +16,7 @@ size_categories:
 
 # GitHub Top Developer Source Code
 
-A curated dataset of **1.3M+ source code files** from **GitHub's ranked developers (2015-2025)**. 
-
-This dataset is filtered to code written by developers who repeatedly appeared on GitHub's trending page (Based on this dataset: https://huggingface.co/datasets/ronantakizawa/github-top-developers)
+A curated dataset of source code from **GitHub's top trending developers (2015-2025)**. Unlike bulk code scrapes, this dataset is filtered to code written by developers who repeatedly appeared on GitHub's trending page, linking each file to rich developer and repository metadata.
 
 ## Dataset Summary
 
@@ -28,13 +26,6 @@ This dataset is filtered to code written by developers who repeatedly appeared o
 - **Permissive licenses only** (MIT, Apache-2.0, BSD, ISC, etc.)
 - **Rich metadata** per file: repo stars, description, primary language, developer company affiliation
 
-## Use Cases                                                                               
-                                                                                          
-  - Code generation fine-tuning — Train or fine-tune LLMs on high-quality source code written by recognized developers, providing a higher signal-to-noise ratio than bulk     GitHub scrapes                                                                          
-  - Cross-language code analysis — Study patterns across 80+ programming languages from the same pool of elite developers to compare idioms, styles, and conventions
-  - Open source contribution patterns — Analyze how prolific open source contributors structure their most popular projects
-
-
 ## Schema
 
 Each row represents a single source file:
@@ -54,20 +45,69 @@ Each row represents a single source file:
 
 **Note on language columns:** `file_language` is determined per-file from the file extension (e.g. a `.py` file is always `Python`). `repo_primary_language` is GitHub's auto-detected primary language for the entire repository. These may differ — for example, a C header file (`.h` → `C/C++ Header`) in a repo that GitHub classifies as `Python`.
 
+## Splits
+
+| Split | Description |
+|-------|-------------|
+| `train` | ~90% of repos — for training |
+| `test` | ~5% of repos — for evaluation |
+| `validation` | ~5% of repos — for hyperparameter tuning |
+
+Splits are assigned **by repository** (deterministic hash), so no repo appears in multiple splits. This prevents data leakage from files in the same project.
+
 ## Usage
 
 ```python
 from datasets import load_dataset
 
-ds = load_dataset("ronantakizawa/github-top-code", split="train")
+# Load a specific split
+train = load_dataset("ronantakizawa/github-top-code", split="train")
+test = load_dataset("ronantakizawa/github-top-code", split="test")
 
 # Filter by language
-python_files = ds.filter(lambda x: x["file_language"] == "Python")
+python_files = train.filter(lambda x: x["file_language"] == "Python")
 
 # Filter by stars
-popular = ds.filter(lambda x: x["repo_stars"] > 1000)
+popular = train.filter(lambda x: x["repo_stars"] > 1000)
 
 # Get files from a specific developer
-dev_files = ds.filter(lambda x: x["developer_username"] == "torvalds")
+dev_files = train.filter(lambda x: x["developer_username"] == "torvalds")
 ```
 
+## What Makes This Dataset Unique
+
+| Feature | This Dataset | The Stack | codeparrot/github-code |
+|---------|-------------|-----------|----------------------|
+| Curated by developer reputation | Yes | No | No |
+| Developer metadata (company) | Yes | No | No |
+| Stars per repo | Yes | Yes | No |
+| Permissive licenses only | Yes | Yes | Mixed |
+
+Existing code datasets are massive bulk scrapes of all public GitHub repos. This dataset is intentionally curated: every file comes from a developer who was recognized on GitHub's trending page, providing a higher signal-to-noise ratio for studying elite developer practices or fine-tuning code models.
+
+## Collection Methodology
+
+1. **Developer sourcing**: 4,763 unique developers extracted from [ronantakizawa/github-top-developers](https://huggingface.co/datasets/ronantakizawa/github-top-developers), which tracks GitHub trending page appearances from 2015-2025 via Wayback Machine snapshots.
+2. **Repository discovery**: For each developer, their top 10 repositories by stars were selected using the GitHub API, filtered to repos they own (not forks or contributions).
+3. **License filtering**: Only repositories with permissive licenses (MIT, Apache-2.0, BSD, ISC, Unlicense, etc.) were included.
+4. **Code extraction**: Repository tarballs were downloaded and source files extracted, skipping binary files, vendored directories (`node_modules`, `vendor`, etc.), generated files, and files exceeding 1MB.
+5. **Source code filtering**: Only source code files are included (80+ language extensions). Configuration files (JSON, YAML, TOML, XML, INI), documentation (Markdown, TXT, RST), and template files are excluded.
+6. **Metadata enrichment**: Each file is linked to repository metadata (stars, language, description) and developer metadata (username, name, company).
+
+## Filtering Applied
+
+- **Repos**: Owner-created only (no forks), permissive license, non-empty
+- **Directories skipped**: `node_modules`, `vendor`, `third_party`, `dist`, `build`, `__pycache__`, `.git`, `venv`, and 30+ more
+- **Files skipped**: Binary files, files >1MB, non-UTF-8 files, config/doc files (JSON, YAML, TOML, XML, Markdown, etc.)
+- **Top 10 repos per developer** by star count
+
+## Limitations
+
+- Star counts and commit counts reflect the time of collection, not real-time values
+- Some developers may have deleted or renamed their accounts since trending
+- Commit counts are for the default branch only
+- The dataset reflects trending developers specifically, which may over-represent certain languages or project types popular on GitHub
+
+## Source
+
+Built from [ronantakizawa/github-top-developers](https://huggingface.co/datasets/ronantakizawa/github-top-developers) using the GitHub REST API.