TeamHSHIntelligence's picture
Update README.md
e77fb39 verified
metadata
license: other
license_name: permissive-mixed
license_link: LICENSE
task_categories:
  - text-generation
  - fill-mask
  - feature-extraction
language:
  - en
tags:
  - code
  - github
  - ai-training
  - llm
  - fine-tuning
  - code-generation
  - python
  - javascript
  - typescript
  - rust
  - go
  - bigcode-standard
  - stack-v2-methodology
  - commercial-safe
  - pii-scrubbed
  - license-audited
pretty_name: HSH Intelligence  GitHub Code AI Training Corpus (5K Sample)
size_categories:
  - 1K<n<10K

HSH Intelligence — GitHub Code AI Training Corpus

5,000-record sample of the HSH Intelligence GitHub Code AI Training Corpus.

A curated, production-grade sample of source code from top-tier public GitHub repositories — engineered for large language model training, fine-tuning, and code understanding research.

The full corpus contains 5.6 TB of source code (211 million+ files, 7.05 billion lines) across 14 production languages.


10/10 Quality Checks

This sample passes all 10 industry-standard quality checks following BigCode / The Stack v2 production methodology.

# Check Tool Result
1 License compliance scancode-toolkit 32.5.0 0% copyleft
2 Secret detection gitleaks 8.18.4 0 leaks
3 Near-duplicate removal MinHash LSH (256-perm, 5-gram, 0.9 threshold) 0% duplicates
4 Code complexity radon 6.0.1 3.92 avg cyclomatic
5 Token diversity tiktoken cl100k_base (GPT-4) 63,712 unique tokens
6 Statistical balance Custom audit 1K per language
7 Benchmark contamination vs HumanEval (164) + MBPP (500) 0 matches
8 PII beyond secrets Custom regex + Luhn validation 0 real PII
9 Syntax validation Babel parser, syn 2.0, tsc, ast, gofmt 98.0% parseable
10 Repo legitimacy GitHub REST API verification 100% verified

Reference: Methodology follows BigCode / The Stack v2 production standards.

Full audit certificate: QUALITY_CERTIFICATE.json


Sample Specifications

Metric Value
Records 5,000 (curated subset)
Languages 5 (Python, JavaScript, TypeScript, Go, Rust)
Records per language 1,000 (perfectly balanced)
Unique repositories 1,499 verified active on GitHub
Format Apache Parquet (zstd compression) + CSV
Schema 19 fields per record
Size 13.4 MB (Parquet) / 49.9 MB (CSV)
License coverage 100% commercial-safe (MIT, Apache-2.0, BSD, ISC)
PII status Fully scrubbed (zero secrets, emails, IPs, SSNs)
Syntax validation 98.0% parseable (industry standard greater than or equal to 95%)

Repository Quality

  • 56.1% from repos with 10,000+ GitHub stars
  • 6.1% archived repos (still valid, just not actively maintained)
  • 0.0% deleted repos
  • Top repos include: facebook/react, ollama/ollama, django/django, AUTOMATIC1111/stable-diffusion-webui

Full Corpus Specifications

Metric Value
Total dataset size 5.6 TB (raw) / 391 GB (Parquet, compressed)
Total records 211 million+ code files
Total lines of code 7.05 billion
Unique repositories 3,710+ permissive-license repos
Programming languages 14 production languages
Updates Daily incremental

Languages covered: Python, JavaScript, TypeScript, Go, Rust, Java, C++, Ruby, Swift, Kotlin, PHP, C#, Scala, Solidity


License Coverage (Commercial-Safe Only)

License Status Notes
MIT INCLUDED Most permissive
Apache-2.0 INCLUDED Permissive with patent grant
BSD-2-Clause INCLUDED Permissive
BSD-3-Clause INCLUDED Permissive
ISC INCLUDED Permissive
GPL-2.0 / GPL-3.0 EXCLUDED Copyleft
AGPL-3.0 EXCLUDED Strong copyleft
LGPL-2.1 / LGPL-3.0 EXCLUDED Copyleft
No license / Proprietary EXCLUDED Default copyright

License detection performed using scancode-toolkit 32.5.0 with per-file SPDX classification.


Schema (19 Fields)

Field Type Description
id string Unique record identifier (sha256-prefixed)
language string Detected programming language
repo_owner string GitHub username or organization
repo_name string Repository name
repo_stars integer GitHub star count
repo_forks integer GitHub fork count
repo_description string Repository description
repo_topics list[string] GitHub repo topics
license string SPDX license identifier
file_path string Relative path within repo
file_name string Filename with extension
file_size integer File size in bytes
code string Raw source code content (PII-scrubbed)
word_count integer Total word count
char_count integer Character count
line_count integer Total lines of code
data_quality_score float Composite quality score (0.0–1.0)
timestamp timestamp Record creation timestamp
scrubbed boolean PII scrubbing flag (always True)

Quick Start

Load with Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset("HSH-Intelligence/github-code-corpus-sample")
print(ds)
print(ds["train"][0])

# Filter to high-quality Python only
python_only = ds["train"].filter(
    lambda x: x["language"] == "Python" and x["data_quality_score"] >= 0.95
)
print(f"High-quality Python records: {len(python_only)}")

Load directly with pandas

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/HSH-Intelligence/github-code-corpus-sample/github_code_sample_5000.parquet"
)
print(df.head())
print(f"Total records: {len(df):,}")
print(f"Languages: {df['language'].value_counts()}")
print(f"Top repos: {df['repo_name'].value_counts().head(10)}")

Live API Demo

Try the full corpus via the live API sandbox (no signup required):

curl -H "X-API-Key: demo-key-12345" \
  "https://api.hshintelligence.com/api/v1/github-code-corpus?language=Rust&license=MIT&page_size=5"

Returns real Parquet records with full metadata: code, license, repo stars, quality score, commit history. Free tier limited to 2 files (~18 records). Full corpus delivered via Backblaze B2 download link after purchase.

API documentation: Or use the interactive docs — click any endpoint, click "Try it out", paste the demo key, and run live queries.

Live endpoint: https://api.hshintelligence.com/api/v1/github-code-corpus

Or run the interactive Google Colab notebook:
https://links.hshintelligence.com/github-demo


Use Cases

  • LLM pre-training — multi-language code corpus for foundation models
  • Code completion fine-tuning — Copilot-style models
  • Code search and retrieval — embedding training
  • Code understanding research — academic benchmarks
  • Vertical AI — domain-specific code assistants
  • Benchmark-safe evaluation — zero contamination vs HumanEval/MBPP

Why This Corpus

vs. Alternative HSH Intelligence Edge
The Stack v2 Per-file license audit + provenance trail + 10-check quality verification
Common Crawl code Pre-filtered, deduplicated, syntax-validated, PII-scrubbed
Custom GitHub scraping Saves 4+ months of engineering work
Internal datasets EU AI Act Article 10 compliance ready
Generic samples Industry-standard 10/10 quality checks documented

Compliance & Provenance

  • EU AI Act Article 10 ready (training data governance)
  • GDPR safe (zero PII verified)
  • CCPA safe (no California resident data)
  • HIPAA considerations addressed (no medical data)
  • Per-record license audit trail
  • Source attribution retained (repo_owner, repo_name)
  • Quality scoring per record
  • Zero PII (emails, phones, IPs, SSNs, credit cards verified)
  • Zero secrets (API keys, tokens, credentials verified via gitleaks)
  • Zero benchmark contamination (HumanEval, MBPP verified)

Methodology

This dataset follows BigCode / The Stack v2 production methodology with additional quality gates.

Tools Used

Category Tools
License detection scancode-toolkit
Secret scanning gitleaks
Deduplication datasketch MinHash LSH
Complexity analysis radon
Tokenization tiktoken (cl100k_base)
Syntax validation Babel parser, syn 2.0, tsc, Python ast, gofmt
Repo verification GitHub REST API v3

Quality Thresholds

  • License compliance: less than 0.1% copyleft (achieved: 0%)
  • Secret leaks: 0 tolerance (achieved: 0)
  • Near-duplicates: less than 5% (achieved: 0%)
  • PII: 0 tolerance (achieved: 0)
  • Syntax validation: greater than or equal to 95% parseable (achieved: 98%)
  • Repo legitimacy: less than 1% deleted (achieved: 0%)

Full quality certificate: QUALITY_CERTIFICATE.json


Full Corpus Access

This is a 5,000-record evaluation sample. The full corpus is available via commercial license:

Tier Records Languages Format
Sample (this dataset) 5,000 5 Parquet + CSV
Standard 10M+ 14 Parquet
Enterprise 211M+ (full) 14 Parquet (+JSONL on request)

Delivery options:

  • Cloud signed URL (Backblaze B2, AWS S3)
  • Cross-cloud transfer (AWS, GCP, Azure)
  • sFTP delivery for on-prem
  • Daily incremental updates (Enterprise tier)

Custom subsets available: Filter by language, license, repo stars, complexity, or quality threshold.

Licensing: 1-year non-exclusive commercial license.


Contact


About HSH Intelligence

HSH Intelligence is the Data Division of Healing Sun Haven LLC, building production-grade AI training datasets and B2B intelligence products.

We engineer datasets across AI training, B2B intelligence, and decision-support — purpose-built for frontier AI labs and enterprise teams who demand industry-standard quality verification.


This dataset is provided for evaluation purposes. The full 5.6 TB corpus is available under commercial license. Quality audit certificate, license documentation, and provenance trail included with all enterprise contracts.

Audit date: 2026-05-07 | Methodology reference: BigCode/Stack v2 | Full quality report: QUALITY_CERTIFICATE.json