midah/hf-dataset-licenses

SPDX License Dataset

Build pipeline for three datasets on Hugging Face:

• midah/hf-dataset-licenses — 700+ SPDX licenses with full text, FSF metadata, HF usage categories, and compatible exceptions
• midah/hf-asset-license — ~95 HF Hub-selectable licenses (SPDX-mapped + HF-only like Llama 2, Gemma, OpenRAIL) with LLM-extracted clause features
• midah/hf-dataset-licenses-exceptions — SPDX license exceptions used in WITH expressions (e.g. GPL-2.0-only WITH Classpath-exception-2.0)

How it works

The pipeline fetches every license from the SPDX License List, enriches each one with FSF classification, Hugging Face Hub usage statistics, temporal metadata, and text metrics, then publishes the result as a HF dataset.

The flow has three phases: build, enrich, and publish.

Build

build_license_dataset.py fetches the SPDX license list, downloads each license's full text and page markdown from spdx.org, parses version/family information, and maps compatible exceptions. It also pulls FSF metadata (libre/non-free, GPL compatibility) and derives a usage category from HF Hub counts. The output is spdx_licenses_dataset/ (Arrow) and spdx_licenses_dataset.parquet.

build_spdx_exceptions.py does the same for SPDX exceptions — the license addenda used in WITH expressions.

Two supporting build scripts produce the reference data that feeds into the main build:

• build_hf_selectable.py — maps SPDX identifiers to Hugging Face Hub license dropdown values and writes schemas/hf_selectable_licenses.json
• build_hf_license_counts.py — scrapes HF Hub for per-license dataset and model counts, writes schemas/hf_license_counts.json and hf_license_counts.parquet

Enrich

enrich_dataset.py adds columns to the existing dataset without rebuilding from scratch. Each enrichment can run independently:

Flag	What it adds	Source
--fsf	fsf_category, is_fsf_libre, compatibility flags	FSF API via license_utils/fsf.py
--temporal	spdx_ref, is_deprecated	SPDX license list via license_utils/temporal.py
--spdx-dates	spdx_added_date, spdx_last_modified	GitHub commit history via license_utils/spdx_dates.py
--lengths	num_chars, num_words	license_texts.json via license_utils/text_metrics.py
--usage	usage_category, dataset_count, model_count	HF counts via license_utils/hf_usage.py

Publish

Any build or enrich step accepts --push to upload the result to Hugging Face. make push-licenses rebuilds and pushes in one step; make push-only pushes the existing local dataset.

export_licenses_metadata.py exports a lightweight licenses_metadata.json (no full texts) for downstream tooling.

Plots

gen_plots.py --all generates all figures into figures/. The individual plot scripts can also run standalone:

Script	Figures	What it shows
gen_openness_plots.py	figures/openness/	Openness score distribution, HF-selectable vs non-selectable
gen_temporal_plots.py	figures/temporal/	Version evolution, calendar-time trends, text complexity
gen_usage_plot.py	figures/usage/	Top HF license usage by artifact type

Project layout

scripts/                    Pipeline scripts (build, enrich, export, plot)
license_utils/              Shared library (FSF, temporal, SPDX dates, HF usage, text metrics)
schemas/                    Reference JSON (features.json, hf_selectable_licenses.json, hf_license_counts.json)
figures/                    Generated plots (openness, temporal, usage, overview, llm)
llm_extractions/            LLM-based feature extraction (separate pipeline)
hf_catalog/                 HF catalog data

Quick start

pip install -r requirements.txt

# Full pipeline
make build-licenses                          # ~15 min
python scripts/build_hf_selectable.py        # SPDX→HF mapping
python scripts/build_hf_license_counts.py    # HF usage counts
make enrich                                  # All enrichments
make push-licenses                           # Push to HF

# Individual steps
make enrich-usage                            # Just usage columns
make gen-plots                               # Generate all figures
make export-metadata                         # Export licenses_metadata.json

Dataset fields (hf-asset-license)

Core: license_name, spdx_id, license_family, version, version_modifier, full_text, source_url, page_markdown, source
FSF: fsf_category, fsf_tags, fsf_gpl_compatibility, is_fsf_libre, is_fsf_nonfree, fsf_gpl2_compatible, fsf_gpl3_compatible, fsf_fdl_compatible, fsf_viewpoint_restricted
Usage: usage_category, dataset_count, model_count
Text: num_chars, num_words, ttr, guiraud_r, dugast_u, mattr
LLM-extracted: prohibits_commercial_use, requires_attribution, requires_sharealike, applies_to_outputs, has_clause_training, has_clause_finetuning, has_provenance_obligations, has_clause_model_redistribution, has_clause_data_redistribution, has_clause_access_mode, allows_training_use, has_patent_grant, has_patent_termination, allows_commercial_use, prohibits_derivatives, is_ml_specific, has_ai_use_restrictions, jurisdiction, patent

Usage

from datasets import load_dataset

# Full SPDX corpus (700+ licenses)
ds = load_dataset("midah/hf-dataset-licenses")
mit = ds["train"].filter(lambda x: x["spdx_id"] == "MIT")[0]

# HF Hub-selectable subset with LLM features (~95 licenses)
ds = load_dataset("midah/hf-asset-license")

License

CC0-1.0. Individual license texts retain their original terms.