onradio/imdb4m: v1.1.0 - IMDB4M code, KG, and Zenodo embeddings

v1.1.0 — IMDB4M code, KG, and Zenodo embeddings This release pairs the IMDB4M source code and the cleaned, disk-aligned knowledge graph with the matching pre-computed embeddings, archived separately on Zenodo. Highlights Cleaned KG: data/kg/imdb_kg_cleaned.ttl (~1.80 M triples), one-for-one aligned with the embedding rows on Zenodo via imdb4m:hasEmbedding records. Pre-computed multi-modal + KG embeddings on Zenodo: 10.5281/zenodo.20057840 — image, video, audio, text, and RotatE knowledge-graph embeddings (full graph + 5 held-out variants), released as parquet files plus a master HDF5 mirror and embedding_metadata.ttl SPARQL pointers. Drop-zone layout: embeddings_output/ ships empty in the repo with a README; drop the Zenodo files into it and every script in the project picks them up automatically — no further configuration needed. Lightweight release bundle: the release/ pipeline (python -m release) now produces a small KG + docs distribution with an empty embeddings_output/ drop-zone instead of trying to bundle the ~20 GB of embedding artefacts. Documentation refresh: README.md, docs/index.html, void.ttl, and the release/ bundle templates all reference the real Zenodo DOI; the KG-entity counts have been homogenised to rdflib (656,121 nodes) across all docs, with a prominent note explaining the 118-node delta against the PyKEEN entity table. Get started git clone https://github.com/onradio/imdb4m.git cd imdb4m pip install -r requirements.txt # Pull every embedding file from Zenodo into embeddings_output/ pip install zenodo_get zenodo_get 10.5281/zenodo.20057840 -o embeddings_output/ # Sanity-check the layout python verify_embeddings_output.py What the released vectors contain | Modality | Rows | Dim | Model | |---|---:|---:|---| | Image | 33,247 | 768 | openai/clip-vit-large-patch14 | | Video | 4,350 | 512 | microsoft/xclip-base-patch32 | | Audio | 4,034 | 512 | laion/larger_clap_music_and_speech | | Text | 4,216 | 1,024 | BAAI/bge-large-en-v1.5 | | KG (RotatE, full + 5 held-out variants) | 139,465 each | 512 (real, from 256-d complex) | PyKEEN RotatE | All vectors are L2-normalised; cosine similarity equals the dot product. Licence CC BY-NC-SA 4.0 for the KG structure and derived embeddings, aligned with IMDb's Conditions of Use. The original images, videos, and audio files are not redistributed — only URLs and the derived vectors are shared. Citation @dataset{imdb4m_embeddings_2026, title = {{IMDB4M} Multi-Modal and KG Embeddings (v1)}, author = {Reklos, Ioannis and de Berardinis, Jacopo and Simperl, Elena and Mero{\~n}o-Pe{\~n}uela, Albert}, year = {2026}, doi = {10.5281/zenodo.20057840}, url = {https://doi.org/10.5281/zenodo.20057840} } Full changelog: https://github.com/onradio/imdb4m/compare/v1.0.0...v1.1.0