LLM-MovieLens: A Portable Pipeline and Testbed for LLM-Synthesized Item Profiles

Abstract

LLM-MovieLens is a portable pipeline that turns a catalogue's structured metadata into LLM-synthesized item features, together with the evaluation harness needed to find out whether those features help a recommender. A language model reads an item's metadata and writes an 80–120 word profile, a ten-axis mood vector and 3–5 key themes; the harness encodes that text once, offline, and evaluates it under a temporal split, five seeds and full ranking. This record archives the exact bytes the accompanying paper's numbers were computed on. How the files fit together The code surface is one archive. The data surface is split into parts, because most of its weight is the embedding arrays and a single large file is a poor thing to download over a slow connection. Every part unpacks to the same hf/ or github/ root: extract them all into one directory and you have the released tree exactly. shasum -a256 -c SHA256SUMS, run from inside github/ and from inside hf/, then verifies all 248 and 55 files — the README.md in this record gives the exact commands and says what a partial download looks like. Each part is independently useful too — take data-core alone for the dataset card, licence, datasheet, Croissant record, prompt manifest and the complete human-evaluation study, then add only the catalogue and encoder you need. What the data archive contains 10,381 film profiles for the genome-covered items of MovieLens 20M, generated with Claude Haiku 4.5, and a second, independent generation of the same catalogue with GPT-4o-mini. 9,289 book profiles for Amazon-Books, a catalogue with no tag-genome equivalent, produced by the same pipeline into the same output schema. Pre-computed embeddings of the generated text — profile, mood and theme matrices — under bge-large-en-v1.5 and e5-large-v2. A 500-profile, three-annotator human evaluation: 15,000 five-point Likert judgments over two generating models and five quality axes, plus 3,000 forced-choice mood judgments, with the blank sheets, the filled sheets and the agreement analysis. A Croissant 1.0 record, a datasheet, the prompt-provenance manifest, and SHA-256 digests of every file. What the code archive contains The profile generator, the embedding generator and the benchmark harness. 14 benchmark configurations × 5 seeds = 70 runs, with the per-seed result files those runs produced, under a temporal split, full ranking and cold-start stratification. A one-command verifier that recomputes the paper's headline numbers from the released result files in about a minute, without a GPU. A split rebuilder: no source interactions are redistributed, so the splits are reconstructed from MovieLens 20M and SHA-256-checked against ours. A new sentence encoder, generating model, catalogue or recommendation method is a path override, not a fork. Licensing — this release is split Zenodo records carry one licence field, so the more restrictive of the two is selected above. In full: CC BY 4.0 — all code, the benchmark result files, the human-evaluation annotations, the metadata and the documentation. These are our own work end to end and carry no upstream restriction. CC BY-NC 4.0 — the generated profiles and their embeddings. They are our output, but generated from sources whose own terms are non-commercial, and we cannot grant rights we do not hold. Research use is unrestricted under both. The full text is in LICENSE inside each archive. Not redistributed No MovieLens interaction data, no tag genome or transformation of it, and no cached third-party item metadata. The harness rebuilds what it needs from the upstream sources. This product uses the TMDB API but is not endorsed, certified, or otherwise approved by TMDB.