recsys-challenge-2026-artifacts

Intermediate artifacts of team FPMs_UMONS's pipeline for the RecSys Challenge 2026. They let you run Blind-B inference without recomputing the expensive preprocessing steps.

These are pipeline intermediates, not a usable dataset on their own. They contain track ids and derived tags keyed to the official challenge catalog, and only mean something to the code that consumes them. There is no standalone loading recipe here on purpose: scripts/download_models.py in the repository fetches them at a pinned revision and puts them where the pipeline expects.

• Code, full pipeline and results: https://github.com/MaximeUM/recsys-challenge-2026
• All released artifacts: RecSys Challenge 2026 — FPMs_UMONS
• Paper: Picking is Not Ranking, and Explanation Quality Has Many Dimensions: Lessons for Conversational Music Recommendation. Maxime Manderlier, Fabian Lecron. RecSys Challenge 2026.

Contents

File	Size	What it is
content_desc_full.json	5.8 MB	per-track {sound | themes} text descriptors, distilled from the challenge's audio and lyrics embeddings — required by the reranker
pools/combined_pool_ctx1024_blindB.parquet	81 KB	top-200 candidate pool per Blind-B turn, from the four-channel RRF retrieval
pools/combined_pool_ctx1024_dev.parquet	15 MB	the same for the 8000-turn dev split

content_desc_full.json

{"desc": {"<track_id>": "sound: hard rock, metal | themes: rebellion, anger", ...}}

sound is the consensus community tags of the track's nearest neighbours in the audio (CLAP) embedding space; themes its closest entries in a lyrical-topic lexicon in the lyrics embedding space. The reranker reads these as words — the paper's finding is that the multimodal signal transfers as distilled text and not as raw vectors.

These are derived tags, not embeddings: the source vectors cannot be reconstructed from them.

Pools

Columns: session_id, turn, pool (JSON list of catalog track ids, top-200), plus gt_idx for the dev split. Pool construction is bit-reproducible, so these files match what the pipeline regenerates.

Everything here can also be rebuilt from your own copy of the challenge data (~25 min on one GPU, no training data needed) with the two commands documented in the repository.

Provenance and terms

These artifacts are derived from the TalkPlayData-Challenge datasets (talkpl-ai on the Hub), which remain the property of the challenge organizers. They are published here following the organizers' code-submission guidelines, which ask participating teams to share their weights, models and datasets on Hugging Face.

They are not a redistribution of the underlying data: no conversation text, no audio, no lyrics and no embedding vectors are included — only track ids and tags derived from them. Anyone using these files still needs the official datasets, and is bound by whatever terms the organizers set for them.

The code that produced these artifacts is MIT-licensed and lives in the GitHub repository linked above.