SM-RS is, to our knowledge, the only public recommender-systems dataset linking users' self-declared propensities toward beyond-accuracy objectives with contextual impressions, item selections, and explicit perceived-quality judgments. This repository is the benchmark code: data loaders and the canonical evaluator for each task, so everyone reports comparable numbers. The data lives on Hugging Face; the leaderboard lives in the dataset card.

Dataset & leaderboard: 🤗 pdokoupil/SM-RS Cite BOTH the SM-RS 2.0 (TORS'26) and SM-RS (SIGIR'24) papers — see Citing.

Install

pip install sm-rs                 # core (numpy / pandas / scikit-learn)
pip install "sm-rs[lstm]"         # + TensorFlow, for the optional LSTM baseline

Quick start (Task 1: propensity estimation)

import numpy as np
from smrs.tasks import task1_propensity as t1

# reproducible 80/20 split (seed 2024), reading a local SM-RS copy for now
X_train, X_test, y_train, y_test = t1.split(data_dir="path/to/sm-rs")

# ... train your estimator, produce an (N, 4) array of [rel, div, nov, exp] ...
predictions = np.full((len(y_test), 4), 0.25)        # placeholder

print(t1.evaluate(predictions, y_test))   # {'MAE': ..., 'MSE': ..., 'KLDiv': ...}

The six tasks

All six draw on the same dataset; they differ in what you predict and how it's scored. The canonical evaluator for each lives in smrs.tasks.*.

¹ Task 2's metric layer is implemented; turning a top-k list into achieved objective proportions needs the derived matrices (see Data) and lands with the source→rating-matrix builder.

Data

Two layers:

1. Core tables (the collected study data, CC-BY, hosted on Hugging Face): behaviors, propensities, objective_perceptions, criteria_values, comparative_diversity, users, movies, books. Items are referenced by ID.

   Auto-downloaded from the Hub and cached — no manual download:

   from smrs import data
   df = data.load("propensities")                 # downloads from HF, cached
   # offline / local copy (e.g. OSF download):
   df = data.load("propensities", data_dir="path/to/sm-rs")   # or set $SMRS_DATA_DIR
   

   Users of the 🤗 datasets library can equivalently do load_dataset("pdokoupil/SM-RS", "behaviors").

2. Derived matrices — recomputed locally, not downloaded. The list-scoring tasks (2, 3) need per-item / per-pair artifacts (relevance via item-item, intra-list diversity via a distance matrix, novelty via popularity). Rather than ship multi-GB blobs — or redistribute the third-party catalogs they come from — the benchmark recomputes them from a rating matrix you build from your own download of the public source datasets (movies: MovieLens 25M — the "Latest" snapshot at collection time — plus MovieLens Tag Genome 2021; books: goodbooks-10k):

   from smrs import derived
   art = derived.build_artifacts(rating_matrix_movies)   # {item_item, distance_matrix, mean_popularities}
   

   derived provides the deterministic pieces: popularity, cosine_distance (1 − cosine over item rating-vectors), and ease_item_item (EASE^R closed form, used as the relevance model). This keeps the benchmark lightweight and license-clean. The bit-exact original artifacts are archived on OSF (v2) for strict reproduction.

   Get the source datasets with the bundled fetcher (downloads from the official hosts — GroupLens, the goodbooks repo — under their licenses; or place them there manually):

   smrs-fetch --list                      # show sources + licenses, no download
   smrs-fetch --dest ./sm-rs-sources      # download MovieLens 25M, Tag Genome 2021, goodbooks-10k
   

Why recompute? MovieLens may not be redistributed, and a 5 GB download hurts adoption. You obtain MovieLens/goodbooks under their own licenses; we ship only the study data, the id-maps (movies.json: movieId→imdbId; books.json: book_index→goodreads_id), and the recompute code. The canonical scorer is pinned (sources above; positive feedback = rating ≥ 3; EASE λ) so results stay comparable. For strict reproduction of the paper's numbers, use the OSF artifacts.

Reproduction check

examples/reproduce_perceived.py reproduces the paper's Linear Regression baseline for Tasks 5 & 6 (the one baseline needing no derived/source data), scored with this package's evaluators — MAE matches the paper exactly:

SMRS_DATA_DIR=/path/to/sm-rs python examples/reproduce_perceived.py

Submitting to the leaderboard

Self-service (no submission server): run the canonical evaluate() for a task, then open a PR adding your row to the leaderboard in the dataset card, with a link to reproduce. The shipped baselines are the rows to beat.

Citing

Please cite both papers (GitHub's "Cite this repository" reads CITATION.cff):

@article{dokoupil2026smrs2,
  author  = {Dokoupil, Patrik and Peska, Ladislav},
  title   = {SM-RS 2.0: User-perceived Qualities of Single- and Multi-Objective Recommender Systems},
  journal = {ACM Transactions on Recommender Systems},
  volume  = {4}, number = {3}, year = {2026},
  doi     = {10.1145/3754459}
}
@inproceedings{dokoupil2024smrs,
  author    = {Dokoupil, Patrik and Peska, Ladislav and Boratto, Ludovico},
  title     = {SM-RS: Single- and Multi-Objective Recommendations with Contextual Impressions and Beyond-Accuracy Propensity Scores},
  booktitle = {Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  series    = {SIGIR '24}, pages = {988--995}, year = {2024},
  doi       = {10.1145/3626772.3657863}
}

License

Code: MIT (see LICENSE). Data: CC-BY-4.0 (see the dataset card).