Jonathan Politzki
Add benchmarks.json: benchmark metadata beside the results, so the UI stays dumb
772850b | { | |
| "version": "2026-09-12", | |
| "status": "seed", | |
| "note": "Seed release. Every row here is an internal Jean run with its experiment id and commit named. No row is marked verified until the rerun script for its snapshot is public and the number has been reproduced from it. Rows are never edited; a rerun adds a row.", | |
| "eval_code": { | |
| "repo": "Jean HFM research repository (private; rerun scripts to be published per snapshot)", | |
| "commit": "682e914" | |
| }, | |
| "primary_metric": "recall@10", | |
| "metrics": [ | |
| { | |
| "id": "recall@10", | |
| "label": "Recall@10", | |
| "kind": "ranking", | |
| "higher_is_better": true, | |
| "definition": "Fraction of persons whose held-out next item is in the model's top ten out of the whole catalog." | |
| }, | |
| { | |
| "id": "ndcg@10", | |
| "label": "NDCG@10", | |
| "kind": "ranking", | |
| "higher_is_better": true, | |
| "definition": "Same target, discounted by rank within the top ten." | |
| }, | |
| { | |
| "id": "js_divergence", | |
| "label": "JS divergence", | |
| "kind": "diversity", | |
| "higher_is_better": false, | |
| "definition": "Jensen-Shannon divergence between the real held-out next-item distribution and the simulator's sampled distribution, over a fixed partition of the catalog (the first-level semantic-ID code). Floor: a collapsed-mode simulator that always emits the most popular item." | |
| }, | |
| { | |
| "id": "entropy_ratio", | |
| "label": "Spread ratio", | |
| "kind": "diversity", | |
| "higher_is_better": null, | |
| "definition": "Entropy of the simulated distribution divided by entropy of the real one. One is the target; below one is mode collapse, above one is noise." | |
| }, | |
| { | |
| "id": "nll", | |
| "label": "Log loss", | |
| "kind": "calibration", | |
| "higher_is_better": false, | |
| "definition": "Negative log probability the simulator assigns to the held-out next item, averaged over persons. Baseline: popularity." | |
| } | |
| ], | |
| "verification_levels": { | |
| "verified": "Rerun by Jean from the public script against the frozen snapshot, with the seed and commit named. Third parties can reproduce it.", | |
| "internal": "Jean's own run, experiment id and commit named, rerun script not yet public.", | |
| "reported": "Submitted by the system's authors and not rerun. Listed for completeness, never ranked above a rerun row." | |
| }, | |
| "benchmarks": [ | |
| { | |
| "id": "amazon2014-beauty", | |
| "name": "Amazon Reviews 2014, Beauty, 5-core", | |
| "short": "Amazon Beauty", | |
| "persons": 22363, | |
| "persons_scored": 22363, | |
| "items": 12101, | |
| "events": 198502, | |
| "cutoff": "2014-07-23", | |
| "license": "no explicit licence; SNAP page states a citation requirement (McAuley et al. 2015); academic use (verified live 2026-09-04). Not redistributed by Jean; obtain the files from the source.", | |
| "source_url": "https://snap.stanford.edu/data/amazon/productGraph/categoryFiles/reviews_Beauty_5.json.gz", | |
| "protocol": "Per person, events sorted by time. The last event is the test target, the second to last is validation, the rest is history. Persons with fewer than five events are dropped. History capped at 20. Full-catalog ranking inside the category, no sampled negatives. Items already in the person's history are masked out.", | |
| "published_reference": "TIGER (arXiv 2305.05065, Table 1) reports SASRec Recall@10 0.0387 on this file. Jean's number is not a win over it: this protocol trains with a full-softmax loss rather than one sampled negative and masks the person's own history out of the candidates, and the literature puts the loss change alone at about 2x on Beauty (arXiv 2309.07602). The published figure is the calibration target, not a competitor row.", | |
| "confounds": [] | |
| }, | |
| { | |
| "id": "steam-2018", | |
| "name": "Steam reviews (Kang and McAuley, 2018)", | |
| "short": "Steam", | |
| "persons": 281678, | |
| "persons_scored": 20000, | |
| "items": 13044, | |
| "events": 3559466, | |
| "cutoff": "2018-01-05", | |
| "license": "unverified (academic use); cite Kang & McAuley, \"Self-Attentive Sequential Recommendation\", ICDM 2018. Not redistributed by Jean; obtain the files from the source.", | |
| "source_url": "https://cseweb.ucsd.edu/~jmcauley/datasets.html#steam_data", | |
| "protocol": "Same leave-one-out protocol as the Amazon snapshot. Items with fewer than five events are dropped. 20,000 persons drawn with a fixed seed are scored; the interval is a 2,000-sample bootstrap clustered by person.", | |
| "published_reference": null, | |
| "confounds": [ | |
| "Steam dates are day resolution.", | |
| "10.7 percent of events were exact duplicates and were dropped." | |
| ] | |
| }, | |
| { | |
| "id": "amazon2014-five-category-transfer", | |
| "name": "Amazon Reviews 2014, five categories, strict transfer", | |
| "short": "Amazon 5-cat. transfer", | |
| "persons": 106189, | |
| "persons_scored": 8348, | |
| "items": 69629, | |
| "events": 1160045, | |
| "cutoff": "2014-07-23", | |
| "license": "no explicit licence; SNAP page states a citation requirement (McAuley et al. 2015); academic use (verified live 2026-09-04). Not redistributed by Jean; obtain the files from the source.", | |
| "source_url": "https://snap.stanford.edu/data/amazon/productGraph/categoryFiles/reviews_Beauty_5.json.gz", | |
| "protocol": "One Amazon account observed in five product categories. For the held-out persons, every event in the target category is deleted from training. The model then scores the person's next item in that category from their history in the other four. Own history masked. The stranger control scores the same model from another held-out person's history.", | |
| "published_reference": null, | |
| "confounds": [ | |
| "Held-out persons stay in training with as few as three events in their other categories.", | |
| "This run predates the cross-domain duplicate fix (L-24); the five categories here share no catalog, so the leak does not apply, but the row is quoted with that history." | |
| ] | |
| }, | |
| { | |
| "id": "amazon2014-taste-transfer", | |
| "name": "Amazon Reviews 2014, taste categories, strict transfer, deduplicated", | |
| "short": "Amazon taste transfer", | |
| "persons": 141000, | |
| "persons_scored": 2686, | |
| "items": 33518, | |
| "events": 2023641, | |
| "cutoff": "2014-07-23", | |
| "license": "no explicit licence; SNAP page states a citation requirement (McAuley et al. 2015); academic use (verified live 2026-09-04). Not redistributed by Jean; obtain the files from the source.", | |
| "source_url": "https://snap.stanford.edu/data/amazon/productGraph/categoryFiles/reviews_Movies_and_TV_5.json.gz", | |
| "protocol": "One account across movies, CDs, video games and digital music, with the same album filed under two categories collapsed to one catalog position before anything is trained. Target category deleted from training for held-out persons; scored from the other categories. Own history masked. 20-core items, so absolute numbers are not comparable to the 5-core snapshots.", | |
| "published_reference": null, | |
| "confounds": [ | |
| "The old-only and per-category slices of this run are small; only the pooled strict transfer cell is shown.", | |
| "20-core is a cost decision, so these numbers are not comparable to the 5-core Amazon rows." | |
| ] | |
| }, | |
| { | |
| "id": "douban-tridomain-transfer", | |
| "name": "Douban movie, book and music, strict transfer", | |
| "short": "Douban transfer", | |
| "persons": 25953, | |
| "persons_scored": 12139, | |
| "items": 33957, | |
| "events": 6959199, | |
| "cutoff": "2017-10-26", | |
| "license": "none stated in the repo; academic use, cite Song et al., \"Session-Based Social Recommendation via Dynamic Graph Attention Networks\", WSDM 2019. Not redistributed by Jean; obtain the files from the source.", | |
| "source_url": "https://github.com/DeepGraphLearning/RecommenderSystems", | |
| "protocol": "One account observed natively in three media domains. For the held-out persons, every event in the target domain is deleted from training. The model then scores that person's next item in the domain it never saw them in, from their history in the other two. Own history masked. The stranger control scores the same model from another held-out person's history, permuted within the target domain so the domain prior is held constant. The training-panel rows keep everything fixed and train on a seeded fraction of the persons.", | |
| "published_reference": null, | |
| "confounds": [ | |
| "Douban ratings are ratings, not plays; timestamps are day resolution.", | |
| "The panel selects the 26,342 persons present in all three domains out of 94,890 movie users.", | |
| "Item ids carry no metadata, so item is the only granularity." | |
| ] | |
| } | |
| ] | |
| } | |