RecoAtlas artifacts

Canonical artifacts for two RecoAtlas environments. Each environment has the same layout and contains only public query/reference data, the five frozen files required to serve the three recommendation tools, and audited leaderboard results. Raw interactions, user identifiers, split keys, mining caches, training checkpoints, and obsolete run directories are deliberately excluded.

Environments

Environment	Catalog items	shared_train	agent_dev	agent_hidden	Leaderboard arms
amazon-musical-instruments	62,961	8,937	2,028	1,960	10
hm-clothing	104,993	9,019	2,028	1,960	10

The Musical Instruments training export contains 8,937 records: 9,019 were generated and 82 were removed by the final positive-set overlap filter. This repository reports the published file count rather than the pre-filter target.

Uniform layout

<environment>/
├── data/
│   ├── shared_train.jsonl
│   ├── agent_dev.jsonl
│   └── agent_hidden.jsonl
├── tools/
│   ├── catalog/{product_texts,subcategories}.json
│   ├── embeddings/{item_embeddings,complementarity_embeddings}.pt
│   └── retrieval/retrieval_model.pt
└── benchmark/
    └── leaderboard/
        ├── metrics.json
        └── arms/<arm>/rollouts.jsonl.gz

A repository commit pins the data, tools, and results together.

Load query data

from datasets import load_dataset

mi = load_dataset("iaouali/recoatlas", "amazon-musical-instruments")
hm = load_dataset("iaouali/recoatlas", "hm-clothing")

Every query record has exactly four fields: candidate_id, partition, query, and positive_items. The agent receives the query and K = len(positive_items); reference IDs stay with the evaluator. The references are observed behavioral positives, not exhaustive judgments of every acceptable recommendation.

Download one environment's tools

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="iaouali/recoatlas",
    repo_type="dataset",
    allow_patterns=["amazon-musical-instruments/tools/**"],
    local_dir="recoatlas-artifacts",
)

The tools use Qwen/Qwen3-Embedding-0.6B at revision 97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3. product_texts.json and subcategories.json form the catalog; item_embeddings.pt supports relevance and near-duplicate checking; complementarity_embeddings.pt supports complement expansion; and retrieval_model.pt is the learned query-side retrieval checkpoint.

Leaderboard metric

The leaderboard uses fractional reference recovery:

SetHit@K = |G ∩ R| / |G|, with K = |G|.

R contains distinct valid recommendations in the first K evaluated positions. Scores are macro-averaged over all 1,960 expected agent_hidden queries; missing or failed reports receive zero. This is not binary any-overlap. Coverage and report-validity diagnostics are stored next to SetHit in benchmark/leaderboard/metrics.json; they are not merged into the SetHit score.

Sources, split policy, and limitations

The Musical Instruments environment derives from Amazon Reviews 2023 review interactions and catalog metadata; reviews rated at least four stars are treated as positive. The H&M Clothing environment derives from H&M Personalized Fashion Recommendations purchases and catalog metadata; every observed purchase is treated as positive because the source has no ratings.

Users were deterministically partitioned before benchmark construction. The partitions are user-disjoint, not product-disjoint or chronological. Candidate sets use seven-day temporal windows and contain three to seven products. They were filtered for bundle coherence and query-set consistency using GPT-5.6 Luna. agent_hidden is the historical held-out evaluation split, but public release means it is no longer secret. Keep it out of training and tuning when reporting held-out results. Temporal co-occurrence is behavioral evidence rather than proof of a single order, exhaustive relevance, or technical compatibility. See LICENSE.md and NOTICE.md for third-party terms.