iaouali/hm-benchmark

H&M query–bundle benchmark

Canonical query/reference-positive data, organized like amazon-benchmark. Reference this repository by commit SHA, category (H_and_M), split, and candidate_id in experiment traces instead of republishing the data.

Split	Examples
shared_train	9,019
agent_dev	2,028
agent_hidden	1,960
Total	13,007

Each record contains an English shopping query and 3–7 reference article IDs, plus generation and judge metadata. Article IDs are zero-padded strings. positive_items and bundle_items contain the same reference set. These are observed reference positives, not exhaustive labels for all valid recommendations.

Construction

Users were assigned to shared_train/agent_dev/agent_hidden with a deterministic user-hash split (70%/15%/15%). The split is user-disjoint, not product-disjoint. All observed purchases are implicit positives; H&M has no source ratings, and a purchase is not proof of satisfaction.

Candidate bundles are mined from seven-day purchase windows. Deterministic checks require nonempty titles, descriptions of at least 15 characters, known product types, distinct product codes and product types, and a single merchandising group (Ladieswear/Divided merged). Dates have day-level resolution, so no unique purchase order within a day is established.

Luna (gpt-5.6-luna, a managed alias without a pinned model revision) judges bundle coherence with three votes, requiring at least two passes. Generated queries pass deterministic checks and an LLM query–bundle consistency check. This is fully automatic, with no human annotation. LLM judgments do not guarantee compatibility or factual correctness.

Candidate ordering and target counts by split and bundle size are fixed independently of agent performance. Targets match the Musical Instruments generation size profile; they are not a claim of matching its product distribution. No retrieval scores, agent wins, or prompt-optimization feedback selected these examples. No queries from the superseded Qwen H&M generation were reused. Exact normalized queries and identical positive sets are unique across the released data; related products and paraphrases can still cross splits.

Evaluation and leakage

shared_train purchases also supply tool training: this split is not an independent evaluation set. Use agent_dev for development. The historical name agent_hidden is retained, but publication means it is no longer a secret test set. Keep it excluded from training and tuning for held-out comparisons. A confidential future test requires a new protected split.

The current benchmark convention defines SetHit@k as macro-averaged recall: for each query, divide the number of distinct exact reference articles recovered in the top k by the reference-set size, then average across queries. Binary any-overlap is a separate metric. Older experiment summaries may use set_hit_at_k for binary any-overlap; do not combine those values. The evaluated agent currently requests k equal to reference size (3–7), not 20.

Provenance and privacy

The source is this third-party H&M tabular mirror, pinned to revision c335faf3386d3c760997c37ba4cc1b0d90265c1f. It was not independently authenticated against the original Kaggle archive. Source hashes, generation policy, aggregate split metadata, validation results, and per-split SHA-256 checksums are included under H_and_M/.

No customer identifiers, demographics, raw transaction histories, private user assignments, split key, or credentials are included. Candidate IDs are opaque bundle identifiers. This repository publishes benchmark examples, not model checkpoints or the raw catalog. Upstream source and model usage terms still apply; this release does not assert a new license over third-party material.