srpone/zoowork-shoprank-bench

ShopRank-Bench

The benchmark of ZooWork-ShopRanker: An Open, Preference-Aligned E-Commerce Reranker.

🚀 An optimized commercial version of the ZooWork-ShopRanker rerankers is available through the zoodata.ai API.

Quick start

from datasets import load_dataset

preference = load_dataset("srpone/zoowork-shoprank-bench", "preference", split="train")
ahp        = load_dataset("srpone/zoowork-shoprank-bench", "ahp", split="train")
budget     = load_dataset("srpone/zoowork-shoprank-bench", "budget", split="train")

row = preference[0]
row["query"], row["tier"], row["preferred_structured"], row["rejected_structured"]

Tested with datasets 5.0.1. Older 3.x releases cannot read the schema the Hub infers for the nested judge_stances field; upgrade with pip install -U datasets.

A reranker is correct on a pair when it scores the preferred product above the rejected one; report accuracy per tier and per text format (*_structured / *_natural), with query-clustered bootstrap intervals.

Evaluate with LookBench

evaluate.py in this repo runs the LookBench reranking harness on all 10,511 preference pairs, in both text formats, and prints accuracy per tier with a query-clustered 95% CI:

pip install "look-bench[reranking]>=0.4.0"
curl -LO https://huggingface.co/datasets/srpone/zoowork-shoprank-bench/resolve/main/evaluate.py
python evaluate.py --model srpone/zoowork-shopranker-0.6b

ShopRank-Bench pairwise accuracy (%), srpone/zoowork-shopranker-0.6b
| Product text | Overall [95% CI] | Gold | Silver | Bronze | Ties |
|---|---|---|---|---|---|
| structured | 76.7 [75.7, 77.7] | 91.0 | 80.0 | 67.1 | 253 |
| natural | 75.5 [74.4, 76.5] | 91.2 | 78.3 | 65.6 | 159 |

The script uses the installed look_bench package; without it, it clones LookBench into ./look-bench (pass --lookbench to reuse a checkout). Use look-bench 0.4.0 or later: earlier PyPI releases lack the reranking track and overwrite Hugging Face datasets. For a ZooWork-ShopRanker repo it loads the base weights and the LoRA adapter from that repo; --model Qwen/Qwen3-Reranker-4B --adapter <path> evaluates your own adapter. Options: --formats, --tie_credit (ties count as wrong by default), --output results.json, --dump_predictions preds_{format}.jsonl. LookBench scores one pair per forward pass, so results can differ from the paper's batched scorer by up to ~0.6 points on a tier (0.2 overall).

A contamination-limited benchmark for judged e-commerce preference — which of two comparably relevant products a shopper is better served by — rather than topical relevance. 13,313 pairs over three tracks, built from private search traffic that has never been published.

Open rerankers score far lower here than their relevance benchmarks suggest, and the gap concentrates on pairs where judges disagree.

Tracks

Track	Pairs	Label source	Purpose
preference	10,511	Cross-family LLM judge panel, tiered by agreement	Headline metric
ahp	1,500	Hand-designed attribute-priority hierarchy	Diagnostic: does the model respect attribute priority?
budget	1,302	Programmatic price ≤ budget oracle	Diagnostic: does the model respect an explicit constraint?

The preference track comes from 2,991 distinct queries (3.5 pairs per query on average), so pairs are not independent — cluster by query when bootstrapping or your intervals will be too narrow. Reported design effects are 1.2–1.4.

Agreement tiers

Each pair is judged by three model families (qwen35-122B, gemma-4-31B-it, deepseek-v4-pro) in both presentation orders. A judge either commits to a candidate or abstains; pairs with an outright conflict are discarded, so committed judges never disagree within a released record.

Tier	Judges committed	Pairs
gold	3 / 3	1,843
silver	2 / 3	4,445
bronze	1 / 3	4,223

Report per tier, not just in aggregate. Every reranker we measured loses 20–25 points from gold to bronze, and gold sits near 96% for everything — an aggregate score is dominated by pairs nothing gets wrong. The panel is not uniformly active: the bronze tier is 82% Gemma-decided.

Fields

preference.jsonl

• query — the shopper's search string
• preferred_structured / rejected_structured — attribute-schema product text
• preferred_natural / rejected_natural — the same two products as prose
• tier — gold | silver | bronze
• judge_stances — {judge: "preferred" | "tie"} per family

ahp.jsonl adds query_constraints, hierarchy_high, hierarchy_low. budget.jsonl adds budget, preferred_price, rejected_price, slice.

The preferred product is always in the preferred_* field, so shuffle or randomise presentation order yourself — do not feed the fields positionally.

Dual rendering

Every preference and ahp pair ships twice: once as a structured attribute schema and once as natural-language prose, with query, products and label held fixed. This isolates serialization format from content. The prose is LLM-rendered from the same attributes, not scraped copy. The budget track is structured only.

What this benchmark does not measure

Labels come from an LLM judge panel. They are not human-verified, and no click, purchase or interleaving signal validates that the panel tracks what shoppers actually choose. A blind audit by a fourth, held-out model family agreed with the released label on 43/45 tier-stratified pairs (15/15 gold, 15/15 silver, 13/15 bronze) — a consistency check, not human validation.

Because candidates are retrieved rather than curated, a minority of pairs are separated by product type rather than preference, and some catalogue attributes are noisy. These pairs are correctly labelled but easy, and are part of why the gold tier sits so high.

The category mix leans toward apparel.

Contamination and privacy

Built from private traffic never published, so it cannot appear in any base model's pretraining corpus. Disjoint from our training corpus: non-overlapping traffic windows, zero shared query strings, zero shared triples. Internally non-redundant: no triple occurs twice, no pair recurs with its label reversed, no product is paired with itself.

Not solvable by surface heuristics — always picking the longer text reaches 53.3%; always picking the cheaper reaches 58.7% (BM25 is 56.2%), against 71.6% for the weakest learned model.

No user, account, session, device, location, timestamp or behavioural field is released, and queries are not linked by user or session. All 2,991 distinct queries were screened for direct identifiers (email, phone, address, postcode, government ID, long digit strings, social handles, URLs); none were found. Product text is rendered from public catalogue attributes.

Reproducing the release

python scripts/release/prepare_bench_release.py \
    --preference    <internal>/hard_v2_release.jsonl \
    --preference_nl <internal>/hard_v2_release_nl.jsonl \
    --ahp           <internal>/ahp_eval_diverse.jsonl \
    --budget        <internal>/price_eval_normalized.jsonl \
    --out_dir       data/release/shoprank_bench

The builder recomputes the three tiers from the per-judge verdicts, whitelists released fields, and aborts on any track-size, tier-count, duplicate or identifier-scan failure.

Citation

@article{xue2026zooworkshopranker,
  title   = {ZooWork-ShopRanker: An Open, Preference-Aligned E-Commerce Reranker},
  author  = {Xue, Siqiao and Liu, Shuxuan and Hu, Ning},
  journal = {arXiv preprint arXiv:2609.31002},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.31002}
}