README.md

---
license: apache-2.0
language:
- en
pretty_name: E-commerce Relevance Evaluation Scoring Suite
tags:
- e-commerce
- information-retrieval
- reranking
- ranking
- evaluation
- search
- relevance
- ndcg
- product-search
task_categories:
- text-classification
- sentence-similarity
size_categories:
- 10K<n<100K
---

# ERESS: E-commerce Relevance Evaluation Scoring Suite

## Dataset Summary

**ERESS (E-commerce Relevance Evaluation Scoring Suite)** is an evaluation dataset for **product discovery / e-commerce search reranking**. It contains **~4.7k unique queries** and **~72k labeled (query, product) pairs**, with **graded relevance** designed to reflect real shopping behavior and modern “assistant-style” query intent (e.g., utility/task framing, gift/audience constraints, attribute-heavy queries).

ERESS is built to stress-test rerankers under common real-world failure modes:
- **Hard negatives** (semantically close but wrong items)
- **Lexical confounders** (high token overlap, wrong intent)
- **Near-substitutes** (plausible alternatives that violate a key constraint like size/color/compatibility)

The dataset is intended primarily for **holistic evaluation** of reranking models using **nDCG** (e.g., nDCG@5, nDCG@10) with graded relevance.

---

## Motivation

Public relevance sets such as Amazon-ESCI and WANDS can under-represent modern e-commerce traffic, especially:
- Gift and audience-specific constraints (“gift for…”, “for my dad who…”, occasions)
- Utility/task queries (“fix squeaky door”, “organize cables”, “reduce glare”)
- Highly attribute-constrained queries and conversational phrasing

ERESS is explicitly stratified to better cover these intents and to include harder confounders that survive high-recall retrieval.

---

## Supported Tasks and Leaderboard-Style Use

**Primary:** Reranking / relevance estimation for product search  
- Input: a query and a candidate product (title/description/attributes)  
- Output: a relevance score (graded)

**Recommended metrics:**
- **nDCG@k** (primary; graded relevance)
- Optionally: MRR@k or Precision@k (secondary)

---

## Dataset Composition

### High-level stats
- **Unique queries:** ~4.7k  
- **Labeled pairs:** ~72k (query, product)  
- **Labels:** graded relevance (scalar in \[0, 1\] as described in the accompanying project materials)

### Query intent families
ERESS queries are designed to reflect a spectrum of shopping intents, including:
1. **Attribute-rich** (high feature density; size/material/compatibility constraints)
2. **Navigational** (brand/store/product-line seeking)
3. **Gift & audience-specific** (recipient + occasion)
4. **Generic** (category-level intent with minimal attributes)
5. **Utility** (task/solution framing)
6. **Short & Books** (very short head queries + targeted books)

---

## How ERESS Was Constructed (Overview)

ERESS is derived from a modern e-commerce relevance pipeline built for high-recall retrieval + reranking evaluation:

1. **Query generation & coverage control**
   - Synthetic queries are generated under a constrained protocol and stratified across intent families.
   - Queries are embedded and clustered to avoid over-representing paraphrases and to preserve semantic diversity.

2. **High-recall candidate retrieval**
   - Candidates are retrieved via a dual-index strategy to improve recall across query types:
     - **Index A:** title-focused view (better for short/navigational)
     - **Index B:** full-text view (better for attribute-rich/utility)
   - Candidate sets are merged and deduplicated per query.

3. **Graded relevance annotation**
   - Each (query, product) pair is labeled using an ensemble (“council”) of LLMs.
   - Annotation uses structured prompts that emit a discrete label as the first token; first-token logits are used as calibrated signals for scoring.

4. **Quality control & leakage prevention**
   - Candidate-level pruning to remove trivial irrelevance
   - Deduplication across splits (no duplicate query strings and no duplicate product IDs leaking across train/eval regimes)
   - Semantic decontamination to reduce benchmark/train overlap via embedding similarity thresholds

---

## Dataset Splits

ERESS is intended as an **evaluation suite**. If the hosted dataset includes splits, they typically correspond to evaluation partitions (e.g., `test` and/or multiple slices). If you create your own splits, ensure:
- no duplicate queries across splits
- no duplicate product IDs across splits (when evaluating generalization)
- avoid semantic contamination between training and ERESS queries (nearest-neighbor overlap)

---

## Usage

### Loading
```python
from datasets import load_dataset

ds = load_dataset("thebajajra/eress")
print(ds)
print(ds[list(ds.keys())[0]][0])