---
license: apache-2.0
task_categories:
- text-classification
- feature-extraction
tags:
- entity-resolution
- record-linkage
- amazon-ml-challenge
size_categories:
- 10M<n<100M
---

# Amazon ML Challenge 2026: Business Entity Resolution Dataset

This repository hosts the official dataset for the **Amazon ML Challenge 2026 — Business Entity Resolution Challenge**, packaged for high-speed multi-server distributed training, EDA, and evaluation.

Both the original **TSV files** and optimized **Parquet files (zstd-compressed)** are provided.

---

## 📊 Dataset Overview & Statistics

Total records across training and test sets: **~26.4 Million records**.

| Split | File | Format | Row Count | Uncompressed Size | Parquet Size | Key Fields |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| **Train** | `train_source1` | Reference (S1) | 2,206,821 | 200.3 MB | 77.3 MB | `entity_id`, `business_name`, `business_address`, `country` |
| **Train** | `train_source2` | Secondary (S2) | 5,034,616 | 466.6 MB | 188.1 MB | `entity_id`, `business_name`, `business_address`, `country` |
| **Train** | `train_source3` | Tertiary (S3) | 5,285,603 | 480.4 MB | 194.0 MB | `entity_id`, `business_name`, `business_address`, `country` |
| **Train** | `train_ground_truth` | Labels | 2,206,821 | 121.1 MB | 55.4 MB | `source1_entity_id`, `matched_entity_ids` |
| **Test** | `test_source1` | Reference (S1) | 1,732,544 | 166.9 MB | 62.2 MB | `entity_id`, `business_name`, `business_address`, `country` |
| **Test** | `test_source2` | Secondary (S2) | 4,887,273 | 485.9 MB | 189.1 MB | `entity_id`, `business_name`, `business_address`, `country` |
| **Test** | `test_source3` | Tertiary (S3) | 5,082,316 | 482.6 MB | 190.9 MB | `entity_id`, `business_name`, `business_address`, `country` |
| **Total** | | | **26,436,001** | **~2.5 GB** | **~957 MB** | |

---

## 🔍 Critical Structural Findings & EDA Highlights

1. **Strict 100% Country Partitioning:**
   - In ground truth (7.64M matched pairs), **0 cross-country matches** exist.
   - US records only match US records; India only matches India.
   - *Strategy:* Partition candidate blocking strictly by `country` to eliminate 50–70% of search space with zero recall loss.

2. **The France Domain Shift:**
   - **Train:** US (60%), India (40%)
   - **Test:** India (46.7%), US (38.3%), **France (15.0% - 259,452 S1 records)**
   - France is **unseen** in the training set. Models and pipelines must not hardcode country-specific assumptions.

3. **1-to-1 Target Match Constraint:**
   - Every S2 or S3 entity links to **at most ONE** S1 entity (0 multi-links).
   - Approximately 25-27% of S2 and S3 entities are non-matching distractors.

4. **Evaluation Metric:**
   - Macro-averaged $F_{0.5}$ score across all S1 entities:
     $$F_{0.5} = \frac{1.25 \times \text{Precision} \times \text{Recall}}{0.25 \times \text{Precision} + \text{Recall}}$$
   - Precision is weighted $2\times$ over Recall: false merges (false positives) penalize score much more than missed links.
   - Singletons (5.58% in train S1): Correctly predicting an empty match list yields $1.0$; a false match drops that entity to $0.0$.

---

## 🚀 Quick Start: Multi-Server Access & Usage

### 1. Fast CLI Download on Any Server

```bash
# Install huggingface_hub
pip install huggingface_hub

# Login (if private repo)
huggingface-cli login --token $HF_TOKEN

# Download entire dataset into ./data folder
huggingface-cli download akshatbakshi/amazon-ml-challenge-2026 --repo-type dataset --local-dir ./data
```

### 2. High-Performance Loading with Polars / DuckDB

Parquet files load in **< 1 second**:

```python
import polars as pl

# Load reference source 1 (Train)
s1 = pl.read_parquet("data/parquet/train/train_source1.parquet")

# Filter strictly by country
s1_us = s1.filter(pl.col("country") == "US")
s1_in = s1.filter(pl.col("country") == "India")

# Load ground truth
gt = pl.read_parquet("data/parquet/train/train_ground_truth.parquet")
```

With **DuckDB** (zero-copy SQL queries across millions of rows):

```python
import duckdb

# Direct SQL query on parquet without loading everything to RAM
con = duckdb.connect()
res = con.execute("""
    SELECT s1.country, COUNT(*) 
    FROM 'data/parquet/train/train_source1.parquet' s1
    GROUP BY s1.country
""").df()
print(res)
```

---

## 📁 Repository Directory Structure

```
.
├── README.md
├── parquet/
│   ├── train/
│   │   ├── train_source1.parquet
│   │   ├── train_source2.parquet
│   │   ├── train_source3.parquet
│   │   └── train_ground_truth.parquet
│   └── test/
│       ├── test_source1.parquet
│       ├── test_source2.parquet
│       └── test_source3.parquet
└── tsv/
    ├── train/
    │   ├── train_source1.tsv
    │   ├── train_source2.tsv
    │   ├── train_source3.tsv
    │   └── train_ground_truth.tsv
    └── test/
        ├── test_source1.tsv
        ├── test_source2.tsv
        └── test_source3.tsv
```