Hplss/wb-review-dataset

WB Review Dataset

A Russian-language product review dataset for sentiment analysis, collected from Wildberries (a large Russian e-commerce marketplace).

~27,000 substantive, human-written product reviews spanning 7 product categories, filtered for length and quality rather than raw volume. Built to provide sentiment analysis training data with real semantic content (multi-sentence reviews with actual arguments), not one-line "great, buy it" / "bad, don't buy" reviews.

Collection code and full methodology: github.com/ProphetSunboy/wb-review-dataset

Also available on Kaggle.

Schema

Column	Description
review_id	Pseudonymous sequential integer, unique per review. Not the original platform review ID.
product_id	Pseudonymous sequential integer, shared by all reviews of the same underlying product. Use this for grouped train/test splits to avoid leaking the same product across splits.
category	Internal category key (e.g. electronics).
category_label	Human-readable category name in Russian.
rating	1-5 stars, as given by the actual reviewer. This is the ground-truth sentiment signal.
text	Main review text.
pros	Separate "pros" field, if the reviewer filled it in (often empty).
cons	Separate "cons" field, if the reviewer filled it in (often empty).
date	Review date (YYYY-MM-DD).

Loading

from datasets import load_dataset

ds = load_dataset("Hplss/wb-review-dataset")

or directly with pandas:

import pandas as pd

df = pd.read_csv("hf://datasets/Hplss/wb-review-dataset/filtered.csv")

Filtering criteria

• Minimum 20 words, maximum 300 words (combined across text + pros + cons)
• At least 70% Cyrillic characters
• At most 50% uppercase characters
• Not emoji-only
• Deduplicated by review ID and by exact combined text (catches templated/copy-paste reviews). Roughly 36% of unique raw reviews were exact-text duplicates of another review, mostly short generic ones already excluded by the length filter regardless.
• Phone numbers and email addresses redacted

Category breakdown

27,248 reviews total across 7 categories:

Category	Reviews
Cosmetics and personal care	6,008
Electronics	4,733
Clothing and footwear	4,534
Sports and outdoors	3,312
Kids' products	3,110
Home goods	2,852
Food and groceries	2,699

Known limitations

Rating as a sentiment proxy is imperfect. The star rating is a real signal from real customers, not inferred or synthetic, but a small fraction of reviews show visible disagreement between the rating and the text (e.g. a clearly negative review left with 5 stars), most likely from reviewer error rather than sarcasm. No automated mismatch flag is applied — an earlier keyword-based heuristic was tested and dropped as unreliable.

Sentiment distribution is naturally skewed toward positive, not an artificially balanced split. An informal 50/15/35 (positive/neutral/negative) split was set as a rough design goal before collection started, mainly to avoid ending up with an all-positive dataset, not a figure derived from any external benchmark. The actual distribution among reviews that pass the length filter: roughly 64% are 5-star, 11% are 4-star, 8% are 3-star, and 17% are 1-2 star. This was not artificially rebalanced, to avoid teaching a model a spurious category-sentiment correlation instead of the real signal in the text.

Length filtering removes most raw scraped reviews. Roughly 72% of unique raw reviews are under 10 words. This dataset trades volume for substance by design.

Identifiers are pseudonymized and cannot be traced back to the original listings.

Date range spans 2015-2026, heavily concentrated in recent years. Over 99% of reviews are from 2022 onward. A small number of older reviews are retained from Wildberries' persisted review history on long-lived product listings — the platform keeps reviews visible on a product page indefinitely, so a handful of old reviews can surface alongside recent ones for products that have been listed for years.

License

CC BY-NC-SA 4.0 — attribution required, non-commercial use only, derivatives must share alike. See the repository's LICENSE-DATA file for full terms.

Source and disclaimer

Reviews were collected from Wildberries' publicly accessible API endpoints (no login or API key required). This project is not affiliated with, endorsed by, or connected to Wildberries in any way. Provided as-is for research, educational, and non-commercial machine learning use.