Acknowledge the W3 Dataset Terms of Use
The W3 dataset is released exclusively for academic, non-commercial research. By requesting access you confirm that you have read the paper's Ethical Use of Data and Informed Consent statement (Appendix G) and agree to the conditions below.
Please provide accurate contact information so the maintainers can reach you if any issue (e.g., privacy concern, data retraction) arises.
Log in or Sign Up to review the conditions and access this dataset content.
W³: A News-to-Alpha Benchmark for Utility-Driven Web3 Recommendation and Trading Backtesting
Source platform: CoinMeta (Bijie, a leading Web3 information aggregator) Time span: January 1, 2022 — January 1, 2026 (4 years)
TL;DR
W³ is the first benchmark dedicated to Web3 news recommendation that goes beyond ranking accuracy. It links a real-world industrial-grade news platform to tick-level cryptocurrency market data and a realistic trading backtesting protocol, enabling researchers to evaluate recommendation models by their realized financial utility (Sharpe ratio, max drawdown, alpha) — not just CTR.
| # Articles | 732,526 |
| # Active users | 9,732 |
| # Keywords | 7,050 |
| # Like / Collect / Comment / Dislike | 61,102 / 189,368 / 22,386 / 4,253 |
| Avg. title / abstract / body length | 46.83 / 126.43 / 416.32 words |
| Languages | English & Chinese (en, zh) |
Supported Tasks
1. Personalized News Recommendation
Standard ranking task using user interaction histories. Reported metrics: AUC, MRR, NDCG@{10,20,50}, Recall@{10,20,50}, Precision@{10,20,50}.
2. Quantitative Trading Backtesting (Novel)
Translates the top-K recommended news for each user into executable sentiment factors via FinBERT, feeds them into an LLM trading agent (with a Market Analysis Memo), and simulates execution under realistic frictions (fees, slippage). Reported metrics: Sharpe Ratio, Max Drawdown (MDD), Profit Factor, Alpha.
Dataset Structure
The dataset is organised as one article table + five behavior tables + impression logs with embedded splits, mirroring the production schema of the source platform.
articles config — cms_article
| Field | Type | Description |
|---|---|---|
id |
int64 | Unique article identifier |
category_id |
int64 | Category identifier for topic classification |
title |
string | Article headline |
ai_synopsis |
string | AI-generated summary |
synopsis |
string | Human-written abstract |
currency_id |
int64 | Associated cryptocurrency identifier |
keywords |
string | Comma-separated keyword tags |
is_show |
bool | Visibility flag (1: visible, 0: hidden) |
content |
string | Cleaned article body (HTML stripped, URLs removed) |
create_time |
timestamp[ms, UTC] | Publication time |
update_time |
timestamp[ms, UTC] | Last modification time |
language |
string | Language code (en / zh) |
data_status |
int32 | Internal validation status |
visit_rank |
float32 | Popularity score based on visit count |
like / collect / comment / dislike / click configs — behavior tables
All five behavior tables share the same schema:
| Field | Type | Description |
|---|---|---|
create_by |
int64 | Anonymized user identifier (SHA-256 + salt) |
article_id |
int64 | Foreign key to cms_article.id |
create_time |
timestamp[ms, UTC] | Timestamp of the interaction |
impressions config — pre-built splits
Impression logs formatted as (user_id, article_id, timestamp, label). The temporal 80 / 10 / 10 split (described in §D.2.4 of the paper) is materialised as three Hugging Face splits — train, validation, test — to enable one-line loading.
Quick Start
from datasets import load_dataset
# Load the article corpus
articles = load_dataset("jining-luan/W3-NewsAlpha", "articles", split="full")
print(articles[0]["title"])
# Load the favorite/collect behavior
collect = load_dataset("jining-luan/W3-NewsAlpha", "collect", split="full")
# Load the pre-split impression logs
impressions = load_dataset("jining-luan/W3-NewsAlpha", "impressions")
train_set = impressions["train"]
test_set = impressions["test"]
See examples/load_with_datasets.py for a full walkthrough and examples/backtest_demo.ipynb for a minimal sentiment-factor → trading-PnL pipeline reproducing parts of Table 4–6 in the paper.
Data Collection & Curation
- Source. Direct backend access to CoinMeta, spanning mainstream financial news, exchange announcements and project updates over four years.
- Filtering. Records flagged invalid by the platform's internal moderation system were removed (promotional spam, low-quality content, misinformation, inflammatory content).
- Text cleaning. HTML tags stripped, external hyperlinks removed, Unicode normalised, articles with < 10 chars excluded.
- User filtering. Users with fewer than 5 interactions removed. Suspected bots (> 500 interactions / hour) removed.
- Temporal processing. All timestamps unified to UTC at tick-level precision. Interactions preceding their article's publication time were dropped as anomalies.
- Splitting. Chronological 80 / 10 / 10 split — no future leakage.
Full details: §3.2 and Appendix D of the paper.
Ethical Considerations & Privacy
This dataset is released under strict ethical guidelines (paper Appendix G):
- ✅ Collected via formal collaboration with CoinMeta; users consented to data collection and analysis upon platform registration per the platform's ToS.
- ✅ All user IDs irreversibly anonymized with SHA-256 + salted keys. No PII (IP, device fingerprint, raw comment text with personal info) is included.
- ✅ Promotional spam, inflammatory content and verified misinformation removed.
- ⛔ Prohibited uses: market manipulation via synthetic sentiment, de-anonymization attacks, any commercial deployment, any use that violates local financial regulations.
If you discover residual sensitive content or an inadvertent privacy leak, please contact the maintainers immediately for retraction.
Limitations
- Only active users (≥5 interactions) are retained — cold-start distribution differs from production.
- The source platform skews toward Chinese-speaking Web3 users; non-CN market behaviors may be under-represented.
- Market data needed for the backtesting protocol (OHLC tick / OI / funding rate) is not redistributed here. The paper's backtest harness queries public exchange APIs (Binance, OKX) at runtime — see the GitHub release for the reference implementation.
License
Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
You are free to share and adapt the dataset for non-commercial research, provided you give appropriate credit. Commercial use is prohibited.
Maintainers
- CoinMeta — data provider
- Corresponding author: Jining Luan (UESTC / CoinMeta)
- For dataset-specific issues, open a discussion in the Community tab of this repository.
中文简介
W³ 是首个面向 Web3 新闻推荐 + 量化交易回测 的工业级基准数据集，由 CoinMeta 平台后端授权采集，覆盖 2022–2026 共 4 年，包含 73 万篇新闻、近万名匿名用户、5 类交互行为以及与 tick 级行情对齐的时间戳。本仓库仅限学术非商用研究，访问需同意条款。
- Downloads last month
- 11