# Dataset Catalog

所有 raw data 存於 `/home/sadoo/Projects/kvcache/research_data/raw/`。本檔記錄每個 dataset 的 **來源 / 下載方式 / 規格 / schema**，方便未來重下載 or 引用 paper。

更新日期：2026-04-19

---

## A. Paper-Reference Datasets (MaxEmbed ASPLOS'24 baseline)

### 1. Criteo (Display Advertising Challenge)
- **Path**: `raw/criteo_kaggle/`（17GB）
- **Source**:
  - Original: Criteo Labs 2014 Kaggle Display Advertising Challenge
  - Files: `train.txt` (11GB, 45.8M rows) + `test.txt` (1.4GB) + `dac.tar.gz` (4.3GB archive)
- **Download**: Kaggle API `kaggle competitions download -c criteo-display-ad-challenge`（現在需 Kaggle 授權；已備份）
- **Schema**：TSV (`\t` 分隔) — `label \t 13×int \t 26×hex_category`（40 columns）
- **Paper spec**: 35M items, 45.8M queries
- **Our stats**: 33.76M unique items, **45.84M queries** ✓

### 2. Criteo Terabyte (1TB Click Logs)
- **Path**: `raw/criteo_terabyte/`（39GB compressed gz，22/24 天）
- **Source**:
  - Original: `https://storage.googleapis.com/criteo-cail-datasets/day_{0..23}.gz`
  - Script: `dlrm/torchrec_dlrm/scripts/download_Criteo_1TB_Click_Logs_dataset.sh`
- **Status**: **缺 day_0 / day_1**（其餘 22 天齊全，足夠做 drift 實驗）
- **Schema**：同 Criteo（40 cols TSV）
- **Paper spec**: 882M items, 4.37B queries (full 24 days)
- **Supplementary subset**: `raw/criteo_terabyte/merci_day_0_subset/` — MERCI 切出的 7 slices (~230MB)
- **Note**: 若要 full paper scale 需要補 day_0/1（~90GB uncompressed）

### 3. Avazu
- **Path**: `raw/avazu/`（7.1GB）
- **Source**: Kaggle Avazu CTR Prediction Competition 2015
- **Download**: `kaggle competitions download -c avazu-ctr-prediction`
- **Files**: `train.csv` (5.9GB) + `avazu-ctr-train.zip` (1.3GB backup)
- **Schema**: CSV with header — `id, click, hour, C1, banner_pos, site_id, site_domain, site_category, app_id, ..., C21` (24 cols)
- **Paper spec**: 9.45M items, 40.4M queries
- **Note**: 需處理到 paper scale（現有 MERCI filtered 版只 1.07M items）

### 4. Amazon M2 (KDD Cup 2023)
- **Path**: `raw/amazon_m2/`（832MB）
- **Source**:
  - Official: https://www.aicrowd.com/challenges/amazon-kdd-cup-23-multilingual-recommendation-challenge（需 login）
  - **Used mirror**: Kaggle `riseserise/kdd-cup-2023`
  - Alternative code repo: `github.com/HaitaoMao/Amazon-M2-Session-Recommendation`（code only, no data）
- **Download**: `kaggle datasets download -d riseserise/kdd-cup-2023 --unzip`
- **Files**:
  - `sessions_train.csv` (248MB, 3.6M sessions, 6 locales: UK/DE/JP/IT/FR/ES)
  - `sessions_test_task1/2/3.csv`
  - `products_train.csv` (562MB, 1.55M product metadata rows)
- **Schema (sessions_train)**: CSV with multi-line quoted fields — `prev_items (numpy repr ['A' 'B'] format), next_item, locale`
- **Paper spec**: 1.39M items, 3.6M queries
- **Our stats**: **1,393,956 items, 3,606,249 sessions** ✓ MATCH
- **Processed**: `processed/amazon_m2/amazon_m2_session_remapped.txt`（已跑 `convert_amazon_m2.py` + MaxEmbed `process.py`）
- **Paper arxiv**: https://arxiv.org/abs/2307.09688

### 5. Alibaba-iFashion (POG KDD'19)
- **Path**: `raw/alibaba_ifashion/`（23GB raw / 9.2GB archived）
- **Source**:
  - Official: `github.com/wenyuer/POG`（Personalized Outfit Generation, Alibaba iFashion）
  - **Download**: Google Drive folder `1xFdx5xuNXHGsUVG2VIohFTXf9S7G5veq` via `gdown`
  - Papers with Code: https://paperswithcode.com/dataset/ifashion-alibaba-pog
- **Files**:
  - `item_data.txt` (1.17GB) — 5.12M rows, 4.75M unique items, 75 categories
  - `outfit_data.txt` (155MB) — 1.01M outfits (each = session-like query)
  - `user_data.txt` (21.75GB) — 19.19M user-click sessions
- **Schema**:
  - `item_data.txt`: CSV — `item_id, cate_id, pic_url, title`
  - `outfit_data.txt`: CSV — `outfit_id, item_id;item_id;...` (semicolon list)
  - `user_data.txt`: CSV — `user_id, item_id;item_id;..., outfit_id`
- **Paper spec**: 4.46M items, 999K queries
- **Our stats**: 4.75M items raw → **4.46M after freq filter**, 1.01M outfits ≈ **999K** ✓
- **Note**: Paper 53.6GB 差額是 POG visual embeddings / images，不在此 text-only 版
- **Paper arxiv**: https://arxiv.org/abs/1905.01866

---

## B. Classic Recommendation Datasets

### 6. Taobao User Behavior
- **Path**: `raw/taobao/`（4.4GB）
- **Source**: Alibaba Tianchi — https://tianchi.aliyun.com/dataset/649
- **Files**: `UserBehavior.csv` (3.5GB) + `userbehavior.zip` (906MB backup)
- **Schema**: CSV — `user_id, item_id, category_id, behavior_type, timestamp`（behavior: pv/buy/cart/fav）
- **Stats**: 100M events, ~1M users, ~4.1M items

### 7. Amazon Reviews (McAuley UCSD)
- **Path**: `raw/amazon/`（46GB）
- **Source**: https://cseweb.ucsd.edu/~jmcauley/datasets.html
- **Files**:
  - `Electronics.json.gz` (473MB) + `meta_Electronics.json.gz` (178MB)
  - `Home_and_Kitchen.json.gz` (132MB) + meta (146MB)
  - `Office_Products.json.gz` (18MB) + meta (46MB)
  - **`All_Amazon_Meta.json`** (12GB uncompressed) — complete Amazon product metadata
  - **`All_Amazon_Review.json.gz`** (34GB) — all review JSON

### 8. MovieLens (GroupLens)
- **Source**: https://files.grouplens.org/datasets/movielens/
- **Paths / Variants**:
  - `raw/movielens_1m/`: 24MB — 1M ratings, 6,040 users, 3,883 movies — `ratings.dat` `::` delimited
  - `raw/movielens_10m/`: 257MB — 10M ratings, ~72K users, 10,681 movies
  - `raw/movielens_25m/`: 1.1GB — **25M ratings, 162,541 users, 62,423 movies**（主要 variant）
  - `raw/movielens_32m/`: 912MB — 32M ratings, ~200K users, 87,585 movies（最新）
- **Schema (25M)**: CSV — `userId, movieId, rating, timestamp`（ratings 0.5-5.0，0.5 step，Unix ts）
- **Extras (25M)**: `genome-scores.csv`, `genome-tags.csv`, `links.csv`, `tags.csv`

### 9. Yelp Open Dataset
- **Path**: `raw/yelp/`（8.7GB）
- **Source**: https://www.yelp.com/dataset（需 signup）；**used mirror**: Kaggle `yelp-dataset/yelp-dataset` (v2022)
- **Download**: `kaggle datasets download -d yelp-dataset/yelp-dataset --unzip`
- **Files** (newline-delimited JSON):
  - `review.json` (5.3GB, **6,990,280 reviews**)
  - `user.json` (3.4GB, 1.99M users)
  - `business.json` (119MB, 150,346 businesses)
  - `checkin.json` (287MB)
  - `tip.json` (181MB)
- **Schema (review)**: `{review_id, user_id, business_id, stars, useful, funny, cool, text, date}`

### 10. LastFM
- **LastFM-1K**: `raw/lastfm_1k/` 2.4GB
  - **Source**: Ocelma Music Recommendation (原站 403)；**used mirror**: Kaggle `japarra27/lastfm-dataset`
  - 19.15M listening records, 992 users, 177K artists
  - Schema: TSV — `user_id, timestamp, artist_id, artist_name, track_id, track_name`
- **LastFM-360K**: `raw/lastfm_360k/` 1.6GB
  - Mirror: Kaggle `dhaatrisanisetty/last-fm`
  - 17.56M records, 359,347 users, ~295K artists
  - Schema: TSV — `user_sha1, artist_mbid, artist_name, plays`

### 11. Gowalla (Location Check-ins)
- **Path**: `raw/gowalla/`（398MB）
- **Source**: Stanford SNAP — https://snap.stanford.edu/data/loc-Gowalla.html
- **Files**:
  - `loc-gowalla_totalCheckins.txt` — 6.44M check-ins
  - `loc-gowalla_edges.txt` — 1.9M social edges
- **Schema (check-ins)**: TSV — `user_id, timestamp, lat, lng, location_id`

### 12. BookCrossing
- **Path**: `raw/bookcrossing/`（103MB）
- **Source**: http://www2.informatik.uni-freiburg.de/~cziegler/BX/（原站）；Kaggle mirror `arashnic/book-recommendation-dataset`
- **Files**: `Ratings.csv` (1.15M), `Users.csv` (278,859), `Books.csv` (271,360)
- **Schema (ratings)**: CSV — `User-ID, ISBN, Book-Rating`(0-10)

### 13. Foursquare (NYC + Tokyo)
- **Path**: `raw/foursquare/`（98MB）
- **Source**: https://sites.google.com/site/yangdingqi/home/foursquare-dataset（TSMC2014）；Kaggle mirror `chetanism/foursquare-nyc-and-tokyo-checkin-dataset`
- **Files**:
  - `dataset_TSMC2014_NYC.csv` (227,428 check-ins)
  - `dataset_TSMC2014_TKY.csv` (573,703 check-ins)
- **Schema**: CSV — `userId, venueId, venueCategoryId, venueCategory, latitude, longitude, timezoneOffset, utcTimestamp`

### 14. Netflix Prize
- **Path**: `raw/netflix/`（2GB）
- **Source**: Netflix Prize dataset (2006)；Kaggle `netflix-inc/netflix-prize-data`
- **Files**: `combined_data_1.txt` + `_2` + `_3` + `_4` (100M ratings total) + `probe.txt` + `qualifying.txt` + `movie_titles.csv`
- **Schema**: 每個 movie 以 `<movie_id>:` header 開始，之後 `user_id, rating (1-5), date` 每行

### 15. Steam Reviews
- **Path**: `raw/steam/`（2.1GB）
- **Source**: McAuley UCSD (404 for both URLs)；**used mirrors**: Kaggle `andrewmvd/steam-reviews` + `tamber/steam-video-games`
- **Files**:
  - `dataset.csv` — 6.4M reviews (columns: `app_id, app_name, review_text, review_score, review_votes`)
  - `steam-200k.csv` — 200K user-game events (`user_id, game, purchase/play, hours, 0`)

### 16. JD (JData 2016)
- **Path**: `raw/jd/`（2.2GB）
- **Source**: JD.com JData 2016 Competition（原站需中國手機驗證）；**used mirror**: Kaggle `owincontext/jdata2016`
- **Files**: 3 個 action files (Feb/Mar/Apr = 11.5+25.9+13.2M rows) + User (105K) + Product (24K) + Comment (558K)
- **Schema (action)**: CSV — `user_id, sku_id, time, model_id, type, cate, brand`（type: 1=browse, 2=add-cart, 3=delete, 4=purchase, 5=favorite, 6=click）

### 17. Tmall (IJCAI16)
- **Path**: `raw/tmall/`（1.7GB）
- **Source**: Alibaba Tianchi（需 Aliyun login）；**used mirror**: Kaggle `galuhramaditya/tmall-ijcai16`
- **Files**: 44.5M interactions
- **Schema**: CSV — `user_id, seller_id, item_id, category_id, timestamp, interaction`（click/purchase）

### 18. Anime Recommendations
- **Path**: `raw/anime/`（108MB）
- **Source**: Kaggle `CooperUnion/anime-recommendations-database`
- **Files**: `anime.csv` (12.3K anime) + `rating.csv` (7.8M ratings)
- **Schema**:
  - anime: `anime_id, name, genre, type, episodes, rating, members`
  - rating: `user_id, anime_id, rating`（-1 = unrated）

### 19. MIND (Microsoft News Recommendation)
- **Path**: `raw/mind/`（153MB）
- **Source**: https://msnews.github.io/（原站 Azure storage 已 disabled）；**used mirror**: Kaggle `arashnic/mind-news-dataset`
- **Files** (MINDsmall_train only, no dev/test):
  - `behaviors.tsv` — 157K impressions — `impr_id, user, time, history, impressions (Nxxx-1/0)`
  - `news.tsv` — 51K news — `news_id, cat, subcat, title, abstract, url, entities, abstract_entities`
  - `entity_embedding.vec`, `relation_embedding.vec`

---

## C. Supplementary / Intermediate

### 20. MERCI day_0_subset
- **Path**: `raw/criteo_terabyte/merci_day_0_subset/`
- **Purpose**: 由 MERCI 從 Criteo Terabyte day_0 切出的 7 slices（142858 × 7 samples），用於 MERCI pipeline benchmark
- **Note**: 雖放在 criteo_terabyte/ 下，實際只來自 day_0，約 day_0 的 1M-subset

### 21. MERCI _1_raw / _downloads (archived)
- **Path**: `raw/_merci_1_raw/`, `raw/_merci_downloads/`
- **Purpose**: MERCI 原 pipeline 前期下載的 raw stage（現已遷移）

---

## D. Download Method 統整

| Method | Used for |
|--------|----------|
| `kaggle datasets download -d <ref> --unzip` | Amazon M2, Yelp, Netflix, Steam, JD, Tmall, Anime, MIND, BookCrossing, Foursquare, LastFM |
| `wget`（直接 HTTP） | Criteo Terabyte, MovieLens（GroupLens 主站仍穩定）, Gowalla（SNAP） |
| `gdown`（Google Drive） | Alibaba-iFashion POG |
| Kaggle Competitions (`kaggle competitions download -c`) | Criteo, Avazu（現需授權） |

**斷鏈提醒**：
- Kaggle API 需要 `~/.kaggle/kaggle.json` token（現狀可能 401）
- Criteo Terabyte Google Cloud Storage URL（公開，穩定）
- McAuley UCSD（Steam / Amazon reviews）近期連線不穩 → 已改用 Kaggle mirror
- Microsoft MIND Azure storage 停止公開 → Kaggle mirror

---

## E. 使用規範

1. **Read-only**：`raw/` 下載後建議 `chmod -R 444` 防誤改
2. **單一副本**：不複製 raw 到 processed/，統一由 `processed/{dataset}/` 引用
3. **引用**：paper 寫作時請標註 paper 來源與 download mirror（reviewers 會檢查 reproducibility）

## F. 空間統計（2026-04-19）

```
raw/ 總計 150GB
├── criteo_terabyte/    39GB
├── criteo_kaggle/      17GB
├── alibaba_ifashion/   23GB
├── avazu/              7.1GB
├── taobao/             4.4GB
├── yelp/               8.7GB
├── lastfm_1k/          2.4GB  ├── lastfm_360k/         1.6GB
├── movielens_25m/      1.1GB  ├── movielens_32m/       912MB
├── movielens_10m/      257MB  ├── movielens_1m/        24MB
├── amazon/             991MB  (+ All_Amazon_{Meta,Review} 46GB)
├── amazon_m2/          832MB
├── netflix/            2.0GB
├── steam/              2.1GB  ├── jd/                  2.2GB
├── tmall/              1.7GB  ├── anime/               108MB
├── mind/               153MB  ├── gowalla/             398MB
├── bookcrossing/       103MB  ├── foursquare/          98MB
└── _merci_{1_raw,downloads}/  (archive)
```