WZDavid/MemoryCD

MemoryVCD — URL edition

323 users · 12 domains · 535,698 interaction records · 276,418 history products

URL edition: product image URLs are provided as strings; no product image bytes are included. Cached images are available in the Full edition. External URLs may expire or fail.

Release: 2026-11-12. The URL edition and Full edition share identical histories, product tables, evaluation windows, and candidates. Full additionally contains the product_images subset. Use file hashes in manifest.json to verify alignment.

Browse the data

The default interactions subset shows one interaction per row: domain, user, product, rating, UTC date, review title/body, and image URL. It is the complete corpus, not a preview sample. Use the subset selector or the links below.

Subset	Rows	Purpose
interactions	535,698	Read, search and filter individual reviews
products	481,333	Product metadata for histories and saved evaluation candidates
users	323	User activity counts and original ordering
evaluation_windows	15,504	Time-axis and domain-axis test windows and memory boundaries
evaluation_candidates	19,751	Saved ranking/choice candidates, with source provenance

The test split label denotes an evaluation corpus, not that every interaction is a held-out test target. Memory and test targets are selected within each user's history using evaluation_windows. There is no random train/validation/test split.

Load

from datasets import load_dataset

repo = "WZDavid/MemoryCD"
reviews = load_dataset(repo, "interactions", split="test")
products = load_dataset(repo, "products", split="test")
windows = load_dataset(repo, "evaluation_windows", split="test")
print(reviews[0])

Original user histories

users/users_12domains.jsonl.gz preserves the original nested histories, one user per line (323 lines). The users preview subset is a compact index, not the nested histories themselves. The archive is provided separately to keep the browser readable.

import gzip, json
from huggingface_hub import hf_hub_download

path = hf_hub_download(repo, "users/users_12domains.jsonl.gz", repo_type="dataset")
with gzip.open(path, "rt") as f:
    users = [json.loads(line) for line in f if line.strip()]

tools/release_dataset_helpers.py provides load_users, split_window and restore_ranking_input. These reconstruct data inputs without invoking any models. This is a data release, not a bundle of model outputs, API credentials, or experiment caches.

Corpus distribution

The first four domains are evaluation targets; the remaining eight provide additional cross-domain memory. The primary four contain 314,116 interactions; all twelve contain 535,698. Users are shared across domains, so domain user counts must not be summed.

Domain	Users	Interactions	History products
Beauty_and_Personal_Care	323	76,094	33,475
Books	323	79,874	48,054
Electronics	323	44,268	25,441
Home_and_Kitchen	323	113,880	60,900
All_Beauty	293	2,438	1,336
Arts_Crafts_and_Sewing	319	18,588	10,508
Automotive	321	6,638	4,348
Grocery_and_Gourmet_Food	323	48,960	19,154
Health_and_Household	323	80,391	35,543
Movies_and_TV	278	9,395	8,433
Office_Products	323	28,325	14,032
Sports_and_Outdoors	323	26,847	15,194

Image and metadata coverage

All 481,333 required product/domain records have metadata. This includes the random ranking candidates and all saved time-cutoff hard candidates, beyond the 276,418 history products.

There are 456,761 product image URLs and 243,394 valid cached product images (50.57% of required product/domain rows). Among history products, 83,565/276,418 have a cached image (30.23%). URL availability is not a live HTTP success guarantee.

Domain	Required products	Image URLs	Cached images
Beauty_and_Personal_Care	82,891	82,890	49,225
Books	100,625	78,730	49,539
Electronics	76,516	76,502	47,696
Home_and_Kitchen	112,753	112,753	72,262
All_Beauty	1,336	1,336	781
Arts_Crafts_and_Sewing	10,508	10,499	2,229
Automotive	4,348	4,348	1,397
Grocery_and_Gourmet_Food	19,154	19,149	2,911
Health_and_Household	35,543	35,543	7,689
Movies_and_TV	8,433	5,840	244
Office_Products	14,032	13,977	3,936
Sports_and_Outdoors	15,194	15,194	5,485

products.cached_image_available describes availability in the companion Full release, including when read from the URL edition. image_shard is relative to the Full repository. Images are embedded as bytes in that shard; the image field's path is only a filename label, not an additional loose file to download. Join using (domain, parent_asin). Missing-cache entries remain in the product and interaction tables; only valid cached images enter product_images. image_sha256 checks image payload integrity.

Product images differ from user-uploaded review images. The primary-domain histories preserve their review-image fields; the eight added domains' extraction did not retain these fields. A missing field there must not be interpreted as proof that the original review had no image. review_images_field_present makes this distinction explicit. Review image bytes are not bundled.

Fields and ordering

• interaction_id: stable identifier formed from source user index, domain, and original position within that domain.
• interaction_index: zero-based index after a stable timestamp sort within one user/domain; used by evaluation windows.
• source_user_index and source_interaction_index: preserve source ordering and allow reconstruction.
• timestamp: original Unix milliseconds; date is its UTC ISO representation.
• review_title and review_text: complete review text. rating preserves the original numeric score.
• image_url: selected MAIN product image (large, then hi_res, then thumb; fallback to another variant).
• review_images_json: JSON representation of original review images; null if absent.
• Product metadata mirrors the evaluator's slim representation: title limited to 200 characters and description to 400. No price field is promised in the new product tables. Per-domain slim JSONL files are in metadata/.

Evaluation protocol and actual candidate coverage

All 323 users have windows for the four primary target domains. Time cutoffs are 0, .25, .5, .75 and 1. A three-interaction window moves along the timestamp-sorted target history; preceding interactions are the initial memory. Interior centers use Python round(c*n), with indices [center-1, center+2).

Domain windows hold the target's final three interactions fixed. D0 has no initial memory. D1–D3 add non-target primary domains in the order Personal Care, Books, Electronics, Home & Kitchen, skipping the target. D4 adds the target; D5 adds the eight additional domains; Dtx uses only the additional eight. Memory must precede the earliest test timestamp strictly, and source-domain segments remain in their specified order. For Image Prediction, earlier test interactions are revealed before later predictions; other tasks retain the same initial memory throughout the window.

Saved candidates have different coverage from windows. The data/v8e/... source files cover the original first 100 users at each cutoff. The base data/cross_domain.jsonl ranking candidates and hard_negatives_*_v3.jsonl choice candidates are separately identified by source_file. Missing saved candidates are not fabricated for the other users. Filter by task, source_file, and cutoff before evaluation; do not mix candidate versions.

The stored hard-negative records contain 3 negatives, although a paper description says 4. This release preserves the actual files rather than silently adding a fourth candidate. Ranking pools retain the saved negatives. evaluation/source_candidates/ also contains unchanged hard-negative JSONL files. The six tasks reuse these histories/windows; they are not six independent corpora.

Integrity and limitations

The original histories contain 10,404 excess records sharing a user/domain/product/timestamp key. They are retained to preserve the original evaluation timelines; this is not a claim that all 535,698 rows are unique events. Stable interaction_id values distinguish the retained rows. Users were selected for substantial overlap/activity across the four primary domains; they are not a representative random sample of Amazon users. Ratings are strongly skewed toward 4–5 stars.

See reports/ for corpus, metadata, image and release validation results. manifest.json lists SHA-256 hashes and sizes for release files. Image caches were checked for valid decoding before export. No external image downloads were performed during packaging.

Source and attribution

Derived from Amazon Reviews 2023. Original reviews and product images retain their applicable source terms and third-party rights; this packaging does not grant a new license to those materials. Refer to the upstream dataset for source attribution and usage terms.

Existing MemoryCD clients

The original four-domain configurations users_interactions, meta_personal_care, meta_books, meta_electronics, and meta_home remain available with the same records, converted to Parquet for compatibility with the new preview tables. Their original JSONL files remain unchanged in the repository. They are legacy subsets, not the new twelve-domain release. The default subset is now interactions; existing callers should specify their legacy configuration explicitly.

Legacy preview limitation: meta_home loads successfully with load_dataset, but its large nested metadata exceeds the Hub viewer's 200 KB first-rows response limit (TooBigContentError). Use the new products subset and filter domain = Home_and_Kitchen for a readable preview. The new five release subsets have passed online preview checks.