Liluo22/mmrev-data

MMRev: Multimodal Personalized Review Generation

Data accompanying What Helps and What Hurts: A Systematic Study of Multimodal Personalized Review Generation (see the GitHub repository). Code: https://github.com/liluo-2/multimodal-review-generation · Model: Liluo22/mmrev-llama3.1-8b-grpo

Task. Given a user's purchase history (past items with the user's reviews and ratings) and a target item with its rating, generate the review this user would write.

Three input scenarios. Each sample belongs to exactly one scenario, which determines what the model may see about the target item:

Scenario	Target-item fields used
both	title, text description, image (as a caption)
text_only	title, text description
img_only	image (as a caption) only — the title is hidden

Scenario membership comes from a data-filtering experiment: a fine-tuned Qwen2.5-VL generated each review from text only, image only, or both, and each sample was placed in the scenario whose input configuration best explained its review. The three subsets are therefore disjoint and differ in content; compare models within a scenario, not across scenarios.

Subsets

Category	both (train/test)	text_only	img_only
Arts (Arts, Crafts & Sewing)	750 / 150	707 / 141	641 / 128
Sports (Sports & Outdoors)	898 / 180	891 / 178	840 / 168
Toys (Toys & Games)	1,216 / 243	1,190 / 238	1,230 / 246

from datasets import load_dataset
ds = load_dataset("Liluo22/mmrev-data", "arts_both")      # configs: {arts,sports,toys}_{both,text_only,img_only}
print(ds["train"][0]["reference_review"])

Fields (data/<Category>/<scenario>/<split>.jsonl)

Field	Description
id	Row index into raw/<Category>/final.jsonl (also the key used by splits/*.npz and image_descriptions/)
dataset, scenario, split	Category, input scenario, train/test
user_id	Anonymized user ID from Amazon Reviews 2023
history	List of past purchases, oldest → newest: product_title, parent_asin, rating, timestamp, review. The paper uses the 5 most recent.
target_product_title, target_product_desc, target_parent_asin, target_image_url, target_rating, target_timestamp	Target item. Fields not visible in the sample's scenario must be dropped when building the prompt.
reference_review	Ground-truth review

Product descriptions are provided for target items only; history items carry titles, reviews, and ratings.

Other files

Path	Content
raw/<Category>/final.jsonl	Full preprocessed pool per category (all users after filtering), in the original nested format used by the training code
splits/<Category>_<scenario>_split.npz	Arrays train_random / test_random of row indices into final.jsonl. <Category>_mix_split.npz is the union of the three scenarios, used for mix-SFT and GRPO training
image_descriptions/<Category>/*.jsonl	Qwen2.5-VL captions of product images: {"id", "image_descriptions": [{"image", "description"}]}, keyed by id
teacher_scores/per_sample_detail_<Category>_train.csv	Per-sample metrics (BLEU, ROUGE-1/2/L, BERTScore, CIDEr) of the single-scenario and mix-scenario SFT models on the training set; used for reward normalization statistics and teacher look-up in Stage II

Image captions were generated with this prompt:

Look at this product image and describe it in 2-4 natural sentences. Include:
- What the product is (category, brand/name if visible)
- Appearance (color, material, shape/design)
- Any visible text or labels
- Notable features or details
Requirements: only describe what is actually visible; be concise; do not guess.

Construction

Source: Amazon Reviews 2023 (McAuley Lab). Users with 10–30 reviews were kept; duplicate, very short, and templated reviews were removed, as were items lacking the fields their scenario requires and samples exceeding the LLaMA-3.1 context window. Scenario membership was assigned with the data-filtering experiment described above; the three scenario subsets do not overlap.

License and intended use

Derived from Amazon Reviews 2023; use is subject to the terms of the original dataset. Intended for research on personalized and multimodal text generation. Do not use to produce or post synthetic reviews on real platforms.

Citation

@techreport{chen2026mmrev,
  title  = {What Helps and What Hurts: A Systematic Study of Multimodal Personalized Review Generation},
  author = {Chen, Chuqiao and Sheng, Xingyue},
  year   = {2026},
  institution = {New York University Shanghai}
}
@article{hou2024bridging,
  title   = {Bridging Language and Items for Retrieval and Recommendation},
  author  = {Hou, Yupeng and Li, Jiacheng and He, Zhankui and Yan, An and Chen, Xiusi and McAuley, Julian},
  journal = {arXiv preprint arXiv:2403.03952},
  year    = {2024}
}