planeB/ecommerce-user-sim

E-commerce User Simulation — corpus and evaluation data

Supporting data for two papers on user simulators for conversational recommender systems:

• Simulating Diverse User Behavioral Stereotypes for Evaluating Agentic Conversational Recommenders (RecSys '26) — teacher–student corpus and evaluation data
• Beyond the Lookup: Simulating Realistic User Uncertainty for the Evaluation of Conversational Agentic Recommenders (Information Systems Frontiers, 2026)

Code: https://github.com/swapUniba/Simulating-Diverse-User-Behavioral-Stereotypes

The repository is split into two families, mirroring the two folders of the code repository.

Family	Mirrors	Contents
corpus/	SyntenticData/	teacher–student training corpus, 10 domains
eval/	PreferenceCoverage/	evaluation protocol data, 4 domains

corpus/ — teacher–student corpus (10 domains)

Domain	items	users	SFT examples	dialogues
All_Beauty	4,731	913	8,708	1,826
Baby_Products	3,247	387	3,662	774
Beauty_and_Personal_Care	4,052	396	3,766	792
CDs_and_Vinyl	10,259	999	9,144	1,998
Cell_Phones_and_Accessories	4,435	440	4,320	880
Electronics	4,781	456	4,469	912
Movies_and_TV	4,686	378	3,546	756
Office_Products	3,663	438	4,166	876
Sports_and_Outdoors	4,320	413	3,945	826
Video_Games	3,497	502	4,769	1,004
Total	47,671	5,322	50,495	10,644

• corpus_sft_simulator — the 50,495 examples the student simulator was distilled on. prompt is the complete Gemma-formatted string, including the model turn: it is exactly the input the released checkpoints saw during training.
• corpus_dialogues — the 10,644 teacher-generated dialogues, with ending_condition (accept / reject), recommend_target_in_list, unnatural_assistant, and the inlined target_item.
• corpus_items — 47,671 items with visual_description (a textual rendering of the product image) and generated_description.
• corpus_users — the 5,322 users the corpus draws on, with their target and interaction_history.

eval/ — evaluation protocol (4 domains)

100 users x 4 domains x 3 stereotypes = 1,200 evaluation episodes.

Domain	items	users	prompt components	catalog
All_Beauty	335	100	300	356
Clothing_Shoes_and_Jewelry	260	100	300	132,043
Electronics	492	100	300	4,117
Movies_and_TV	1,457	100	300	4,283
Total	2,544	400	1,200	140,799

• eval_users — the 100 users per domain: inferred profile (profile_*), opening instruction, target item, and history.
• eval_items — items with visual_description + generated_description, filtered to the ASINs the episodes reference (targets and history). This is the artifact required by PreferenceCoverage/phase1_extract_aspects.py.
• eval_catalog — the full all_products.json: the retrieval pool the evaluated recommenders search over. Plain Amazon metadata, without the generated descriptions.
• eval_prompt_components — prompt components per stereotype, 100 per cell; user_persona distinguishes the stereotype within each split.
• eval/personalwab/<Domain>/ — the four original JSON files (users_ids, user_profiles, user_history, user_instructions) verbatim, for PersonalWAB-style loaders. user_instructions has test (100) and train (0).

Quickstart

from datasets import load_dataset

# the 100 evaluation users of one domain
users = load_dataset("swap-uniba/ecommerce-user-sim", "eval_users", split="Electronics")

# items with the descriptions Preference Coverage needs
items = load_dataset("swap-uniba/ecommerce-user-sim", "eval_items", split="Electronics")

# the exact string the simulator was trained on
sft = load_dataset("swap-uniba/ecommerce-user-sim", "corpus_sft_simulator", split="Electronics")
print(sft[0]["prompt"])

The simulator prompt has the following structure, and it is the only format the released checkpoints expect:

<bos><start_of_turn>user
You are an advanced user simulator. [...]
# Category
<domain>
# User Persona
<SPECIFIC | VAGUE | VAGUE_TO_SPECIFIC>
# Interaction history
## Item Name: ...
## Description: ...
## Visual Description: ...
## Review: ...
  (repeated for every item in the history)
# Target item
## Item Name: ...
## Description: ...
## Visual Description: ...
## Review: ...

Generate the first user message based on the above information.<end_of_turn>
<start_of_turn>model

The simulator consumes text only: product images enter exclusively as visual_description.

Stereotypes: code labels vs. paper labels

In the code and in this data	In the paper
SPECIFIC	Direct
VAGUE_TO_SPECIFIC	Vague-Proactive
VAGUE	Vague-Reactive

Mapping to the names used by the scripts

The scripts in the code repository read directories named historically. This is the map:

Name in the scripts / on disk	Config here
items_used_with_descriptions_with_descriptions	eval_items
items_used_with_images_with_descriptions_with_descriptions	corpus_items
items_used_with_descriptions (in the PreferenceCoverage README)	eval_items
user_profiles_cleaned	eval_users
user_profiles_filtered_cleaned	corpus_users
processed_prompts_new.jsonl	corpus_sft_simulator
generated_responses_new.jsonl	corpus_dialogues
all_products.json	eval_catalog

Product images

Product images are not redistributed. Their textual rendering (visual_description) is in the *_items configs and is all the pipeline ever consumes. The images column preserves the original Amazon URLs (hi_res, large, thumb, variant); to regenerate the JPEGs locally, use download_images.py in the code repository. The local_image_path column present in the sources has been dropped: it held absolute paths from the development machine.

Normalizations applied

The sources diverged across domains; here every config has an identical schema on all splits.

• details converted to a JSON string (in the sources it was a struct with hundreds of fields in 2 of 10 domains, and a Python repr in the remaining 8).
• description converted to list<string>; rating_number to int64 (lossless cast: zero values had a fractional part); categories and bought_together to list<string>.
• images reduced to the four canonical keys. The resolution variants (360w...1920w), present only in Movies_and_TV, were all null.

Known limitations

• corpus_users contains only the users the corpus actually draws on. The sources held 5,949, of which 627 are leftovers that no SFT example or dialogue references (621 of them in Movies_and_TV) and whose items were never enriched. Filtering them yields the 5,322 users reported in the paper.
• Every target resolves in *_items, in every domain. A few interaction_history ASINs, however, have no corresponding item row; the history_unresolved column counts them per user. Affected users: Movies_and_TV 134, CDs_and_Vinyl 11, Office_Products 6, Cell_Phones_and_Accessories 2, Sports_and_Outdoors 2.
• eval_items is filtered to the ASINs the episodes reference. The full pool can be regenerated with process_personalwab/main.py.
• eval_catalog has no visual_description / generated_description: it is the retrieval pool, not the item dataset the metric needs.
• Multi-turn simulation logs are not included; they are reproducible with the code.

Simulator checkpoints

Model	Hugging Face
Gemma-3 270M	https://huggingface.co/planeB/user_gemma_3_270m_it
Gemma-3 1B	https://huggingface.co/planeB/user_gemma_3_1b_it
Gemma-3 4B	https://huggingface.co/planeB/user_gemma_3_4b_it
Gemma-3 12B	https://huggingface.co/planeB/user_gemma_3_12b_it
Gemma-3 27B	https://huggingface.co/swap-uniba/user_gemma_3_27b_it

The weights are subject to the Gemma license terms, independent of the license of this data.

Provenance and license

Derived from Amazon Reviews 2023 (McAuley Lab, UCSD). Reviews and product metadata remain subject to the terms of the original source; the textual descriptions (generated_description, visual_description) and the dialogues are LLM-generated. Released under CC BY-NC 4.0 for research use.

Citation

@inproceedings{PetruzzelliMartinaSimulating,
  author    = {Petruzzelli, Alessandro and Martina, Alessandro Francesco Maria and Musto, Cataldo and de Gemmis, Marco and Lops, Pasquale and Semeraro, Giovanni},
  editor    = {Konstan, Joseph A. and Karypis, George and Adomavicius, Gediminas and Chen, Minmin and Goethals, Bart and Willemsen, Martijn C.},
  title     = {{Simulating Diverse User Behavioral Stereotypes for Evaluating Agentic Conversational Recommenders}},
  booktitle = {Proceedings of the 20th {ACM} Conference on Recommender Systems, RecSys 2026, Minneapolis, Minnesota, USA, September 28-October 2, 2026},
  publisher = {{ACM}},
  year      = {2026},
  doi       = {10.1145/3773078.3831840},
  url       = {https://doi.org/10.1145/3773078.3831840}
}

@article{BeyondLookup,
  author     = {Petruzzelli, Alessandro and Martina, Alessandro Francesco Maria and Musto, Cataldo and de Gemmis, Marco and Lops, Pasquale and Semeraro, Giovanni},
  journal    = {Information Systems Frontiers},
  title      = {Beyond the Lookup: Simulating Realistic User Uncertainty for the Evaluation of Conversational Agentic Recommenders},
  year       = {2026},
  url        = {https://doi.org/10.1007/s10796-026-10787-3}
}

Mirror: planeB/ecommerce-user-sim