user_gemma_3_27b_it

user_gemma_3_27b_it

A user simulator for evaluating agentic conversational recommender systems. SFT fine-tune of google/gemma-3-27b-it on the teacher–student corpus described in Simulating Diverse User Behavioral Stereotypes for Evaluating Agentic Conversational Recommenders (RecSys '26).

Given a behavioral stereotype, a user's interaction history, and a target item, the model generates the user turns of a dialogue with a recommender system. It exists to evaluate recommenders against simulated users who express vague or evolving preferences, rather than perfectly specified queries.

This repository also ships dataset/training_dataset.jsonl, the SFT corpus in JSON-lines form. The same corpus is now released, in Parquet and with a dataset card, as the corpus_sft_simulator config of swap-uniba/ecommerce-user-sim — prefer that one.

Behavioral stereotypes

Prompt format

The model was trained on a single format and needs it to behave as intended:

<bos><start_of_turn>user
You are an advanced user simulator. Your objective is to generate realistic and coherent user responses in a dialogue with a recommendation system based on the user persona, interaction history and target item provided.
[...]
# Category
Electronics

# User Persona
VAGUE_TO_SPECIFIC

# Interaction history
## Item Name: ...
## Description: ...
## Visual Description: ...
## Review: ...

# 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

Product images enter only as the textual ## Visual Description: the simulator does not consume pixels.

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("swap-uniba/user_gemma_3_27b_it")
model = AutoModelForCausalLM.from_pretrained("swap-uniba/user_gemma_3_27b_it", device_map="auto")

prompt = build_prompt(category, persona, history, target)   # in the format above
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device),
                     max_new_tokens=128)
print(tok.decode(out[0][len(tok(prompt).input_ids):], skip_special_tokens=True))

Fully materialized prompts, identical to the training ones, are in the corpus_sft_simulator config of swap-uniba/ecommerce-user-sim.

Training data

A teacher–student corpus of 10,644 dialogues across 10 Amazon domains, yielding 50,495 SFT examples. The released checkpoints were trained on exactly this corpus, published as the corpus_sft_simulator config. Data, evaluation protocol, and pipeline: swap-uniba/ecommerce-user-sim. Code: https://github.com/swapUniba/Simulating-Diverse-User-Behavioral-Stereotypes

• LoRA: rank 64, alpha 32, dropout 0.1, 4 target modules

Framework versions

TRL 0.24.0 · Transformers 4.57.1 · PyTorch 2.9.0 · Datasets 4.2.0 · Tokenizers 0.22.1

Other checkpoints

License

Weights derive from Gemma-3 and are subject to the Gemma license terms. The training data is released separately under CC BY-NC 4.0.

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}
}

The simulator is also used in:

@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}
}