Reindex-Then-Adapt

Implementation of WSDM'25 Reindex-Then-Adapt: Improving Large Language Models for Conversational Recommendation.

Dependencies

pip install -r requirements.txt

Data

1. Data from Orignal Datasets

These real-world datasets are used to (1) generate data for the Reindex step; (2) adjust model weights in the Adapt step.

Download the data and move it to data directory.

git clone https://huggingface.co/datasets/ZhankuiHe/Reindex-Then-Adapt-Real-World-Data
mv Reindex-Then-Adapt-Real-World-Data/* data
rm -r Reindex-Then-Adapt-Real-World-Data

Note: the RedditV1.5 data is also included in the link.

2. Data from LLMs Generation

These genereated datasets are used for the Reindex step.

Download the data and move it to reindex_step/data directory.

git clone https://huggingface.co/datasets/ZhankuiHe/Reindex-Then-Adapt-LLM-Generated-Data
mv Reindex-Then-Adapt-LLM-Generated-Data/* reindex_step/data
rm -r Reindex-Then-Adapt-LLM-Generated-Data

Note: since data for Reindex-Step are generated by Llama2, we also shared the scripts in tools for how we did data synthesis with the help of exllamav2.

Experiment Steps

Reindex Step

bash reindex_step.sh

Move the best checkpoint (i.e., the checkpoints folder in reindex_step/logs/*) to ckpts/best_aggregator for our Adapt step.

Adapt Step

for data in inspired redial redditv1.5; do
    bash adapt_step_for_${data}.sh
done

Note: we shared our best configs in scripts/best_*.yaml.

BibTeX

Please cite our paper if using this code, and feel free to contact [email protected] if any questions.

@inproceedings{he2025rta,
    author = {He, Zhankui and Xie, Zhouhang and Steck, Harald and Liang, Dawen and Jha, Rahul and Kallus, Nathan and McAuley, Julian},
    title = {Reindex-Then-Adapt: Improving Large Language Models for Conversational Recommendation},
    year = {2025},
    isbn = {9798400713293},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3701551.3703573},
    booktitle = {Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining},
    pages = {866–875}
}