KAMR

Code for "KAMR: Grounding Generation via Knowledge-Aligned Multi-hop Retrieval".

Datasets

• PathQuestion: https://github.com/zmtkeke/IRN/tree/master/PathQuestion
• ComplexWebQuestions (CWQ): https://huggingface.co/datasets/rmanluo/RoG-cwq

Workflow

1) Build the pretraining dataset

python augment.py

2) Pretrain the retriever

python pretrain.py

3) Run retrieval inference (batch)

bash run_retr.sh

4) Run LLM generation (batch)

bash run_generation.sh

Notes

Additional documentation and usage details will be released when the repository is made public.