Reinforced-Retrieval

enviroment prepare

install requirements

pip install -r requirements.txt

pre-download model weight

python save_pretrained.py

Before start, you need to manual download Natural Quastion dataset and put them into ./data

v1.0-simplified_simplified-nq-train.jsonl and v1.0-simplified_nq-dev-all.jsonl'

from https://ai.google.com/research/NaturalQuestions/download

config

All the config is in config.py

data preprocess

you should run "dataclean" first

python data_preprocess.py

train retriever

python train_ret_1.py

python train_ret_2.py

data preprocess

then run "process Q-A-Doc"

python data_preprocess.py

pre-train encoder

python train_enc.py

train agent

python train_RL_multi.py

experimental result

python Env_inference.py