liujianyu0824 / primo

Progressive Induction for Multi-hop Open Rule Generation

Progressive Induction for Multi-hop Open Rule Generation

Code Structure

The source code is organized as follows:

data: contains multi-hop open rule dataset.

checkpoint: contains the model weight files.

data_process: contains the code for preprocessing dataset.

trl: contains the utility functions.

gen_openrule.py: contains the main executable code for genarating open rule.

inference_ppo_gpt.py: contains the code for testing the performance of gpt after ppo.

inference_reward.py: contains the code for testing the performance of reward model.

iTrainingLogger.py: contains the code for loading dataset.

ppo_openrule.py: contains the code for reinforcement learning of G_net and E_net.

train_gpt.py: contains the code for training gpt for Generation and Extraction.

train_reward_model.py: contains the code for training reward model.

Commands

requirements

pip install -r requirements.txt

Training gpt

python train_gpt.py

Training RM

python train_reward_model.py

Training gpt with ppo

python ppo_openrule.py

Generating multi-hop open rule

python gen_openrule.py