Multi-hop Reasoning

The official implementation code of the paper "Improving Multi-hop Logical Reasoning in Small LMs with LoRA Training" (Link).

Methodology

Usage

• Run navset_lora_train.py to train the models.

python navset_lora_train.py --model_name meta-llama/Llama-3.1-8B-Instruct

• Run navset_lora_test.py and navset_quan_test.py for testing. The LoRA models should be in the same folder.

python navset_lora_test.py --model_name meta-llama/Llama-3.1-8B-Instruct --train_dataset navset

python navset_quan_test.py --model_name meta-llama/Llama-3.1-8B-Instruct

• Run navset_facts.py for evaluation:

python navset_facts.py --model_name meta-llama_Llama-3.1-8B-Instruct_navset_lora