Assessing consistent logical reasoning in Large Language Models

Zoe Tzifa Kratira, Daniël van Dijk, Oliver Neut, Theofanis Aslanidis and Vasilis Karlis

Supervised by Alina Leidinger

Set up

The provided pipeline requires a gpu environment to run

Dependencies for running the code

• The required dependencies can be found in env_global.yml.
• To create the required conda environment run: conda env create -f env_global.yml
• The name of the created env will be "llm_venv_final"

Provided data background and usage

We include in our repo two folders containing annotated_data and modeling_data.

• annotated_data includes model outputs manually annotated by human annotators
• modeling_data includes the required transofrmed training and evaluation sets so that we can finetune and evaluate the respective models for zero-shot and zero-shot chain-of-thought

Download Input Data

The data that should be downloaded for the inference tasks can be found in https://github.com/microsoft/LoNLI/tree/main/data named "data_v2.zip" should be unzipped and placed in the "data" folder outside src

Download finetuned models to use as LLM evaluators

In order to evaluate the output of the models we have finetuned and provide 2 mistral7B models so that we can perform evaluation of zero-shot tasks and zero-shot chain-of-thought tasks respectively.

• The finetuned model for zero-shot evaluation can be downloaded from https://amsuni-my.sharepoint.com/:u:/g/personal/vasilis_karlis_student_uva_nl/EV8kSS9oc8BEmh4bfDQlQM4BKhasSwH_PDxepjjvfjU_zg?e=M4DhBg
• The finetuned model for zero-shot chain-of-thought evaluation ca by downloaded from https://amsuni-my.sharepoint.com/:u:/g/personal/vasilis_karlis_student_uva_nl/EXaJXkjGi19GkDfVlXxWwMgB4VimVoLQag4FbHV85YJLfg?e=9ArpHI

The above folders should be unzipped and placed in src/finetuning folder

Main modules of our work

Run inference for the provided models

Main script for inference is located in 'src/inference.py'.

Example usage to run inference for Starling7B on the spatial subset with zero_shot and CoT reasoning:

srun python -u src/inference.py --model starling7B --run_tasks spatial --prompt-type zero_shot_cot

--model: starling7B, llama3_8B and mistral7B
--run_tasks: spatial, numerical, quantifier, comparative and temporal
--prompt-type: zero_shot and zero_shot_cot

Perform evaluation (and reproducing results)

Main script for evaluation is located in "src/evaluate.py".

Reproducing results in paper

Perform the following steps to obtain the prediction files to reproduce the logical consistency performance results displayed in the paper

1. Download the folders zero_shot/ and zero_shot_cot/ from https://drive.google.com/drive/folders/1uHPl-d9m9D3hgymkw8Cm0XKaOkeCV9qT?usp=drive_link
2. Unzip zero_shot.zip and zero_shot_cot.zip in the predictions/ folder
3. Unzip all subfolders: CD into zero_shot/ folder and run from terminal: for file in *.zip; do unzip "$file" -d "${file%.*}"; done
4. Repeat step 3 for zero_shot_cot/

Example usage to reproduce results for Starling7B on the numerical predictions subset with zero_shot reasoning evaluated with LLM

python -u src/evaluate.py --model starling7B --task numerical --prompt_type zero_shot_cot --evaluation_type llm

--model: starling7B, llama3_8B and mistral7B
--task: spatial, numerical, quantifier, comparative and temporal
--prompt_type: zero_shot and zero_shot_cot
--evaluation_type: logprob and llm

Note: only use logprob evaluation for zero_shot

Prepare data for LLM evaluator finetuning

In order to perform finetuning, we manually annotated LLMs outputs and placed the outputs in "src/finetuning/annotated_data" folder. Following that, "src/finetuning/ft_data_processing.py" is the module responsible for performing the necessary data splits and structure so that the we can perform finetuning and evaluation. The outputs are stored in "src/finetuning/modeling_data"