VerityMath

Official repo for "VerityMath: Advancing Mathematical Reasoning by Self-Verification Through Unit Consistency"

Quick Start

git clone https://github.com/vernontoh/VerityMath
cd VerityMath
conda create -n verity python=3.10 -y
conda activate verity
pip install -r requirements.txt

NOTE: You will need OPENAI KEY if you want to obtain the annotations yourself.

Convert UCPs to HF Dataset

python convert_to_dataset.py \
--synthetic_data_path gsm8k_program \
--output_data_path gsm8k_ucp \

Finetune with UCPs

accelerate launch finetune.py \
--load_in_4bit \
--bnb_4bit_use_double_quant \
--model codellama/CodeLlama-7b-hf \
--dataset_name gsm8k_ucp \
--hf_auth_token <AUTH TOKEN> \

Evaluation

accelerate launch evaluate.py \
--hf_auth_token <AUTH TOKEN> \
--run saved/<Saved Model Directory> \

(Optional)

Generate UCPs Dataset

python generate_annotations.py \
--task program \
--dataset gsm8k \
--model gpt-4-1106-preview \
--openai_key <OPENAI KEY> \

Generate Classification

python generate_annotations.py \
--task classification \
--dataset gsm8k \
--model gpt-3.5-turbo \
--openai_key <OPENAI KEY> \

Acceleration Config

compute_environment: LOCAL_MACHINE
distributed_type: MULTI_GPU
downcast_bf16: 'no'
gpu_ids: 0,1,2,3
machine_rank: 0
main_training_function: main
mixed_precision: 'no'
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false