Pre-Calc – Learning to Use the Calculator Improves Numeracy in Language Models

Paper: https://arxiv.org/abs/2404.14355
Published/Presented at SemEval @ NAACL 2024 and AIForMath workshop @ ICML 2024

Source code for the project titled as above

You can find our datasets and models at https://huggingface.co/collections/Calc-CMU/pre-calc-657a5ad5f1ae42fb12364563

Encoder-Only Approach

cd Encoder-Only

data_preprocessing.py preprocesses the Calc-MAWPS data https://huggingface.co/datasets/MU-NLPC/Calc-mawps with annotations required for Calc-BERT continued pretraining. Preprocessed data pushed to https://huggingface.co/vishruthnath/Calc_BERT_20.

train.py contains the script for continued pretraining with the dual objective. The data used is Calc_BERT_20 linked above from Huggingface.

qnli_finetuine_inference.py is the script for finetuning Calc-BERT on downstream QNLI tasks.

inference_awpnli.py for inference of Pre-Calc model on AWPNLI Tasks

Encoder- Decoder Based approach

cd Encoder-Decoder

For continued pretraining on reframed MAWPS and Multi-NLI data, download the data from https://huggingface.co/datasets/vishwa27/MAWPS-MNLI-CalX-NumEval and place them under the data folder as

data/train_preft_flant5.csv data/test_preft_flant5.csv

Change the appropraite repo_name at line 165 and run

python train_flanT5.py

To evaluate the trained models on specific tasks, firstly download the datasets from AWPNLI, NewsNLI and RTE-Quant data from https://huggingface.co/datasets/NLPFin/Quantitative101

Then run

python eval_ft_flanT5.py to test FlanT5 (TB-PT ours)

and

python eval_fs_flanT5.py to test FlanT5 with few-shot prompting