semeval2024-task7-subtask1

Final project for CS6120 NLP, by Junxin Zheng, Divyank Singh, Aakarsh Arora

Task Description

https://sites.google.com/view/numeval/tasks?authuser=0#h.uin1dj2oxo8t

To download the dataset for this subtask (Quantitative-101)

https://drive.google.com/drive/folders/10uQI2BZrtzaUejtdqNU9Sp1h0H9zhLUE?usp=sharing

Original paper

[1] Chen, Chung-Chi, et al. "Improving Numeracy by Input Reframing and Quantitative Pre-Finetuning Task." Findings of the Association for Computational Linguistics: EACL 2023. 2023.

Reproduce the experiments with Flan-T5 from:

[2] Chen, Kaiyuan, Jin Wang, and Xuejie Zhang. "YNU-HPCC at SemEval-2024 Task 7: Instruction Fine-tuning Models for Numerical Understanding and Generation." Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024). June 20-21, 2024.

To Reproduce our experiments for BERT, T5, and Llama models:

BERT:

Navigate to the ./BERT, and execute notebooks (BERT_QNLI.ipynb or Bert_QP.ipynb).

T5:

1. Navigate to the ./T5.
2. Install necessary dependencies as specified in colab_requirements.txt.
3. Run training and evaluation by executing the Jupyter notebooks or scripts ({task}_train_reproduction.ipynb, {task}_test_reproduction.ipynb/.py)in this directory.

Llama:

1. Navigate to the ./Llama.
2. Make sure to run it with Linux machines or WSL since Unsloth only supports Linux environments.
3. Follow the instructions in the llama_* notebooks (e.g., llama_qnli_train.ipynb, llama_qqa_train.ipynb).
4. Execute these notebooks or scripts to train and evaluate Llama models on the specified tasks.