This repository contains the code, data, and models for the paper "TIMO: Towards Better Temporal Reasoning for Language Models", accepted at COLM 2024.

• 📌 Introduction
• 🚀 Models
• 📊 Datasets
• 🌟 Highlights
• ⚙️ Installation
• 🛠️ Training and Inference
• 📜 License
• 🙏 Acknowledgements
• 📖 Citation

We introduce TIMO 🌱, a series of open-source large language models (LLMs) designed for temporal reasoning. TIMO models are trained on self-generated temporal preference pairs and optimized with a novel self-critic temporal optimization method, enabling the models to excel in both temporal reasoning and general tasks. TIMO is the new state-of-the-art for temporal reasoning across 19 tasks while maintaining robust general task performance.

Our models are available on Hugging Face:

• TIMO-7B
• TIMO-13B

We have uploaded all datasets used in various stages of training to Hugging Face. You can access them via the links below:

• Default pure-time datasets: TRAM-Temporal
• Temporal preference pairs generated by LLaMA2-7B: TRAM-Temporal-DPO-7B
• Temporal preference pairs generated by LLaMA2-13B: TRAM-Temporal-DPO-13B

TIMO achieves state-of-the-art results in temporal reasoning tasks. Here are the key results for 7B and 13B models:

Clone this repository and install the required dependencies:

git clone https://github.com/zhaochen0110/Timo.git
cd Timo
pip install -r requirements.txt

To quickly start using TIMO, run the following code:

from transformers import pipeline
pipeline = pipeline("text-generation", "Warrieryes/timo-7b-hf")

template = '''Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{query}\n\n### Response:'''

query = "What is 08:32 AM - 04:28?\n (A) 6:10 AM\n (B) 2:49 AM\n (C) 6:17 AM\n (D) 4:04 AM"

input = template.format(query=query)

output = pipeline(input)[0]['generated_text']

print(output)

To replicate the experimental results in our paper, run:

python inference.py \
    --model_path $model_path \
    --data_path $data_path \
    --excel_folder $excel_folder \
    --output_path $output_path 

We use the MAmmoTH project's code to train mathematical models. Then we use the following code to generate Temporal Preference pairs:

python generate.py \
    --model_path $model_path \
    --generate True \
    --train_data_path $train_data_path \
    --score True \
    --save_path $save_path

After generating preference pairs, we use Direct Preference Optimization (DPO) to train the model:

python tdpo.py \
    --model_name_or_path $model_name_or_path \
    --json_path $json_path \
    --output_dir $output_dir 

This project is licensed under the Apache 2.0 license - see the LICENSE file for details.

This project is partly based on the work done in MAmmoTH. Special thanks to their authors for valuable contributions.

Please cite our paper if you use our data, model or code. Please also kindly cite the original dataset papers.

@article{su2024timo,
  title={Timo: Towards Better Temporal Reasoning for Language Models},
  author={Su, Zhaochen and Zhang, Jun and Zhu, Tong and Qu, Xiaoye and Li, Juntao and Zhang, Min and Cheng, Yu},
  journal={arXiv preprint arXiv:2406.14192},
  year={2024}
}