mCoT: Multilingual Instruction Tuning for Reasoning Consistency in Language Models

Paper: https://arxiv.org/abs/2406.02301

Code: https://github.com/laihuiyuan/mCoT

Dataset: https://huggingface.co/datasets/laihuiyuan/mCoT-MATH

Introduction

We introduce mCoT, a 7B parameter model for multilingual math reasoning that achieves impressive multilingual reasoning consistency across multiple languages. Based on Mistral-7B-v0.1, mCoT is trained on mCoT-MATH, the first large-scale multilingual math CoT reasoning dataset containing around 6.3 million samples for 11 diverse languages.

Results on MGSM

Results on MSVAMP

Prompt Template

# Template
template = "Question: \n{question} \nAnswer: \n{language}\n"

# Language prompt
bn = "আসুন ধাপে ধাপে চিন্তা করি।"
de = "Denken wir Schritt für Schritt."
en = "Let's think step by step."
es = "Pensemos paso a paso."
fr = "Réfléchissons étape par étape."
ja = "段階的に考えてみましょう。"
ru = "Давайте думать поэтапно."
sw = "Hebu fikiria hatua kwa hatua."
te = "అంచెలంచెలుగా ఆలోచిద్దాం."
th = "ลองคิดทีละขั้นตอน"
zh = "让我们一步步思考。"

# Math question
math_en = "A robe takes 2 bolts of blue fiber and half that much white fiber.  How many bolts in total does it take?"

# An example for the English question
prompt = template.format(question=math_en, language=en)

Citation

If you use any content from this repository, please cite our paper:

@inproceedings{lai-etal-2024-mcot,
    title = "mCoT: Multilingual Instruction Tuning for Reasoning Consistency
    in Language Models",
    author = "Lai, Huiyuan and Nissim, Malvina",
    booktitle = "Proceedings of the 62nd Annual Meeting of the Association
    for Computational Linguistics,
    month = aug,
    address = "Bangkok, Thailand",
    year = "2024",
    publisher = "Association for Computational Linguistics"
}