Update README.md
Browse files
README.md
CHANGED
@@ -17,23 +17,16 @@ tags:
|
|
17 |
---
|
18 |
# mCoT: Multilingual Instruction Tuning for Reasoning Consistency in Language Models
|
19 |
|
20 |
-
Paper:
|
21 |
|
22 |
-
Code:
|
23 |
|
24 |
-
Dataset:
|
25 |
|
26 |
### Introduction
|
27 |
-
We introduce mCoT
|
|
|
28 |
|
29 |
-
### 🤗 Dataset: [mCoT-MATH](https://huggingface.co/datasets/laihuiyuan/mCoT-MATH)
|
30 |
-
|
31 |
-
Based on [MetaMathQA](https://github.com/meta-math/MetaMath) and [MathInstruct](https://github.com/TIGER-AI-Lab/MAmmoTH)
|
32 |
-
, we compile [mCoT-MATH](https://huggingface.co/datasets/laihuiyuan/mCoT-MATH) using machine translation.
|
33 |
-
|
34 |
-
| Language | SW | BN | TE | TH | JA | ZH | RU | ES | FR | DE | DE |Overall |
|
35 |
-
|:----------|:------|:------|:------|:------|:------|:------|:------|:------|:------|:------|:------|--------|
|
36 |
-
| mCoT-MATH | ~580K | ~580K | ~580K | ~580K | ~580K | ~580K | ~580K | ~580K | ~580K | ~580K | ~580K | ~6.3M |
|
37 |
|
38 |
### Results on [MGSM](https://arxiv.org/abs/2210.03057v1)
|
39 |
| Language | SW | BN | TE | TH | JA | ZH | RU | ES | FR | DE | EN |
|
@@ -68,7 +61,10 @@ Based on [MetaMathQA](https://github.com/meta-math/MetaMath) and [MathInstruct](
|
|
68 |
|
69 |
### Prompt Template
|
70 |
```bash
|
71 |
-
#
|
|
|
|
|
|
|
72 |
bn = "আসুন ধাপে ধাপে চিন্তা করি।"
|
73 |
de = "Denken wir Schritt für Schritt."
|
74 |
en = "Let's think step by step."
|
@@ -81,10 +77,11 @@ te = "అంచెలంచెలుగా ఆలోచిద్దాం."
|
|
81 |
th = "ลองคิดทีละขั้นตอน"
|
82 |
zh = "让我们一步步思考。"
|
83 |
|
84 |
-
# Math
|
85 |
-
|
86 |
|
87 |
-
|
|
|
88 |
```
|
89 |
|
90 |
### Citation
|
|
|
17 |
---
|
18 |
# mCoT: Multilingual Instruction Tuning for Reasoning Consistency in Language Models
|
19 |
|
20 |
+
Paper: https://arxiv.org/abs/2406.02301
|
21 |
|
22 |
+
Code: https://github.com/laihuiyuan/mCoT
|
23 |
|
24 |
+
Dataset: https://huggingface.co/datasets/laihuiyuan/mCoT-MATH
|
25 |
|
26 |
### Introduction
|
27 |
+
We introduce mCoT, a 7B parameter model for multilingual math reasoning that achieves impressive multilingual reasoning consistency across multiple languages.
|
28 |
+
Based on [Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1), mCoT is trained on [mCoT-MATH](https://huggingface.co/datasets/laihuiyuan/mCoT-MATH), the first large-scale multilingual math CoT reasoning dataset containing around 6.3 million samples for 11 diverse languages.
|
29 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
|
31 |
### Results on [MGSM](https://arxiv.org/abs/2210.03057v1)
|
32 |
| Language | SW | BN | TE | TH | JA | ZH | RU | ES | FR | DE | EN |
|
|
|
61 |
|
62 |
### Prompt Template
|
63 |
```bash
|
64 |
+
# Template
|
65 |
+
template = "Question: \n{question} \nAnswer: \n{language}\n"
|
66 |
+
|
67 |
+
# Language prompt
|
68 |
bn = "আসুন ধাপে ধাপে চিন্তা করি।"
|
69 |
de = "Denken wir Schritt für Schritt."
|
70 |
en = "Let's think step by step."
|
|
|
77 |
th = "ลองคิดทีละขั้นตอน"
|
78 |
zh = "让我们一步步思考。"
|
79 |
|
80 |
+
# Math question
|
81 |
+
math_en = "A robe takes 2 bolts of blue fiber and half that much white fiber. How many bolts in total does it take?"
|
82 |
|
83 |
+
# An example for the English question
|
84 |
+
prompt = template.format(question=math_en, language=en)
|
85 |
```
|
86 |
|
87 |
### Citation
|