Commit
•
fc46879
1
Parent(s):
fb3c3be
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,184 @@
|
|
1 |
---
|
2 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: mit
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
tags:
|
6 |
+
- Mathematical Reasoning
|
7 |
+
datasets:
|
8 |
+
- akjindal53244/Arithmo-Data
|
9 |
---
|
10 |
+
# Model Card for Model ID
|
11 |
+
|
12 |
+
[![Code License](https://img.shields.io/badge/Code%20License-Apache_2.0-green.svg)](CODE_LICENSE)
|
13 |
+
[![Model Weight License](https://img.shields.io/badge/Model%20Weights%20License-Apache_2.0-green.svg)](LICENSE)
|
14 |
+
[![Python 3.9+](https://img.shields.io/badge/python-3.9+-blue.svg)](https://www.python.org/downloads/release/python-390/)
|
15 |
+
|
16 |
+
**P.S.:** Please reach out to [Ashvini Jindal](https://www.linkedin.com/in/ashvini-jindal-26653262/) if you would be interested in supporting compute need. We are looking for small-scale support so we'd appreciate any kind of help! :)
|
17 |
+
|
18 |
+
## Model Details
|
19 |
+
|
20 |
+
**Arithmo2-7B** is improved version of [Arithmi-Mistral-7B](https://huggingface.co/akjindal53244/Arithmo-Mistral-7B) model and is trained to reason and answer mathematical problems and is also capable of writing a Python program that upon execution prints answer to the question. We used [Mistral-7B](https://huggingface.co/mistralai/Mistral-7B-v0.1) as a base model and used **QLoRA to fine-tune it on a single GPU**.
|
21 |
+
|
22 |
+
### Model Description
|
23 |
+
|
24 |
+
- **Project GitHub Page:** https://github.com/akjindal53244/Arithmo-Mistral-7B
|
25 |
+
- **Developed by:** [Ashvini Kumar Jindal](https://www.linkedin.com/in/ashvini-jindal-26653262/)
|
26 |
+
- **Funded by:** self-work
|
27 |
+
- **Model type:** fine-tuned using QLoRA on Single GPU
|
28 |
+
- **Language(s) (NLP):** English
|
29 |
+
- **Finetuned from model:** mistralai/Mistral-7B-v0.1
|
30 |
+
|
31 |
+
## Results
|
32 |
+
|
33 |
+
Arithmo2-7B is improved version of [Arithmi-Mistral-7B](https://huggingface.co/akjindal53244/Arithmo-Mistral-7B) model and is competitive with full fine-tuned state-of-the-art 7B Mathematical Reasoning models. Refer to [Comparing Arithmo-Mistral-7B with other LLM models](https://github.com/akjindal53244/Arithmo-Mistral-7B/tree/master#comparing-arithmo-mistral-7b-with-other-llm-models) section for more details.
|
34 |
+
|
35 |
+
<table>
|
36 |
+
<thead>
|
37 |
+
<tr>
|
38 |
+
<th>Prompt Approach</th>
|
39 |
+
<th>GSM8k</th>
|
40 |
+
<th>MATH</th>
|
41 |
+
</tr>
|
42 |
+
</thead>
|
43 |
+
<tbody>
|
44 |
+
<tr>
|
45 |
+
<td>Zero-Shot CoT</td>
|
46 |
+
<td><b>76.4</b></td>
|
47 |
+
<td><b>27.2</b></td>
|
48 |
+
</tr>
|
49 |
+
<tr>
|
50 |
+
<td>Zero-Shot PoT</td>
|
51 |
+
<td><b>74.2</b></td>
|
52 |
+
<td>-</td>
|
53 |
+
</tr>
|
54 |
+
</tbody>
|
55 |
+
</table>
|
56 |
+
|
57 |
+
- **Zero-Shot CoT**: On providing a question as prompt, model generates reasoning steps to solve the question along with answer. We check if answer matches with ground-truth.
|
58 |
+
- **Zero-Shot PoT**: We prompt the model to generate a Python program for the given question. During inference, we execute the Python program generated by the model and check if the program output matches with ground-truth answer.
|
59 |
+
|
60 |
+
|
61 |
+
## Installation
|
62 |
+
|
63 |
+
```
|
64 |
+
pip install transformers >=4.34.0
|
65 |
+
pip install accelerate
|
66 |
+
pip install sentencepiece
|
67 |
+
pip install protobuf
|
68 |
+
|
69 |
+
# If you are GPU poor like me
|
70 |
+
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
|
71 |
+
|
72 |
+
# If you have a GPU.
|
73 |
+
pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu118
|
74 |
+
pip install scipy
|
75 |
+
pip install bitsandbytes
|
76 |
+
```
|
77 |
+
|
78 |
+
|
79 |
+
## How to query the model
|
80 |
+
|
81 |
+
```
|
82 |
+
# Set `run_model_on_gpu` to `False` if you are running on CPU. Model will generate reasoning steps with answer for your question. If you want to generate Python program, uncomment line-69 that adds a Python prompt.
|
83 |
+
# This script automatically does formatting for you, so you just need to type question (eg: `What is 2+2?`) without any prefix like `Question:`, etc.**
|
84 |
+
|
85 |
+
$ python query_model.py
|
86 |
+
```
|
87 |
+
**Note:** Above script automatically does formatting for you, so you just need to type question (eg: `What is 2+2?`) without any prefix like `Question:`, etc. Checkout [query_model.py](https://github.com/akjindal53244/Arithmo-Mistral-7B/blob/master/query_model.py) for more details. <br><br>
|
88 |
+
|
89 |
+
##### Sample Input:
|
90 |
+
```
|
91 |
+
Question: There are total 10 children. I have to give 1 apple to first child, 2 apples to second child, 3 apples to third child, and so on. How many apples do I need?
|
92 |
+
```
|
93 |
+
##### Model Output:
|
94 |
+
```
|
95 |
+
Answer: The total number of apples needed is the sum of the first 10 positive integers.
|
96 |
+
This can be calculated using the formula for the sum of an arithmetic series:
|
97 |
+
\[S = \frac{n}{2}(a_1 + a_n),\]
|
98 |
+
where $S$ is the sum, $n$ is the number of terms, $a_1$ is the first term, and $a_n$ is the last term.
|
99 |
+
In this case, $n = 10$, $a_1 = 1$, and $a_n = 10$.
|
100 |
+
Plugging these values into the formula, we get:
|
101 |
+
\[S = \frac{10}{2}(1 + 10) = 5(11) = \boxed{55}.\]
|
102 |
+
The answer is: 55
|
103 |
+
```
|
104 |
+
|
105 |
+
Arithmo-Mistral-7B is trained with the following format:
|
106 |
+
#### CoT Format (generate reasoning steps with answer):
|
107 |
+
```
|
108 |
+
Question: <question>
|
109 |
+
|
110 |
+
Answer:
|
111 |
+
```
|
112 |
+
|
113 |
+
#### PoT Format (generate a python program):
|
114 |
+
```
|
115 |
+
Question: <question> <python_prompt>
|
116 |
+
|
117 |
+
Answer:
|
118 |
+
```
|
119 |
+
It will perform best if queried in this way with your own script.
|
120 |
+
|
121 |
+
## Comparing Arithmo-Mistral-7B with other LLM models.
|
122 |
+
Results for all models except `Arithmo2-7B` are taken from [MetaMath](https://github.com/meta-math/MetaMath/blob/main/README.MD) repository.
|
123 |
+
|
124 |
+
| Model | GSM8k Pass@1 | MATH Pass@1 | Fine-tuning |
|
125 |
+
|---------------------|--------------|-------------|-------------|
|
126 |
+
| MPT-7B | 6.8 | 3.0 |
|
127 |
+
| Falcon-7B | 6.8 | 2.3 |
|
128 |
+
| LLaMA-1-7B | 11.0 | 2.9 |
|
129 |
+
| LLaMA-2-7B | 14.6 | 2.5 |
|
130 |
+
| MPT-30B | 15.2 | 3.1 |
|
131 |
+
| LLaMA-1-13B | 17.8 | 3.9 |
|
132 |
+
| GPT-Neo-2.7B | 19.5 | -- |
|
133 |
+
| Falcon-40B | 19.6 | 2.5 |
|
134 |
+
| Baichuan-chat-13B | 23.9 | -- |
|
135 |
+
| Vicuna-v1.3-13B | 27.6 | -- |
|
136 |
+
| LLaMA-2-13B | 28.7 | 3.9 |
|
137 |
+
| InternLM-7B | 31.2 | -- |
|
138 |
+
| ChatGLM-2-6B | 32.4 | -- |
|
139 |
+
| GPT-J-6B | 34.9 | -- |
|
140 |
+
| LLaMA-1-33B | 35.6 | 3.9 |
|
141 |
+
| LLaMA-2-34B | 42.2 | 6.24 |
|
142 |
+
| RFT-7B | 50.3 | -- |
|
143 |
+
| LLaMA-1-65B | 50.9 | 10.6 |
|
144 |
+
| Qwen-7B | 51.6 | -- |
|
145 |
+
| WizardMath-7B | 54.9 | 10.7 |
|
146 |
+
| LLaMA-2-70B | 56.8 | 13.5 |
|
147 |
+
| WizardMath-13B | 63.9 | 14.0 |
|
148 |
+
| MetaMath-7B | 66.5 | 19.8 |
|
149 |
+
| MetaMath-13B | 72.3 | 22.4 |
|
150 |
+
| Arithmo-Mistral-7B Zero-Shot PoT | 71.2 | -- | SFT: 4-bit QLoRA |
|
151 |
+
| Arithmo-Mistral-7B Zero-Shot CoT | 74.7 | 25.3 | SFT: 4-bit QLoRA |
|
152 |
+
| MetaMath-Mistral-7B | 77.7 | 27.2 | SFT: Full fine-tuned |
|
153 |
+
| 🔥 **Arithmo2-7B Zero-Shot PoT** | **74.2** | -- | **SFT: 4-bit QLoRA** |
|
154 |
+
| 🔥 **Arithmo2-7B Zero-Shot CoT** | **76.4** | **27.2** | **SFT: 4-bit QLoRA** |
|
155 |
+
|
156 |
+
|
157 |
+
|
158 |
+
If you are interested in reproducing the resullts, visit https://github.com/akjindal53244/Arithmo-Mistral-7B#reproducing-results section.
|
159 |
+
|
160 |
+
<h2 id="References">References</h2>
|
161 |
+
|
162 |
+
```
|
163 |
+
@article{yu2023metamath,
|
164 |
+
title={MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models},
|
165 |
+
author={Yu, Longhui and Jiang, Weisen and Shi, Han and Yu, Jincheng and Liu, Zhengying and Zhang, Yu and Kwok, James T and Li, Zhenguo and Weller, Adrian and Liu, Weiyang},
|
166 |
+
journal={arXiv preprint arXiv:2309.12284},
|
167 |
+
year={2023}
|
168 |
+
}
|
169 |
+
|
170 |
+
@article{Yue2023mammoth,
|
171 |
+
title={MAmmoTH: Building math generalist models through hybrid instruction tuning},
|
172 |
+
author={Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen},
|
173 |
+
journal={arXiv preprint arXiv:2309.05653},
|
174 |
+
year={2023}
|
175 |
+
}
|
176 |
+
|
177 |
+
@article{mishra2022lila,
|
178 |
+
title={Lila: A unified benchmark for mathematical reasoning},
|
179 |
+
author={Swaroop Mishra, Matthew Finlayson, Pan Lu, Leonard Tang, Sean Welleck, Chitta Baral, Tanmay Rajpurohit, Oyvind Tafjord, Ashish Sabharwal, Peter Clark, and Ashwin Kalyan},
|
180 |
+
journal={arXiv preprint arXiv:2210.17517},
|
181 |
+
year={2022}
|
182 |
+
}
|
183 |
+
|
184 |
+
```
|