Commit
•
5d2bf3f
1
Parent(s):
ee2ced6
Update README.md
Browse files
README.md
CHANGED
@@ -7,25 +7,18 @@ tags:
|
|
7 |
datasets:
|
8 |
- akjindal53244/Arithmo-Data
|
9 |
---
|
10 |
-
# Model Card for Model ID
|
11 |
|
12 |
-
|
13 |
-
|
14 |
-
|
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 [Arithmo-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 |
-
<span style="color:red"><ins>Note</ins></span>: LoRA adapter of Arithmo2-7B model is also available here: https://huggingface.co/upaya07/Arithmo2-7B-adapter
|
23 |
|
|
|
24 |
|
25 |
|
26 |
### Model Description
|
27 |
|
28 |
-
- **Project GitHub Page:** https://github.com/akjindal53244/Arithmo
|
29 |
- **Developed by:** [Ashvini Kumar Jindal](https://www.linkedin.com/in/ashvini-jindal-26653262/)
|
30 |
- **Funded by:** self-work
|
31 |
- **Model type:** fine-tuned using QLoRA on Single GPU
|
@@ -34,7 +27,7 @@ datasets:
|
|
34 |
|
35 |
## Results
|
36 |
|
37 |
-
Arithmo2-7B is improved version of [Arithmo-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
|
38 |
|
39 |
<table>
|
40 |
<thead>
|
@@ -88,7 +81,7 @@ pip install bitsandbytes
|
|
88 |
|
89 |
$ python query_model.py
|
90 |
```
|
91 |
-
**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
|
92 |
|
93 |
##### Sample Input:
|
94 |
```
|
@@ -106,7 +99,7 @@ Plugging these values into the formula, we get:
|
|
106 |
The answer is: 55
|
107 |
```
|
108 |
|
109 |
-
Arithmo2-7B is trained with the following format:
|
110 |
#### CoT Format (generate reasoning steps with answer):
|
111 |
```
|
112 |
Question: <question>
|
@@ -122,8 +115,8 @@ Answer:
|
|
122 |
```
|
123 |
It will perform best if queried in this way with your own script.
|
124 |
|
125 |
-
## Comparing
|
126 |
-
Results for all models except `Arithmo2-7B` are taken from [MetaMath](https://github.com/meta-math/MetaMath/blob/main/README.MD) repository.
|
127 |
|
128 |
| Model | GSM8k Pass@1 | MATH Pass@1 | Fine-tuning |
|
129 |
|---------------------|--------------|-------------|-------------|
|
@@ -154,12 +147,12 @@ Results for all models except `Arithmo2-7B` are taken from [MetaMath](https://gi
|
|
154 |
| Arithmo-Mistral-7B Zero-Shot PoT | 71.2 | -- | SFT: 4-bit QLoRA |
|
155 |
| Arithmo-Mistral-7B Zero-Shot CoT | 74.7 | 25.3 | SFT: 4-bit QLoRA |
|
156 |
| MetaMath-Mistral-7B | 77.7 | 28.2 | SFT: Full fine-tuned |
|
157 |
-
| 🔥 **Arithmo2-7B Zero-Shot PoT** | **74.2** | -- | **SFT: 4-bit QLoRA** |
|
158 |
-
| 🔥 **Arithmo2-7B Zero-Shot CoT** | **76.4** | **27.2** | **SFT: 4-bit QLoRA** |
|
159 |
|
160 |
|
161 |
|
162 |
-
If you are interested in reproducing the resullts, visit https://github.com/akjindal53244/Arithmo
|
163 |
|
164 |
<h2 id="References">References</h2>
|
165 |
|
|
|
7 |
datasets:
|
8 |
- akjindal53244/Arithmo-Data
|
9 |
---
|
|
|
10 |
|
11 |
+
**Arithmo2-Mistral-7B** model improves initially released Arithmo-Mistral-7B model on both GSM8K and MATH benchmarks. Specifically, there is **absolute** improvement of:
|
12 |
+
- +1.7% on GSM8K
|
13 |
+
- +3.0% on GSM8K PoT
|
14 |
+
- +1.9% on MATH
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
|
16 |
+
We release both [merged model](https://huggingface.co/upaya07/Arithmo2-Mistral-7B) and [LoRA Adapter](https://huggingface.co/upaya07/Arithmo2-Mistral-7B-adapter).
|
17 |
|
18 |
|
19 |
### Model Description
|
20 |
|
21 |
+
- **Project GitHub Page:** https://github.com/akjindal53244/Arithmo
|
22 |
- **Developed by:** [Ashvini Kumar Jindal](https://www.linkedin.com/in/ashvini-jindal-26653262/)
|
23 |
- **Funded by:** self-work
|
24 |
- **Model type:** fine-tuned using QLoRA on Single GPU
|
|
|
27 |
|
28 |
## Results
|
29 |
|
30 |
+
Arithmo2-Mistral-7B is improved version of [Arithmo-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 models with other SFT LLM models](https://github.com/akjindal53244/Arithmo/tree/master?tab=readme-ov-file#comparing-arithmo-models-with-other-sft-llm-models) section for more details.
|
31 |
|
32 |
<table>
|
33 |
<thead>
|
|
|
81 |
|
82 |
$ python query_model.py
|
83 |
```
|
84 |
+
**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/blob/master/query_model.py) for more details. <br><br>
|
85 |
|
86 |
##### Sample Input:
|
87 |
```
|
|
|
99 |
The answer is: 55
|
100 |
```
|
101 |
|
102 |
+
Arithmo2-Mistral-7B is trained with the following format:
|
103 |
#### CoT Format (generate reasoning steps with answer):
|
104 |
```
|
105 |
Question: <question>
|
|
|
115 |
```
|
116 |
It will perform best if queried in this way with your own script.
|
117 |
|
118 |
+
## Comparing Arithmo models with other SFT LLM models
|
119 |
+
Results for all models except `Arithmo2-Mistral-7B` are taken from [MetaMath](https://github.com/meta-math/MetaMath/blob/main/README.MD) repository.
|
120 |
|
121 |
| Model | GSM8k Pass@1 | MATH Pass@1 | Fine-tuning |
|
122 |
|---------------------|--------------|-------------|-------------|
|
|
|
147 |
| Arithmo-Mistral-7B Zero-Shot PoT | 71.2 | -- | SFT: 4-bit QLoRA |
|
148 |
| Arithmo-Mistral-7B Zero-Shot CoT | 74.7 | 25.3 | SFT: 4-bit QLoRA |
|
149 |
| MetaMath-Mistral-7B | 77.7 | 28.2 | SFT: Full fine-tuned |
|
150 |
+
| 🔥 **Arithmo2-Mistral-7B Zero-Shot PoT** | **74.2** | -- | **SFT: 4-bit QLoRA** |
|
151 |
+
| 🔥 **Arithmo2-Mistral-7B Zero-Shot CoT** | **76.4** | **27.2** | **SFT: 4-bit QLoRA** |
|
152 |
|
153 |
|
154 |
|
155 |
+
If you are interested in reproducing the resullts, visit https://github.com/akjindal53244/Arithmo#reproducing-results section.
|
156 |
|
157 |
<h2 id="References">References</h2>
|
158 |
|