Update README.md
Browse files
README.md
CHANGED
@@ -1,199 +1,174 @@
|
|
1 |
---
|
|
|
|
|
|
|
2 |
library_name: transformers
|
3 |
-
tags:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
---
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
-
|
21 |
-
- **
|
22 |
-
- **
|
23 |
-
-
|
24 |
-
-
|
25 |
-
- **
|
26 |
-
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
#### Speeds, Sizes, Times [optional]
|
98 |
-
|
99 |
-
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
|
100 |
-
|
101 |
-
[More Information Needed]
|
102 |
-
|
103 |
-
## Evaluation
|
104 |
-
|
105 |
-
<!-- This section describes the evaluation protocols and provides the results. -->
|
106 |
-
|
107 |
-
### Testing Data, Factors & Metrics
|
108 |
-
|
109 |
-
#### Testing Data
|
110 |
-
|
111 |
-
<!-- This should link to a Dataset Card if possible. -->
|
112 |
-
|
113 |
-
[More Information Needed]
|
114 |
-
|
115 |
-
#### Factors
|
116 |
-
|
117 |
-
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
|
118 |
-
|
119 |
-
[More Information Needed]
|
120 |
-
|
121 |
-
#### Metrics
|
122 |
-
|
123 |
-
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
|
124 |
-
|
125 |
-
[More Information Needed]
|
126 |
-
|
127 |
-
### Results
|
128 |
-
|
129 |
-
[More Information Needed]
|
130 |
-
|
131 |
-
#### Summary
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
## Model Examination [optional]
|
136 |
-
|
137 |
-
<!-- Relevant interpretability work for the model goes here -->
|
138 |
-
|
139 |
-
[More Information Needed]
|
140 |
-
|
141 |
-
## Environmental Impact
|
142 |
-
|
143 |
-
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
|
144 |
-
|
145 |
-
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
|
146 |
-
|
147 |
-
- **Hardware Type:** [More Information Needed]
|
148 |
-
- **Hours used:** [More Information Needed]
|
149 |
-
- **Cloud Provider:** [More Information Needed]
|
150 |
-
- **Compute Region:** [More Information Needed]
|
151 |
-
- **Carbon Emitted:** [More Information Needed]
|
152 |
-
|
153 |
-
## Technical Specifications [optional]
|
154 |
-
|
155 |
-
### Model Architecture and Objective
|
156 |
-
|
157 |
-
[More Information Needed]
|
158 |
-
|
159 |
-
### Compute Infrastructure
|
160 |
-
|
161 |
-
[More Information Needed]
|
162 |
-
|
163 |
-
#### Hardware
|
164 |
-
|
165 |
-
[More Information Needed]
|
166 |
-
|
167 |
-
#### Software
|
168 |
-
|
169 |
-
[More Information Needed]
|
170 |
-
|
171 |
-
## Citation [optional]
|
172 |
-
|
173 |
-
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
|
174 |
-
|
175 |
-
**BibTeX:**
|
176 |
-
|
177 |
-
[More Information Needed]
|
178 |
-
|
179 |
-
**APA:**
|
180 |
-
|
181 |
-
[More Information Needed]
|
182 |
-
|
183 |
-
## Glossary [optional]
|
184 |
-
|
185 |
-
<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
|
186 |
-
|
187 |
-
[More Information Needed]
|
188 |
-
|
189 |
-
## More Information [optional]
|
190 |
-
|
191 |
-
[More Information Needed]
|
192 |
-
|
193 |
-
## Model Card Authors [optional]
|
194 |
-
|
195 |
-
[More Information Needed]
|
196 |
-
|
197 |
-
## Model Card Contact
|
198 |
-
|
199 |
-
[More Information Needed]
|
|
|
1 |
---
|
2 |
+
language:
|
3 |
+
- en
|
4 |
+
license: llama3
|
5 |
library_name: transformers
|
6 |
+
tags:
|
7 |
+
- mathematics
|
8 |
+
datasets:
|
9 |
+
- hkust-nlp/dart-math-uniform
|
10 |
+
metrics:
|
11 |
+
- accuracy
|
12 |
+
pipeline_tag: text-generation
|
13 |
+
base_model: meta-llama/Meta-Llama-3-70B
|
14 |
+
model-index:
|
15 |
+
- name: dart-math-llama3-70b-uniform
|
16 |
+
results:
|
17 |
+
- task:
|
18 |
+
type: text-generation
|
19 |
+
name: Mathematical Problem-Solving
|
20 |
+
dataset:
|
21 |
+
type: hendrycks/competition_math
|
22 |
+
name: MATH
|
23 |
+
split: test
|
24 |
+
metrics:
|
25 |
+
- type: accuracy
|
26 |
+
name: Pass@1 (0-shot CoT)
|
27 |
+
value: 54.9
|
28 |
+
- task:
|
29 |
+
type: text-generation
|
30 |
+
name: Mathematical Problem-Solving
|
31 |
+
dataset:
|
32 |
+
type: openai/gsm8k
|
33 |
+
name: GSM8K
|
34 |
+
config: main
|
35 |
+
split: test
|
36 |
+
metrics:
|
37 |
+
- type: accuracy
|
38 |
+
name: Pass@1 (0-shot CoT)
|
39 |
+
value: 90.4
|
40 |
+
- task:
|
41 |
+
type: text-generation
|
42 |
+
name: Mathematical Problem-Solving
|
43 |
+
dataset:
|
44 |
+
type: college-math
|
45 |
+
name: CollegeMath
|
46 |
+
metrics:
|
47 |
+
- type: accuracy
|
48 |
+
name: Pass@1 (0-shot CoT)
|
49 |
+
value: 38.5
|
50 |
+
- task:
|
51 |
+
type: text-generation
|
52 |
+
name: Mathematical Problem-Solving
|
53 |
+
dataset:
|
54 |
+
type: deepmind-mathematics
|
55 |
+
name: DeepMind-Mathematics
|
56 |
+
metrics:
|
57 |
+
- type: accuracy
|
58 |
+
name: Pass@1 (0-shot CoT)
|
59 |
+
value: 64.1
|
60 |
+
- task:
|
61 |
+
type: text-generation
|
62 |
+
name: Mathematical Problem-Solving
|
63 |
+
dataset:
|
64 |
+
type: Hothan/OlympiadBench
|
65 |
+
name: OlympiadBench-OE_TO_maths_en_COMP
|
66 |
+
config: OE_TO_maths_en_COMP
|
67 |
+
split: train
|
68 |
+
metrics:
|
69 |
+
- type: accuracy
|
70 |
+
name: Pass@1 (0-shot CoT)
|
71 |
+
value: 19.1
|
72 |
+
- task:
|
73 |
+
type: text-generation
|
74 |
+
name: Mathematical Problem-Solving
|
75 |
+
dataset:
|
76 |
+
type: TIGER-Lab/TheoremQA
|
77 |
+
name: TheoremQA
|
78 |
+
split: test
|
79 |
+
metrics:
|
80 |
+
- type: accuracy
|
81 |
+
name: Pass@1 (0-shot CoT)
|
82 |
+
value: 27.4
|
83 |
---
|
84 |
+
# DART-Math: Difficulty-Aware Rejection Tuning for Mathematical Problem-Solving
|
85 |
+
|
86 |
+
📝 [Paper@arXiv](https://arxiv.org/abs/2407.13690) | 🤗 [Datasets&Models@HF](https://huggingface.co/collections/hkust-nlp/dart-math-665704599b35de59f8fdf6c1) | 🐱 [Code@GitHub](https://github.com/hkust-nlp/dart-math) | 🐦 [Thread@X(Twitter)](https://x.com/tongyx361/status/1811413243350454455) | 🐶 [中文博客@知乎](https://zhuanlan.zhihu.com/p/708371895) | 📑 [BibTeX](https://github.com/hkust-nlp/dart-math?tab=readme-ov-file#citation)
|
87 |
+
|
88 |
+
## Models: `DART-Math`
|
89 |
+
|
90 |
+
`DART-Math` models achieve performance **superior or competitive to previous SOTAs** on 2 in-domain and 4 challenging out-of-domain mathematical reasoning benchmarks, despite using **much smaller datasets** and **no proprietary model like GPT-4**.
|
91 |
+
|
92 |
+
| Model | [MATH](https://huggingface.co/datasets/hendrycks/competition_math) | [GSM8K](https://huggingface.co/datasets/gsm8k) | [College](https://github.com/hkust-nlp/dart-math/tree/main/data/eval-dsets/mwpbench/college-math-test.jsonl) | [DM](https://github.com/hkust-nlp/dart-math/tree/main/data/eval-dsets/deepmind-mathematics.json) | [Olympiad](https://github.com/hkust-nlp/dart-math/tree/main/data/eval-dsets/olympiadbench/OE_TO_maths_en_COMP.json) | [Theorem](https://github.com/hkust-nlp/dart-math/tree/main/data/eval-dsets/theoremqa.json) | AVG |
|
93 |
+
| :----------------------------------------------------------------------------------------------------- | -----------------------------------------------------------------: | ---------------------------------------------: | -----------------------------------------------------------------------------------------------------------: | -----------------------------------------------------------------------------------------------: | ------------------------------------------------------------------------------------------------------------------: | -----------------------------------------------------------------------------------------: | -------: |
|
94 |
+
| GPT-4 (0314) | [52.6](https://arxiv.org/abs/2403.04706) | [94.7](https://arxiv.org/abs/2403.04706) | [24.4](https://arxiv.org/abs/2403.02884) | -- | -- | -- | -- |
|
95 |
+
| Llama-3-70B-MetaMath | 44.9 | 88.0 | 31.9 | 53.2 | 11.6 | 21.9 | 41.9 |
|
96 |
+
| [`DART-Math-Llama-3-70B` (Uniform)](https://huggingface.co/hkust-nlp/dart-math-llama3-70b-uniform) | 54.9 | **90.4** | **38.5** | **64.1** | 19.1 | 27.4 | 49.1 |
|
97 |
+
| [`DART-Math-Llama-3-70B` (Prop2Diff)](https://huggingface.co/hkust-nlp/dart-math-llama3-70b-prop2diff) | **56.1** | 89.6 | 37.9 | **64.1** | **20.0** | **28.2** | **49.3** |
|
98 |
+
| DeepSeekMath-7B-MetaMath | 43.7 | 81.8 | 33.7 | 53.0 | 13.6 | 23.2 | 41.5 |
|
99 |
+
| [DeepSeekMath-7B-RL](https://huggingface.co/deepseek-ai/deepseek-math-7b-rl) | 53.1 | 88.4 | 41.3 | 58.3 | 18.7 | 35.9 | 49.3 |
|
100 |
+
| [`DART-Math-DSMath-7B` (Uniform)](https://huggingface.co/hkust-nlp/dart-math-dsmath-7b-uniform) | 52.9 | **88.2** | 40.1 | 60.2 | 21.3 | **32.5** | 49.2 |
|
101 |
+
| [`DART-Math-DSMath-7B` (Prop2Diff)](https://huggingface.co/hkust-nlp/dart-math-dsmath-7b-prop2diff) | **53.6** | 86.8 | **40.7** | **61.6** | **21.7** | 32.2 | **49.4** |
|
102 |
+
| Mistral-7B-MetaMath | 29.8 | 76.5 | 19.3 | 28.0 | 5.9 | 14.0 | 28.9 |
|
103 |
+
| [`DART-Math-Mistral-7B` (Uniform)](https://huggingface.co/hkust-nlp/dart-math-mistral-7b-uniform) | 43.5 | **82.6** | 26.9 | 42.0 | 13.2 | 16.4 | 27.4 |
|
104 |
+
| [`DART-Math-Mistral-7B` (Prop2Diff)](https://huggingface.co/hkust-nlp/dart-math-mistral-7b-prop2diff) | **45.5** | 81.1 | **29.4** | **45.1** | **14.7** | **17.0** | **38.8** |
|
105 |
+
| Llama-3-8B-MetaMath | 32.5 | 77.3 | 20.6 | 35.0 | 5.5 | 13.8 | 30.8 |
|
106 |
+
| [`DART-Math-Llama-3-8B` (Uniform)](https://huggingface.co/hkust-nlp/dart-math-llama3-8b-uniform) | 45.3 | **82.5** | 27.1 | **48.2** | 13.6 | 15.4 | 38.7 |
|
107 |
+
| [`DART-Math-Llama-3-8B` (Prop2Diff)](https://huggingface.co/hkust-nlp/dart-math-llama3-8b-prop2diff) | **46.6** | 81.1 | **28.8** | 48.0 | **14.5** | **19.4** | **39.7** |
|
108 |
+
|
109 |
+
***Abbreviations**: College (CollegeMath), DM (DeepMind Mathematics), Olympiad (OlympiadBench-Math), Theorem (TheoremQA).
|
110 |
+
**Bold** means the best score by SFT on the respective base model here.
|
111 |
+
To reproduce our results, please refer to [the `DART-Math` GitHub repository](https://github.com/hkust-nlp/dart-math).*
|
112 |
+
## Prompt Template
|
113 |
+
All the `DART-Math` models use the [Alpaca](https://github.com/tatsu-lab/stanford_alpaca) prompt template:
|
114 |
+
```
|
115 |
+
Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n###Instruction:\n{query}\n\n### Response:\n
|
116 |
+
```
|
117 |
+
## Training Dataset
|
118 |
+
We construct our traning datasets by applying **Difficulty-Aware Rejection Sampling** (`DARS`) to the **MATH and GSM8K** training sets.
|
119 |
+
|
120 |
+
`DARS` tackle **severe biases towards easy queries, with frequent failures to generate any correct response for the most challenging queries**, in previous datasets.
|
121 |
+
|
122 |
+
These biases are primarily caused by vanilla rejection sampling, where **the same number of responses is
|
123 |
+
sampled for each query**, yet the likelihood of obtaining correct responses for difficult queries is significantly lower, sometimes even zero.
|
124 |
+
|
125 |
+
Please refer to [`DART-Math-Hard`](https://huggingface.co/datasets/hkust-nlp/dart-math-hard) / [`DART-Math-Uniform`](https://huggingface.co/datasets/hkust-nlp/dart-math-uniform) for more details.
|
126 |
+
|
127 |
+
## Training Setup
|
128 |
+
|
129 |
+
We perform standard instruction tuning to several base models including Llama3-8B & Mistral-7B & Llama3-70B as representatives of general models and DeepSeekMath-
|
130 |
+
7B as the representative of math-specialized model
|
131 |
+
on our synthetic datasets [`DART-Math-Hard`](https://huggingface.co/datasets/hkust-nlp/dart-math-hard) & [`DART-Math-Uniform`](https://huggingface.co/datasets/hkust-nlp/dart-math-uniform),
|
132 |
+
leading to `DART-Math (Prop2Diff)` & `DART-Math (Uniform)` respectively.
|
133 |
+
|
134 |
+
For simplicity, we keep most hyper-parameters the same across different models and datasets:
|
135 |
+
|
136 |
+
- Model max length (of [packed](https://github.com/MeetKai/functionary/tree/main/functionary/train/packing) sequence): 4096
|
137 |
+
- Batch size: 64
|
138 |
+
- Warm-up ratio: 0.03
|
139 |
+
- Learning rate scheduler: cosine
|
140 |
+
- Prompt template: [Alpaca](https://github.com/tatsu-lab/stanford_alpaca)
|
141 |
+
|
142 |
+
Several other key hyper-parameters are tuned as follow:
|
143 |
+
|
144 |
+
| Base Model | Max. L.R. | # of Epochs | # of Grad. Acc. Steps | # of A100 GPUs |
|
145 |
+
|:--------------- | ---------:| -----------:| ---------------------:| --------------:|
|
146 |
+
| Mistral-7B | `1e-5` | 3 | 1 | 8 |
|
147 |
+
| Llama3-8B | `5e-5` | 1 | 2 | 8 |
|
148 |
+
| Llama3-70B | `2e-5` | 1 | 1 | 32 |
|
149 |
+
| DeepSeekMath-7B | `5e-5` | 3 | 1 | 8 |
|
150 |
+
|
151 |
+
- For **maximum learning rate**, we determine the values by **searching** through `1e-6,5e-6,1e-5,2e-5,5e-5,1e-4` according to the MATH performance after training on MMIQC for 1 epoch, except for Llama3-70B that is so expensive to search for that we derive from Llama3-8B’s learning rate in analogy to the relationship of (per-training) learning rates between [Llama2-7B](https://huggingface.co/meta-llama/Llama-2-7b-hf) and [Llama2-70B](https://huggingface.co/meta-llama/Llama-2-70b-hf) (\~2:1).
|
152 |
+
- For **Llama3** models, preliminary experiments indicate that **training for 1 epoch consistently outperforms 3 epochs**.
|
153 |
+
|
154 |
+
Please refer to [Appendix A.1 of our paper](https://tongyx361.github.io/assets/dart-math/paper-dart-math.pdf) for more details.
|
155 |
+
|
156 |
+
## Other Details
|
157 |
+
|
158 |
+
- For Mistral-7B-based models, we disable `sliding_window` by default following [the newest Mistral-7B-Instruct](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3/blob/main/config.json) (Flash Attention 2 does not support `sliding_window` and XFormer backend in vLLM has throughput \~10% lower in our experiments.)
|
159 |
+
|
160 |
+
## Citation
|
161 |
+
|
162 |
+
If you find our data, model or code useful for your work, please kindly cite [our paper](https://arxiv.org/abs/2407.13690):
|
163 |
+
|
164 |
+
```latex
|
165 |
+
@article{tong2024dartmath,
|
166 |
+
title={DART-Math: Difficulty-Aware Rejection Tuning for Mathematical Problem-Solving},
|
167 |
+
author={Yuxuan Tong and Xiwen Zhang and Rui Wang and Ruidong Wu and Junxian He},
|
168 |
+
year={2024},
|
169 |
+
eprint={2407.13690},
|
170 |
+
archivePrefix={arXiv},
|
171 |
+
primaryClass={cs.CL},
|
172 |
+
url={https://arxiv.org/abs/2407.13690},
|
173 |
+
}
|
174 |
+
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|