WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instruct (RLEIF)
π Home Page
π€ HF Repo β’π± Github Repo β’ π¦ Twitter
π [WizardLM] β’ π [WizardCoder] β’ π [WizardMath]
π Join our Discord
News
[12/19/2023] π₯ We released WizardMath-7B-V1.1 trained from Mistral-7B, the SOTA 7B math LLM, achieves 83.2 pass@1 on GSM8k, and 33.0 pass@1 on MATH.
[12/19/2023] π₯ WizardMath-7B-V1.1 outperforms ChatGPT 3.5, Gemini Pro, Mixtral MOE, and Claude Instant on GSM8K pass@1.
[12/19/2023] π₯ WizardMath-7B-V1.1 is comparable with ChatGPT 3.5, Gemini Pro, and surpasses Mixtral MOE on MATH pass@1.
| Model | Checkpoint | Paper | GSM8k | MATH |
|---|---|---|---|---|
| WizardMath-7B-V1.1 | π€ HF Link | π [WizardMath] | 83.2 | 33.0 |
| WizardMath-70B-V1.0 | π€ HF Link | π [WizardMath] | 81.6 | 22.7 |
| WizardMath-13B-V1.0 | π€ HF Link | π [WizardMath] | 63.9 | 14.0 |
| WizardMath-7B-V1.0 | π€ HF Link | π [WizardMath] | 54.9 | 10.7 |
[12/19/2023] Comparing WizardMath-7B-V1.1 with other open source 7B size math LLMs.
| Model | GSM8k Pass@1 | MATH Pass@1 |
|---|---|---|
| MPT-7B | 6.8 | 3.0 |
| Llama 1-7B | 11.0 | 2.9 |
| Llama 2-7B | 12.3 | 2.8 |
| Yi-6b | 32.6 | 5.8 |
| Mistral-7B | 37.8 | 9.1 |
| Qwen-7b | 47.8 | 9.3 |
| RFT-7B | 50.3 | -- |
| MAmmoTH-7B (COT) | 50.5 | 10.4 |
| WizardMath-7B-V1.0 | 54.9 | 10.7 |
| Abel-7B-001 | 59.7 | 13 |
| MetaMath-7B | 66.5 | 19.8 |
| Arithmo-Mistral-7B | 74.7 | 25.3 |
| MetaMath-Mistral-7B | 77.7 | 28.2 |
| Abel-7B-002 | 80.4 | 29.5 |
| WizardMath-7B-V1.1 | 83.2 | 33.0 |
[12/19/2023] Comparing WizardMath-7B-V1.1 with large open source (30B~70B) LLMs.
| Model | GSM8k Pass@1 | MATH Pass@1 |
|---|---|---|
| Llemma-34B | 51.5 | 25.0 |
| Minerva-62B | 52.4 | 27.6 |
| Llama 2-70B | 56.8 | 13.5 |
| DeepSeek 67B | 63.4 | -- |
| Gork 33B | 62.9 | 23.9 |
| MAmmoTH-70B | 72.4 | 21.1 |
| Yi-34B | 67.9 | 15.9 |
| Mixtral 8x7B | 74.4 | 28.4 |
| MetaMath-70B | 82.3 | 26.6 |
| WizardMath-7B-V1.1 | 83.2 | 33.0 |
β Data Contamination Check:
Before model training, we carefully and rigorously checked all the training data, and used multiple deduplication methods to verify and prevent data leakage on GSM8k and MATH test set.
| Model | Checkpoint | Paper | MT-Bench | AlpacaEval | GSM8k | HumanEval | License |
|---|---|---|---|---|---|---|---|
| WizardLM-70B-V1.0 | π€ HF Link | πComing Soon | 7.78 | 92.91% | 77.6% | 50.6 pass@1 | Llama 2 License |
| WizardLM-13B-V1.2 | π€ HF Link | 7.06 | 89.17% | 55.3% | 36.6 pass@1 | Llama 2 License | |
| WizardLM-13B-V1.1 | π€ HF Link | 6.76 | 86.32% | 25.0 pass@1 | Non-commercial | ||
| WizardLM-30B-V1.0 | π€ HF Link | 7.01 | 37.8 pass@1 | Non-commercial | |||
| WizardLM-13B-V1.0 | π€ HF Link | 6.35 | 75.31% | 24.0 pass@1 | Non-commercial | ||
| WizardLM-7B-V1.0 | π€ HF Link | π [WizardLM] | 19.1 pass@1 | Non-commercial | |||
| Model | Checkpoint | Paper | HumanEval | MBPP | Demo | License |
|---|---|---|---|---|---|---|
| WizardCoder-Python-34B-V1.0 | π€ HF Link | π [WizardCoder] | 73.2 | 61.2 | Demo | Llama2 |
| WizardCoder-15B-V1.0 | π€ HF Link | π [WizardCoder] | 59.8 | 50.6 | -- | OpenRAIL-M |
| WizardCoder-Python-13B-V1.0 | π€ HF Link | π [WizardCoder] | 64.0 | 55.6 | -- | Llama2 |
| WizardCoder-Python-7B-V1.0 | π€ HF Link | π [WizardCoder] | 55.5 | 51.6 | Demo | Llama2 |
| WizardCoder-3B-V1.0 | π€ HF Link | π [WizardCoder] | 34.8 | 37.4 | -- | OpenRAIL-M |
| WizardCoder-1B-V1.0 | π€ HF Link | π [WizardCoder] | 23.8 | 28.6 | -- | OpenRAIL-M |
Github Repo: https://github.com/nlpxucan/WizardLM/tree/main/WizardMath
Twitter: https://twitter.com/WizardLM_AI/status/1689998428200112128
Discord: https://discord.gg/VZjjHtWrKs
Comparing WizardMath-V1.0 with Other LLMs.
π₯ The following figure shows that our WizardMath-70B-V1.0 attains the fifth position in this benchmark, surpassing ChatGPT (81.6 vs. 80.8) , Claude Instant (81.6 vs. 80.9), PaLM 2 540B (81.6 vs. 80.7).
βNote for model system prompts usage:
Please use the same systems prompts strictly with us, and we do not guarantee the accuracy of the quantified versions.
Default version:
"Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:"
CoT Version: οΌβFor the simple math questions, we do NOT recommend to use the CoT prompt.οΌ
"Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response: Let's think step by step."
Inference WizardMath Demo Script
We provide the WizardMath inference demo code here.
βTo commen concern about dataset:
Recently, there have been clear changes in the open-source policy and regulations of our overall organization's code, data, and models. Despite this, we have still worked hard to obtain opening the weights of the model first, but the data involves stricter auditing and is in review with our legal team . Our researchers have no authority to publicly release them without authorization. Thank you for your understanding.
Citation
Please cite the repo if you use the data, method or code in this repo.
@article{luo2023wizardmath,
title={WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instruct},
author={Luo, Haipeng and Sun, Qingfeng and Xu, Can and Zhao, Pu and Lou, Jianguang and Tao, Chongyang and Geng, Xiubo and Lin, Qingwei and Chen, Shifeng and Zhang, Dongmei},
journal={arXiv preprint arXiv:2308.09583},
year={2023}
}
- Downloads last month
- 280