MetaMath Models by MetaMath
About
The MetaMath family of large language models (LLMs) are fine-tuned models specializing in mathematical reasoning 125. Developed by a team of researchers from various institutions, including the University of Cambridge and Huawei Noah's Ark Lab, MetaMath models are trained on a dataset called MetaMathQA, created by bootstrapping mathematical questions from existing benchmarks like GSM8K and MATH 15. This process involves rewriting questions from multiple perspectives to create a richer and more diverse training set 15. The MetaMath family includes models of varying sizes, such as MetaMath-7B, MetaMath-13B, and MetaMath-70B, each demonstrating improved performance on mathematical reasoning benchmarks compared to other open-source LLMs of similar size 125. The MetaMath-70B model, in particular, achieves accuracy on the GSM8K benchmark that is comparable to GPT-3.5-Turbo. The models and dataset are publicly available 125.
Current Variants
Use-when guidance is derived from seed capabilities, context, release, and replacement fields.
Use when the workload needs 32k context and 7B parameters.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| MetaMath 70B | Use when the workload needs 70B parameters. | 2023-10 | 70B parameters | Current |
| MetaMath 13B | Use when the workload needs 13B parameters. | 2023-10 | 13B parameters | Current |
| MetaMath 7B | Use when the workload needs 7B parameters. | 2023-10 | 7B parameters | Current |
| MetaMath Mistral 7B | Use when the workload needs 32k context and 7B parameters. | 2023-10 | 32k context7B parameters | Current |
| MetaMath Llemma 7B | Use when the workload needs 7B parameters. | 2023-10 | 7B parameters | Current |
Release Timeline
1 release groupSpecifications(5 models)
| Model | Released | Context | Parameters |
|---|---|---|---|
| MetaMath 70B | 2023-10 | — | 70B |
| MetaMath 13B | 2023-10 | — | 13B |
| MetaMath 7B | 2023-10 | — | 7B |
| MetaMath Mistral 7B | 2023-10 | 32k | 7B |
| MetaMath Llemma 7B | 2023-10 | — | 7B |
Frequently Asked Questions
- What is MetaMath used for?
- MetaMath is used for mathematics and math-heavy prompts. The family description and listed model capabilities point to those workloads as the best fit.
- How does MetaMath compare to InternLM2-Math?
- MetaMath by MetaMath is strongest where you need mathematics, while InternLM2-Math by Intern-AI is the closest related family to check for mathematics. MetaMath has 5 listed variants and reaches up to 32k context, while InternLM2-Math reaches up to 64k context, so compare the specs and pricing tables before choosing a production model.
- Which MetaMath model should I use?
- If price is the main constraint, use the pricing table first because MetaMath does not have complete provider pricing in the local data. For the most capable/latest local choice, evaluate MetaMath Mistral 7B with 32k context.
