|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- EleutherAI/muInstruct |
|
- camel-ai/math |
|
language: |
|
- en |
|
tags: |
|
- math |
|
--- |
|
|
|
`llemma_7b_muinstruct_camelmath` is an instruction-following finetune of [Llemma 7B](https://huggingface.co/EleutherAI/llemma_7b), trained on the [μInstruct](https://huggingface.co/datasets/EleutherAI/muInstruct) and [camel-ai/math](https://huggingface.co/datasets/camel-ai/math) datasets. |
|
|
|
## Input Formatting |
|
Format input queries as follows: |
|
``` |
|
input_text = f"Input:{input}\n\nResponse:" |
|
``` |
|
|
|
Note that due to an error during training, this model's end-of-sequence token ID is `0` instead of the `2` which is standard for Llama-2 based models. Inference APIs should handle this automatically by reading this repo's `config.json`, but be aware of this difference if you are doing token surgery. |
|
|
|
## Evals |
|
` |
|
llemma_7b_muinstruct_camelmath` compares favorably to other 7B parameter models on the [Hungarian Math Exam](https://huggingface.co/datasets/keirp/hungarian_national_hs_finals_exam/blob/main/README.md). It surpasses the few-shot performance of Llemma 7B whilst being the strongest Llama-2 7B based model. |
|
|
|
| Model | Exam Score | |
|
| ------------------------------------------------------------------------------ | ---------- | |
|
| [Code Llama 7B](https://huggingface.co/codellama/CodeLlama-7b-hf) (few-shot) | 8\% | |
|
| [MetaMath 7B](https://huggingface.co/meta-math/MetaMath-7B-V1.0) | 20\% | |
|
| [MAmmoTH 7B](https://huggingface.co/TIGER-Lab/MAmmoTH-7B) | 17\% | |
|
| [MAmmoTH Coder 7B](https://huggingface.co/TIGER-Lab/MAmmoTH-Coder-7B) | 11\% | |
|
| [Llemma 7B](https://huggingface.co/EleutherAI/llemma_7b) (few-shot) | 23\% | |
|
| Llemma_7B_muinstruct_camelmath | 25\% | |
|
| - | - | |
|
| [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1) (few-shot) | 22\% | |
|
| [MetaMath Mistral 7B](https://huggingface.co/meta-math/MetaMath-Mistral-7B) | 29\% | |
|
| [OpenChat 3.5](https://huggingface.co/openchat/openchat_3.5) | 37\% | |
|
|
|
|
|
|