|
|
--- |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- merge |
|
|
- mergekit |
|
|
- lazymergekit |
|
|
- NousResearch/Meta-Llama-3.1-8B-Instruct |
|
|
- EpistemeAI/Fireball-Alpaca-Llama3.1.07-8B-Philos-Math-KTO-beta |
|
|
- nvidia/OpenMath2-Llama3.1-8B |
|
|
base_model: |
|
|
- NousResearch/Meta-Llama-3.1-8B-Instruct |
|
|
- EpistemeAI/Fireball-Alpaca-Llama3.1.07-8B-Philos-Math-KTO-beta |
|
|
- nvidia/OpenMath2-Llama3.1-8B |
|
|
--- |
|
|
|
|
|
# Llama-3.1-8B-Squareroot |
|
|
|
|
|
This is a TIES merge that combines the performance of the following models: |
|
|
* [NousResearch/Meta-Llama-3.1-8B-Instruct](https://huggingface.co/NousResearch/Meta-Llama-3.1-8B-Instruct) |
|
|
* [EpistemeAI/Fireball-Alpaca-Llama3.1.07-8B-Philos-Math-KTO-beta](https://huggingface.co/EpistemeAI/Fireball-Alpaca-Llama3.1.07-8B-Philos-Math-KTO-beta) |
|
|
* [nvidia/OpenMath2-Llama3.1-8B](https://huggingface.co/nvidia/OpenMath2-Llama3.1-8B) |
|
|
|
|
|
%3C!-- HTML_TAG_END --> |
|
|
|
|
|
# Description |
|
|
|
|
|
I observed that when a model is trained to do just math, it does badly on everything else. So my plan was to merge a “math” model with a strong reasoning/inference model and a general instruction-following model. The result should be a model that's steerable (able to follow instructions) and still good at math. |
|
|
|
|
|
# Examples |
|
|
|
|
|
# Benchmarks |
|
|
|
|
|
This model ranks in the top 5 for MATH benchmarks, but is severely bad on the rest (which isn't quite what I was expecting). I’m hoping to improve its general abilities without losing its math skills. Qwen still has the top spot :( |
|
|
|
|
|
|
|
|
%3C!-- HTML_TAG_END --> |
|
|
|
|
|
|
