|
--- |
|
library_name: transformers |
|
tags: |
|
- mergekit |
|
- merge |
|
|
|
--- |
|
# arco ultra |
|
|
|
arco ultra is an improvement over [WizardLM-Evol-V2-Unfiltered](https://huggingface.co./trollek/danube2-1.8b-WizardLM-Evol-V2-Unfiltered) which at the time of writting is a state-of-the-art 1.8 billion parameters mistral language model. The model was first merged with another sota model ([Synthia-v1.3](https://huggingface.co./trollek/danube2-1.8b-Synthia-v1.3)) in order to get a balanced, performant model. |
|
|
|
Post-merging was used just like with palmer models (merge, finetune later). |
|
|
|
#### benchmarks |
|
|
|
zero-shot evaluations performed on current sota ~0.5b models. |
|
|
|
| Parameters | Model | MMLU | ARC | HellaSwag | PIQA | Winogrande | Average | |
|
| -----------|--------------------------------|-------|-------|-----------|--------|------------|---------| |
|
| 0.5b | arco |26.17|37.29|62.88|74.37|62.27|52.60| |
|
| 1.8B | wizard |40.79|40.87|71.85|**78.02**|64.33| 59.17| |
|
| 1.8B | niro |**41.75**|**40.96**|**72.07**|77.97|**65.51**|**59.65**| |
|
|
