metadata
license: ms-pl
tags:
- merge
This model is a result of merging three Orca2-13B models with itself using 'mergekit-legacy'. Merge parameters were passthrough for one and --weight 0.5 --density 0.5 for the second merge.
This merged model showed marginal improvement in perplexity scores:
The perplexity for Orca-2-13b is: 7.595028877258301 The perplexity for orca2-26B-self-merge is: 7.550178050994873 The perplexity for orca2-39B-self-merge is: NC
The following table summarizes the model performance across a range of benchmarks:
| Model | Average | ARC | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K |
|---|---|---|---|---|---|---|---|
| microsoft/Orca-2-13b | 58.64 | 60.67 | 79.81 | 60.37 | 56.41 | 76.64 | 17.97 |
| vmajor/Orca2-13B-selfmerge-26B | 62.24 | 60.84 | 79.84 | 60.32 | 56.38 | 76.87 | 39.2 |
| vmajor/Orca2-13B-selfmerge-39B | 62.24 | 60.84 | 79.84 | 60.32 | 56.38 | 76.87 | 39.2 |
Interestingly the GSM8K performance more than doubled with the first self merge. Second self merge resulting in the 39B model did not produce any further gains.